Multi-source data fusion-based risk assessment method for agricultural ecological environments

Abstract

Introduction:

Agricultural ecological environments represent intricate systems shaped by diverse factors, climate, soil conditions, crop health, and anthropogenic influences. Traditional risk assessment methods often struggle to effectively integrate heterogeneous data sources and account for the dynamic interdependencies within these systems, resulting in limited accuracy and applicability. To overcome these challenges, this study introduces a novel Multi Source Data Fusion Based Risk Assessment Method designed for agricultural ecological environments, leveraging advanced computational techniques to enhance precision and relevance.

Methods:

The proposed methodology consists of three key components preliminaries, the Agricultural Risk Fusion Network (ARFN), and an Adaptive Fusion Strategy (AFS). ARFN integrates feature extraction, graph based fusion, temporal modeling, and risk prediction to deliver accurate and reliable assessments. It captures multidimensional interactions among climatic factors, soil indicators, vegetation indices, and human activities by constructing graph based relationships and modeling temporal dependencies. Spatial features are extracted via convolutional layers, while temporal variations are encoded using recurrent or attention based structures, enabling simultaneous representation of both short term fluctuations and long term trends. The AFS further strengthens the model by dynamically adjusting the fusion weights of heterogeneous inputs based on environmental context and temporal relevance. This adaptive mechanism ensures consistency across varying agricultural conditions and maintains data integrity.

Results and discussion:

Experimental results demonstrate the efficacy of the proposed method, showing that it achieves superior prediction accuracy and interpretability compared to existing approaches. This work provides a robust and adaptive framework for risk assessment in complex agricultural ecological environments, contributing significant advancements to sustainable environmental management.

1 Introduction

Agricultural ecological environments are critical to global food security and sustainable development, yet they face increasing risks due to climate change, human activities, and environmental degradation. Assessing these risks accurately is essential for implementing effective mitigation strategies and ensuring the long term viability of agricultural systems. Not only does risk assessment provide insights into potential threats, but it also enables policymakers and stakeholders to make informed decisions regarding resource allocation and environmental management. In this study, agricultural ecological environments are considered as multi scale systems composed of farmland plots, soil units, crop growth areas, climatic zones, and surrounding ecological components. The associated risks are evaluated not only at the level of individual fields, but also at regional and landscape scales where climate variability, soil degradation, vegetation dynamics, and anthropogenic pressures interact. At the spatial scale, the proposed framework integrates plot level soil information, grid level climate and ecological indicators, and regional crop productivity patterns. At the temporal scale, it considers both short term disturbances, such as drought, flood events, pest outbreaks, and abnormal vegetation stress, and long term processes, such as soil degradation, land use change, biodiversity loss, and climate induced changes in agricultural productivity. This multi scale perspective allows the risk assessment framework to characterize localized ecological stress as well as broader agricultural vulnerability, thereby supporting field management, regional planning, and environmental decision making. Traditional methods often fail to capture the complexity and dynamic nature of agricultural ecosystems, necessitating the development of advanced approaches that integrate diverse data sources and leverage advanced computational techniques. By fusing data from multiple sources, researchers can achieve a more comprehensive understanding of ecological risks, which not only enhances prediction accuracy but also facilitates proactive interventions to safeguard agricultural productivity and environmental health (Firoozshahian et al., 2023).

Early efforts to assess risks in agricultural ecological environments focused on manually defining relationships between environmental factors and potential threats. These methods relied heavily on expert knowledge and predefined rules to model the interactions within ecosystems. While these approaches provided interpretability and consistency, they were constrained by their inability to adapt to the dynamic and multifaceted nature of agricultural systems (Dhelim et al., 2022). The reliance on static frameworks limited their scalability and effectiveness in handling large datasets or rapidly changing conditions (Khanal et al., 2019). These challenges highlighted the need for more flexible and data-driven methodologies capable of capturing the complexity of ecological risks.

To address these limitations, researchers began employing statistical and algorithmic techniques that could learn patterns from structured datasets. Methods such as regression analysis, decision trees, and clustering algorithms demonstrated improved adaptability and scalability in risk assessment tasks (Chaney et al., 2017). These approaches allowed for the identification of correlations and trends within agricultural data, offering valuable insights into ecosystem dynamics. The effectiveness of these models was often hindered by the need for extensive feature engineering and domain-specific expertise, which could be resource-intensive and time-consuming (Elkahky et al., 2015). Traditional models struggled to capture intricate, non-linear interactions between variables, which are essential for understanding the complex nature of ecological risks.

The introduction of advanced computational models has transformed risk assessment in agricultural ecological environments. Techniques such as neural networks and ensemble learning have enabled the automatic extraction of meaningful patterns from diverse data sources, including satellite imagery, sensor data, and environmental records (Zhang et al., 2023). These models have significantly enhanced prediction accuracy and reduced the dependency on manual feature engineering, allowing researchers to focus on higher-level analysis and decision-making (Wei et al., 2021). Despite their advantages, these methods often require substantial computational resources and large labeled datasets, which can be challenging to obtain in agricultural contexts. Moreover, concerns about model interpretability and transparency remain, particularly in scenarios where critical decisions must be made (Shi et al., 2020). Addressing these challenges necessitates the development of innovative approaches that balance computational efficiency, accuracy, and interpretability.

Based on the aforementioned limitations, this study proposes a multi source data fusion based risk assessment method for agricultural ecological environments. The proposed approach integrates symbolic AI, machine learning, and deep learning techniques to leverage their respective strengths while mitigating their weaknesses. By incorporating domain knowledge into the model design, the proposed framework enhances interpretability and ensures that the system aligns with real world agricultural practices. In this study, interpretability refers to the ability of the model to provide understandable explanations of how different environmental factors, such as climate variables, soil indicators, vegetation conditions, and human activities, contribute to the final ecological risk prediction. Specifically, the model is expected not only to generate a risk score, but also to indicate the relative importance of different data sources, the relationships among risk factors, and the temporal changes that lead to increased or reduced ecological vulnerability. Interpretability is important because agricultural ecological risk assessment is closely related to practical decision making. Policymakers, environmental managers, agricultural extension services, and farmers need to understand the main drivers of predicted risks before taking actions such as soil conservation, irrigation adjustment, pest control, or ecological restoration. An interpretable model can improve stakeholder trust, support evidence based intervention, and reduce the risk of applying mitigation strategies without understanding their ecological basis. The proposed method employs advanced data fusion techniques to combine heterogeneous data sources, such as satellite imagery, sensor readings, and historical records, enabling a comprehensive analysis of ecological risks. This approach not only improves prediction accuracy but also enhances the model's adaptability to dynamic environmental conditions. The computational challenges associated with deep learning are addressed by optimizing model architectures and utilizing efficient training strategies, ensuring that the method is both scalable and accessible to researchers and practitioners in the field. By bridging the gap between traditional and modern approaches, the proposed method represents a significant step forward in risk assessment for agricultural ecological environments.

The main contributions of this study are summarized as follows:

• This study develops an integrated risk assessment framework for agricultural ecological environments by combining multi source data fusion, graph based representation learning, temporal modeling, and domain knowledge. Rather than claiming an entirely new paradigm, the framework adapts and integrates existing computational techniques to address the practical characteristics of agricultural ecological risk assessment.

• The proposed framework incorporates heterogeneous information from climate, soil, crop growth, remote sensing, and human activity related data sources. By modeling spatial relationships, temporal dependencies, and cross source interactions, the method provides a more comprehensive representation of agricultural ecological risk than single source or static assessment approaches.

• The study evaluates the proposed framework on multiple agricultural ecological datasets and compares it with representative baseline models. The results show that the integrated framework improves prediction performance and provides more interpretable risk information, which can support ecological monitoring, agricultural management, and environmental decision making.

2 Related work

2.1 Multi source data fusion techniques

2.1.1 Fusion strategies in agricultural ecological assessment

The integration of multi source data has emerged as a fundamental approach for improving agricultural ecological risk assessment, as it enables diverse datasets to be jointly analyzed to enhance analytical precision and reliability. Multi source data fusion combines heterogeneous information, including remote sensing imagery, meteorological observations, soil characteristics, and crop health metrics, to construct a comprehensive understanding of ecological risks (Zhang et al., 2023). By integrating these complementary sources, risk assessment models can move beyond isolated indicators and better reflect the interactions among climate, soil, vegetation, and human activities in agricultural ecosystems. Existing fusion strategies can generally be categorized into data level fusion, feature level fusion, and decision level fusion (Sun et al., 2022). Data level fusion emphasizes the preservation of spatial, temporal, and spectral attributes during the integration of raw data from multiple sources (Wei et al., 2021). Feature level fusion focuses on extracting and synthesizing relevant features from disparate datasets to improve the predictive capability of risk models (Yang et al., 2020). Decision level fusion aggregates outputs from different models or algorithms, thereby enhancing the robustness and reliability of assessments (Shi et al., 2020).

2.1.2 Learning based fusion methods

Machine learning and deep learning algorithms have significantly contributed to the evolution of these fusion techniques, with convolutional neural networks excelling in spatial data analysis and recurrent neural networks adept at modeling temporal dynamics (Tang, 2019). Hybrid architectures that integrate these neural network models have demonstrated efficacy in capturing complex interdependencies among diverse data sources (Son and Kim, 2017). Ensemble learning methods, including random forests and gradient boosting machines, are frequently employed to consolidate predictions from multiple models, further improving the accuracy of risk assessments (Vartak et al., 2016).

2.1.3 Fusion definition and remaining challenges

In this study, fusion refers to the structured integration of heterogeneous agricultural ecological information from multiple sources into a unified representation for risk assessment. The data sources considered include remote sensing imagery, meteorological observations, soil quality measurements, crop growth indicators, and human activity related ecological pressure variables. These sources describe different aspects of agricultural ecological environments, including vegetation status, climate variability, soil fertility, land use change, water stress, pesticide application, and other anthropogenic disturbances. The proposed framework mainly performs fusion at the feature level and model level. At the feature level, modality specific encoders transform raw inputs from different sources into latent feature representations, such as spectral vegetation features from remote sensing images, temporal climate features from meteorological records, physicochemical soil features from field measurements, and ecological pressure features from land use or agricultural management data. These representations are projected into a shared latent space for joint analysis. At the model level, graph based propagation is used to model relationships among different data sources and environmental factors, while the temporal prediction module captures the evolution of fused risk representations over time. The Adaptive Fusion Strategy further adjusts the contribution of different sources according to data quality, temporal relevance, and environmental context. Therefore, fusion in this study does not refer only to simple concatenation of raw data, but to a structured integration process involving feature transformation, inter source relationship modeling, temporal alignment, and adaptive weighting for ecological risk prediction. Challenges such as data heterogeneity, missing data, and computational demands persist, necessitating the development of standardized preprocessing protocols, imputation strategies, and advanced computational frameworks (Kim and Chen, 2015). Despite these obstacles, multi source data fusion remains a pivotal area of research, offering transformative potential for enhancing agricultural ecological risk assessments (Chen et al., 2025).

2.2 Risk assessment models and frameworks

2.2.1 Probabilistic and process based risk models

Risk assessment models and frameworks provide quantitative tools for estimating the likelihood, intensity, and spatial distribution of adverse events in agricultural ecological systems. Early and widely used approaches include probabilistic models, such as Bayesian networks and Monte Carlo simulations, which are designed to represent uncertainty in climate, soil, crop, and management variables (Forouzandeh et al., 2022). Bayesian networks are particularly useful for describing conditional dependencies among risk drivers, such as the relationship between precipitation deficits, soil moisture decline, crop stress, and yield loss (Yang et al., 2022). Their graphical structure makes it possible to identify critical risk factors and interpret causal or probabilistic pathways. Monte Carlo simulations support scenario based risk estimation by repeatedly sampling from predefined probability distributions, which allows researchers to evaluate the possible range of ecological outcomes under uncertain environmental conditions (Cabrera-Sánchez et al., 2020). Process based models represent another important line of research. Models such as the Soil and Water Assessment Tool and the Decision Support System for Agrotechnology Transfer simulate physical and biological processes, including hydrological flow, soil erosion, nutrient cycling, crop growth, and management responses (Ginart et al., 2021). These models have contributed important milestones to agricultural risk assessment because they connect environmental mechanisms with observable ecological outcomes. For example, by integrating soil properties, climatic records, and crop management practices, process based models can estimate the effects of drought, nutrient depletion, or land management change on agricultural productivity and ecosystem stability (Peake and Wang, 2018). However, these models usually require extensive calibration, detailed local parameters, and continuous validation before they can be transferred to new regions or different agricultural systems (Zhao and Guo, 2017). This limits their scalability when heterogeneous data sources and rapidly changing environmental conditions are involved.

2.2.2 Data driven and explainable risk assessment models

With the increasing availability of remote sensing imagery, sensor observations, meteorological records, and field survey data, data driven models have become central to agricultural ecological risk assessment. Machine learning based models, including support vector machines, random forests, and deep neural networks, can process high dimensional datasets and identify nonlinear relationships among environmental variables (Alhamid et al., 2016). Compared with rule based or purely process based methods, these models reduce the dependence on manually specified relationships and can capture complex patterns in large scale agricultural data. Deep learning models have further expanded the capacity of risk assessment frameworks by extracting spatial features from satellite imagery, temporal features from climate records, and latent representations from heterogeneous ecological indicators. Nevertheless, these methods also introduce new challenges. Many models rely on large labeled datasets, may perform poorly when data are missing or temporally misaligned, and often provide limited interpretability for ecological decision making. To address this issue, explainable artificial intelligence techniques have been incorporated into risk models to reveal feature importance, identify dominant ecological stressors, and improve stakeholder trust in model outputs (Wang et al., 2015). However, existing explainable methods often focus on post hoc interpretation and do not fully resolve the need for transparent integration of multi source ecological information.

2.2.3 Framework level requirements and research gap

Comprehensive agricultural ecological risk assessment frameworks generally include data collection, preprocessing, feature extraction, model development, validation, interpretation, and decision support (Trichopoulos et al., 2023). These steps are necessary because risk assessment outputs must not only be accurate, but also usable for practical interventions such as irrigation scheduling, soil conservation, pest control, and regional ecological planning. Stakeholder engagement and domain expertise are also important for ensuring that model outputs reflect real agricultural management needs and local ecological conditions (Kreutz and Schenkel, 2022). Despite these advances, existing frameworks still face several limitations. Probabilistic models are interpretable but often depend on simplified assumptions. Process based models are mechanistically meaningful but require intensive calibration. Conventional machine learning models improve prediction performance but may inadequately represent spatial and temporal dependencies. Deep learning models can extract complex patterns but often lack ecological transparency and robustness under heterogeneous or incomplete data conditions. These limitations indicate the need for a framework that can integrate multi source data, model relationships among ecological factors, capture temporal risk evolution, and retain sufficient interpretability for decision support. The proposed method is developed to address these requirements by combining multimodal feature representation, graph based relationship modeling, temporal risk prediction, and adaptive fusion.

2.3 Applications in agricultural sustainability

2.3.1 Precision management and climate adaptation

The application of multi source data fusion in agricultural ecological risk assessment has direct implications for sustainable agriculture, especially in precision management and climate adaptation. In precision agriculture, remote sensing imagery, soil measurements, meteorological observations, and crop growth indicators are jointly used to support decisions on irrigation, fertilization, pest control, and field level resource allocation (Zhang et al., 2023). For example, vegetation indices derived from satellite imagery can indicate crop stress, while soil moisture and nutrient measurements can help determine whether the stress is caused by water shortage, soil degradation, or management practices. The integration of these indicators enables more targeted interventions and helps reduce excessive use of water, fertilizer, and pesticides (Sun et al., 2022). Climate change impact assessment is another important application area. Agricultural systems are increasingly exposed to drought, floods, heat stress, and unstable precipitation patterns. Multi source data fusion supports the analysis of historical climate variability, the projection of future risk scenarios, and the identification of regions that are vulnerable to extreme weather events (Wei et al., 2021). These analyses can guide adaptation strategies such as drought resistant crop selection, water saving irrigation, adjustment of planting schedules, and regional risk zoning (Yang et al., 2020). Therefore, data fusion based risk assessment provides a technical basis for shifting agricultural management from reactive responses to proactive prevention.

2.3.2 Soil health, ecological pressure, and policy support

Soil health monitoring and ecological pressure assessment are also central to agricultural sustainability. The integration of soil sensor data, laboratory measurements, satellite imagery, and field surveys enables the detection of soil degradation, nutrient imbalance, salinization, and erosion risk (Shi et al., 2020). Such information can guide conservation practices, including contour farming, cover cropping, crop rotation, reduced tillage, and agroforestry, which help maintain soil fertility and prevent further ecological degradation (Tang, 2019). In this context, fused data provide more reliable evidence than single indicators because soil degradation is often associated with interacting drivers, such as climate stress, land use change, irrigation intensity, and fertilizer application. The incorporation of socioeconomic and management related data further expands the role of risk assessment in sustainability planning. Variables such as land use patterns, market conditions, pesticide application, irrigation practices, and policy interventions can be analyzed together with environmental indicators to reveal trade offs between agricultural production and ecological conservation (Son and Kim, 2017). This makes it possible to identify areas where high productivity depends on unsustainable resource use, or where ecological restoration may require economic incentives and technical support (Vartak et al., 2016). Thus, multi source risk assessment can inform policy design, resource prioritization, and regional environmental governance.

2.3.3 Ecological theory and motivation for the proposed framework

From an ecological theoretical perspective, agricultural ecosystems are complex systems characterized by interconnectivity, feedback mechanisms, and resilience (Wang et al., 2022). Ecosystem interconnectivity means that changes in one component can influence other components. For example, soil degradation may affect water retention, vegetation growth, biodiversity, and crop productivity at the same time. This theoretical principle motivates the use of graph based modeling in the proposed framework, because graph structures can represent relationships among climate, soil, vegetation, and human activity related variables. Feedback mechanisms are also important in agricultural ecological systems. A short term climatic anomaly may gradually influence soil moisture, crop growth, pest occurrence, and final yield, while management responses may further alter future ecological conditions. This motivates the use of temporal modeling in the Agricultural Risk Fusion Network, which captures the evolution of risk over time (Zhao et al., 2023). Resilience theory emphasizes the ability of ecosystems to absorb disturbances and reorganize under changing conditions. The proposed framework reflects this concept by incorporating both short term fluctuations and long term trends into risk prediction. By linking ecological theory with computational modeling, the framework aims to produce risk estimates that are not only predictive, but also consistent with the systemic nature of agricultural ecological environments (Shuai and Pang, 2025).

3 Method

3.1 Overview

The proposed method takes heterogeneous agricultural ecological data as input, including remote sensing imagery, meteorological observations, soil quality indicators, crop growth information, and human activity related ecological pressure variables. These data sources are first transformed into modality specific feature representations and then integrated into a unified risk representation. The final output is a spatially and temporally defined ecological risk score, which indicates the probability or severity of agricultural ecological risk in a given region and time period.

The overall workflow consists of three main steps. Section 3.2 defines the problem setting, input data representation, fusion objective, and risk prediction target. This step provides the notation and mathematical basis for multi source ecological risk assessment. Section 3.3 introduces the Agricultural Risk Fusion Network. This module extracts features from different data sources, models relationships among environmental factors through graph based propagation, and captures temporal risk evolution through a recurrent or attention based prediction module. It is designed to represent both cross source dependencies and temporal changes in agricultural ecological conditions. Section 3.4 presents the Adaptive Fusion Strategy. This module adjusts the contribution of each data source according to data quality, temporal relevance, and environmental context. By combining feature transformation, temporal alignment, graph based relationship modeling, and adaptive weighting, the proposed framework supports robust ecological risk prediction under heterogeneous and dynamic agricultural conditions.

3.2 Preliminaries

This section formalizes the problem of risk assessment in agricultural ecological environments through the integration of multi-source data. The objective is to construct a framework that synthesizes various data sources to evaluate and predict risks within agricultural ecosystems. This involves defining mathematical representations of the data, elucidating the relationships between different data sources, and establishing the principles guiding the fusion process.

Consider the set of data sources denoted by , where each D_i signifies a distinct data source providing pertinent information about the agricultural environment. These sources encompass satellite imagery, weather data, soil quality measurements, and crop health indicators. Each data source D_i is represented as a multidimensional array , with m_i and n_i indicating the dimensions of the data.

The fusion process integrates these data sources into a unified representation that encapsulates the essential features for risk assessment. A fusion function is defined, where k represents the dimensionality of the fused data representation. The function F is crafted to extract and amalgamate relevant features from each data source, considering spatial, temporal, and contextual relationships.

To model interactions between different data sources, transformation functions {T₁, T₂, …, T_n} are introduced, where each maps data from source D_i to a feature space of dimension p_i. The transformed features are concatenated to form a composite feature vector Z = [T₁(X₁), T₂(X₂), …, T_n(X_n)].

The risk assessment model is constructed upon this composite feature vector Z. A risk function R:ℝ^k → ℝ is defined to quantify the risk level associated with the current state of the agricultural environment. The risk function is parameterized by weights w ∈ ℝ^k, learned from historical data using an optimization criterion. The risk score is computed as Equation 1:

where b is a bias term. The objective is to minimize the discrepancy between predicted risk scores and actual observed risks, formulated as an optimization problem Equation 2:

where y_j denotes the true risk level for the j-th observation, and N is the total number of observations.

To enhance the robustness of the fusion process, a regularization term is incorporated into the optimization objective. This term penalizes overly complex models and mitigates overfitting. The regularized objective function is expressed as Equation 3:

where λ is a regularization parameter that balances fitting the data and maintaining model simplicity.

To connect the above formulation with the actual implementation, the variables and functions are explicitly mapped to the data and model components used in this study. Each data source D_i corresponds to a concrete agricultural ecological modality, including climate variables such as temperature and precipitation, soil properties such as pH and nutrient levels, remote sensing features such as NDVI and land use patterns, crop related indicators such as yield and growth status, and ecological pressure variables such as pesticide use and irrigation intensity. The input matrix represents the observed features of modality i, where m_i denotes the number of spatial or temporal samples and d_i denotes the feature dimension. The transformation function T_i(·) or f_i(·) is implemented as a modality specific encoder in the Agricultural Risk Fusion Network, which maps heterogeneous inputs into latent representations Z_i. The fused representation Z = [T₁(X₁), T₂(X₂), …, T_n(X_n)] captures the integrated ecological state of a region. The risk function R(Z) is realized by the prediction layers of the model, producing risk scores or categories that are compared with ground truth labels from the datasets used in the experiments. The optimization objectives in Equations 2 and 3 are used to learn model parameters that minimize prediction error while controlling model complexity. Agricultural ecological risk in this study refers to the probability and severity of negative impacts on agricultural ecosystems caused by environmental stressors and human activities, including climate extremes, soil degradation, water scarcity, and ecological disturbance. The model outputs spatially and temporally defined risk scores that can be interpreted as indicators of system vulnerability. These outputs support applications such as early warning, resource allocation, and adaptive management. Within the proposed framework, the fused representation is further used to construct graph based relationships among ecological factors, while the temporal module captures the evolution of risk over time, forming a complete pipeline from multi source data integration to practical risk prediction and decision support.

3.3 Agricultural Risk Fusion Network (ARFN)

In this subsection, we introduce the Agricultural Risk Fusion Network (ARFN), a novel model designed to address the challenges of multi-source data fusion for risk assessment in agricultural ecological environments. As shown in Figure 1, the Agricultural Risk Fusion Network consists of three major modules: a multimodal encoder architecture, a graphical propagation layer, and a temporal risk prediction module. The model first receives heterogeneous agricultural ecological data sources, including climate data, soil data, and remote sensing data. Each data source is represented as an input matrix and processed by a modality specific encoder to obtain latent feature representations. These encoded features are then concatenated to form the initial node embeddings for graph based fusion. The graphical propagation layer constructs a graph among data sources and environmental factors, where the adjacency matrix describes the strength of inter source relationships. Through graph convolutional propagation, the model updates node embeddings and captures dependencies among climate, soil, vegetation, and remote sensing variables. The resulting graph enhanced representations are then passed into the temporal risk prediction module, where recurrent units such as LSTM or GRU are used to model risk evolution across time steps. A fully connected layer with a softmax function generates probabilistic risk scores for different risk categories, and the model is trained using cross entropy loss with ground truth labels.

Figure 1

3.3.1 Multimodal encoder architecture

The ARFN is structured as a multi-layer architecture, where each layer is responsible for processing specific types of data and extracting meaningful features. Let represent the set of n heterogeneous data sources, where each D_i corresponds to a specific type of data. Each data source D_i is represented as a feature matrix , where m_i is the number of samples and d_i is the dimensionality of the features. For each data source D_i, a feature transformation function maps the raw features into a latent space. The transformed features are denoted as Z_i = f_i(X_i), where . The transformation function f_i can be implemented using neural networks, kernel methods, or other machine learning techniques, depending on the nature of the data. This multimodal encoder architecture ensures that diverse data sources are effectively represented in a unified latent space, enabling subsequent fusion and analysis. As shown in Figure 2, the multimodal encoder maps heterogeneous inputs, including climate data, soil data, remote sensing data, and crop indicators, into a shared latent space. Each data source D_i is represented by a feature matrix and transformed by a modality specific encoder f_i(·) to obtain Z_i = f_i(X_i). The encoded representations are then concatenated as Z = [Z₁, Z₂, …, Z_n] and passed to the downstream fusion module for graph based fusion and temporal risk prediction.

Figure 2

3.3.2 Graphical propagation layer

The ARFN employs a graph-based fusion mechanism to integrate the transformed features from all data sources. Let represent a graph, where is the set of nodes corresponding to the data sources and is the set of edges representing relationships between the sources. The adjacency matrix A ∈ ℝ^n×n encodes the connectivity of the graph, with A_ij indicating the strength of the relationship between D_i and D_j. The graph-based fusion is performed using a graph convolutional network (GCN), which updates the node embeddings H^(l) at layer l as follows Equation 4:

where is the node embedding matrix at layer l, is the learnable weight matrix, and σ(·) is a non-linear activation function such as ReLU. The initial node embeddings H⁽⁰⁾ are set to the concatenated feature vectors Z_i from all data sources. This graphical propagation layer captures the interdependencies between data sources, enabling the model to leverage relational information for enhanced feature integration.

3.3.3 Temporal risk prediction module

To capture temporal dependencies in the data, the ARFN incorporates a recurrent neural network (RNN) module. The output of the GCN, H^(L), where L is the number of GCN layers, is fed into the RNN. The RNN processes the sequential data and generates a temporal representation S_t for each time step tEquation 5:

where S_t−1 is the hidden state from the previous time step. The RNN can be implemented using Long Short-Term Memory (LSTM) networks or Gated Recurrent Units (GRUs) to effectively model temporal patterns. The temporal representations S_t are passed through a fully connected layer to produce the risk scores , where k is the number of risk categories. The risk scores are computed as Equation 6:

where and are the learnable parameters of the fully connected layer, and d_s is the dimensionality of S_t. The softmax function ensures that the risk scores are normalized and can be interpreted as probabilities. The ARFN is trained using a supervised learning approach, where the objective is to minimize the cross-entropy loss between the predicted risk scores r_t and the ground truth labels y_t. The loss function is defined as Equation 7:

where T is the total number of time steps, K is the number of risk categories, and y_{t, k} is the ground truth label for category k at time step t. This temporal risk prediction module integrates sequential data modeling and probabilistic reasoning to provide accurate and interpretable risk assessments.

The ARFN integrates feature extraction, graph-based fusion, temporal modeling, and risk prediction into a unified framework. By leveraging the complementary strengths of these components, the ARFN provides a powerful tool for multi-source data fusion and risk assessment in agricultural ecological environments.

3.4 Adaptive fusion strategy for multi-source data integration

In this subsection, we introduce our novel Adaptive Fusion Strategy (AFS), which is specifically designed to address the challenges associated with multi-source data integration for risk assessment in agricultural ecological environments. As shown in the Figure 3, the strategy leverages domain-specific knowledge and incorporates a dynamic optimization framework to ensure effective fusion of heterogeneous data sources while maintaining the integrity and relevance of the information.

Figure 3

3.4.1 Multimodal encoder architecture

The primary objective of AFS is to dynamically adapt to the varying characteristics of multi-source data, such as spatial resolution, temporal frequency, and data quality, to generate a unified representation that facilitates accurate risk assessment. As shown in the Figure 4, to achieve this, AFS employs a multi-layered approach that integrates probabilistic modeling, graph-based representation, and iterative refinement techniques.

Figure 4

Let represent the set of N data sources, where each d_i corresponds to a specific data modality. Each data source d_i is characterized by a feature space and a temporal domain . The fusion process begins by constructing a unified feature space that encapsulates the essential attributes of all data sources. This is achieved through a feature transformation function , defined as Equation 8:

where f_i represents the feature vector of data source d_i, and f_u is the transformed feature vector in the unified space.

To account for the temporal dynamics of the data, we define a temporal alignment function , where is the unified temporal domain. The alignment function ensures that data from different sources are synchronized, enabling coherent analysis. The temporal alignment is expressed as Equation 9:

where t_i and t_u denote the timestamps in the original and unified temporal domains, respectively.

3.4.2 Graphical propagation layer

Once the data is transformed and aligned, AFS employs a graph-based representation to model the relationships between data sources. Let represent the graph, where is the set of nodes corresponding to data sources, and is the set of edges representing inter-source dependencies. The edge weights w_ij are computed using a similarity metric σ(f_i, f_j), defined as Equation 10:

where τ is a scaling parameter that controls the sensitivity of the similarity measure.

The graph is then subjected to iterative refinement using a probabilistic optimization framework. The goal is to maximize the consistency and relevance of the fused data while minimizing redundancy and noise. The optimization problem is formulated as Equation 11:

where is a consistency function that quantifies the agreement between feature vectors f_i and f_j.

3.4.3 Probabilistic optimization framework

To further enhance the robustness of the fusion process, AFS incorporates domain-specific constraints , which are derived from expert knowledge and empirical observations. These constraints are integrated into the optimization framework as penalty terms Equation 12:

where is the original objective function, represents the penalty terms, and λ is a regularization parameter.

The final fused representation F_u is obtained by solving the optimization problem using an iterative algorithm, such as gradient descent or alternating direction method of multipliers (ADMM). The fused data is then utilized for risk assessment through predictive modeling and decision-making processes.

The Adaptive Fusion Strategy provides a comprehensive framework for integrating multi-source data in agricultural ecological environments. By leveraging advanced techniques such as feature transformation, temporal alignment, graph-based modeling, and probabilistic optimization, AFS ensures the generation of a unified representation that is both accurate and informative for risk assessment tasks.

4 Experimental setup

4.1 Task definition, risk factors, and dataset integration

4.1.1 Task definition

The objective of this study is to predict agricultural ecological risk based on multi source data. The model takes heterogeneous inputs derived from climate observations, soil measurements, remote sensing data, crop yield records, and ecological pressure indicators. The output is a spatially and temporally defined risk score or risk category, representing the likelihood or severity of ecological stress in a given region. Depending on the dataset, the prediction targets correspond to labeled risk levels or derived indicators such as climate impact risk, soil degradation risk, crop yield risk, or composite ecological risk indices.

4.1.2 Risk factor definition

Agricultural ecological risk is modeled as the combined effect of multiple interacting factors. Climate variables represent short term environmental variability and extreme events. Soil properties capture long term fertility and degradation conditions. Crop related variables reflect agricultural productivity and system response. Ecological pressure variables describe anthropogenic influences such as irrigation intensity, fertilizer application, pesticide use, and land use change. These factors are treated as complementary components that jointly determine the risk state of agricultural ecosystems.

4.1.3 Dataset relationship and fusion rationale

The four datasets are integrated through spatial and temporal alignment. All data sources are mapped to consistent spatial units, such as grid cells or regions, using georeferenced coordinates. Temporal alignment is achieved by aggregating or interpolating observations to shared time steps, ensuring that climate, soil, crop, and ecological variables correspond to the same observation period. Climate data provide dynamic environmental context, soil data represent baseline ecological conditions, crop yield data reflect agricultural system outcomes, and ecological risk factors capture external environmental and human activity related pressures. The multi source fusion framework is designed to capture interactions among heterogeneous risk factors that cannot be represented by single source models. Feature level fusion maps modality specific data into a shared latent space, graph based modeling captures dependencies among environmental variables and data sources, and temporal modeling represents the evolution of ecological risk over time. By integrating these components, the proposed framework links diverse datasets to a unified prediction task and supports more robust agricultural ecological risk assessment.

4.2 Dataset

The Agricultural Climate Impact Dataset (Wu et al., 2025) is a comprehensive collection of data designed to analyze the effects of climate variability on agricultural systems. It includes detailed records of temperature, precipitation, humidity, and other meteorological factors across diverse geographical regions. The dataset spans multiple decades, enabling researchers to study long-term trends and seasonal patterns. It incorporates information on crop types, planting schedules, and harvest yields, facilitating the exploration of correlations between climatic conditions and agricultural productivity. The dataset is curated to ensure high-quality data integrity, with rigorous preprocessing steps to remove anomalies and fill missing values. Its wide applicability makes it a valuable resource for climate impact modeling, risk assessment, and policy-making in agriculture. The Soil Quality Monitoring Dataset (de Oliveira et al., 2024) provides extensive data on soil properties and health across various agricultural zones. It includes measurements of soil pH, nutrient levels, organic matter content, and microbial activity, among other parameters. The dataset is collected using standardized sampling techniques and advanced laboratory analyses to ensure accuracy and consistency. It also incorporates spatial data, allowing researchers to map soil quality variations and identify areas requiring intervention. This dataset is particularly useful for studies on sustainable land management, soil conservation, and the impact of agricultural practices on soil health. Its detailed structure supports the development of predictive models for soil degradation and fertility assessment. The Crop Yield Prediction Dataset (Cui et al., 2023) is designed to facilitate the modeling and forecasting of crop yields under varying environmental and management conditions. It includes historical data on crop yields, weather patterns, soil characteristics, and farming practices. The dataset is enriched with satellite imagery and remote sensing data, providing spatial and temporal insights into crop growth dynamics. It also incorporates socioeconomic factors such as market prices, labor availability, and policy interventions, enabling holistic analyses of yield determinants. Researchers can leverage this dataset to develop machine learning models for yield prediction, optimize resource allocation, and assess the impact of climate change on food security. The Ecological Risk Factors Dataset (Sukarjo et al., 2022) focuses on identifying and quantifying risks to ecosystems and biodiversity arising from agricultural activities. It includes data on pesticide usage, water consumption, habitat fragmentation, and species population dynamics. The dataset is compiled from multiple sources, including field surveys, remote sensing, and government reports, ensuring a comprehensive representation of ecological risks. It also integrates climate data to assess the compounded effects of environmental stressors on ecosystems. This dataset is instrumental for studies on sustainable agriculture, conservation planning, and the development of strategies to mitigate ecological risks. Its robust structure and diverse data points make it a critical resource for interdisciplinary research on agriculture and environmental sustainability.

Table 1 summarizes the datasets used in this study, including their spatial resolution, temporal coverage, geographic scope, and preprocessing procedures. The datasets span multiple agricultural regions, primarily across Asia and Europe, and include both large-scale gridded observations and plot-level measurements. Preprocessing steps such as missing value interpolation, anomaly removal, normalization, spatial resampling, and temporal alignment are applied to ensure consistency across heterogeneous data sources. Four core datasets are utilized, each selected for its relevance to agricultural ecological risk assessment and multi-source data fusion. The Agricultural Climate Impact Dataset provides over 30 years of meteorological observations from multi-country weather stations in Asia and Europe, including temperature, precipitation, humidity, wind speed, and solar radiation, supporting spatiotemporal climate risk modeling. The Soil Quality Monitoring Dataset consists of field-collected samples from over 600 agricultural plots across multiple agro-ecological zones, including soil pH, organic carbon, and nutrient indicators, enabling the analysis of long-term soil fertility and degradation. The Crop Yield Prediction Dataset integrates satellite imagery and field survey data from three countries, namely China, India, and Spain, combining MODIS and Sentinel-2 vegetation indices with reported crop yields for major crops such as maize, wheat, and rice. The data are preprocessed using cloud masking, radiometric correction, and spatial resampling. The Ecological Risk Factors Dataset integrates environmental and management-related variables from government agencies, FAO databases, and remote sensing products across multiple regions, including pesticide use, irrigation intensity, land-use change, and biodiversity indicators. These datasets collectively provide a comprehensive foundation for modeling spatial heterogeneity, temporal dynamics, and anthropogenic pressures in agricultural ecological risk assessment.

4.3 Experimental details

The experiments were conducted using a state-of-the-art deep learning framework implemented in PyTorch. All models were trained on a high-performance computing cluster equipped with NVIDIA A100 GPUs, each with 40 GB of memory. The training process utilized mixed precision to optimize computational efficiency and reduce memory consumption. The backbone architecture employed in our experiments was ResNet-50, pre-trained on ImageNet, which served as the feature extractor. For temporal modeling, we integrated a Transformer-based module, leveraging its self-attention mechanism to capture long-range dependencies across frames. The batch size was set to 64, and the models were trained for 100 epochs. The initial learning rate was set to 0.001 and decayed using a cosine annealing schedule. The Adam optimizer was employed with β₁ = 0.9 and β₂ = 0.999, ensuring stable convergence during training.

Data augmentation techniques were applied to enhance the robustness of the models. These included random cropping, horizontal flipping, color jittering, and Gaussian noise injection. Temporal augmentation was performed by randomly sampling frames from video sequences to simulate variations in motion dynamics. To ensure fair evaluation, all datasets were split into training, validation, and test sets following standard protocols. The evaluation metrics included top-1 and top-5 accuracy, mean average precision (mAP), and F1-score, depending on the dataset and task requirements. For reproducibility, all random seeds were fixed, and the codebase was made publicly available, including detailed documentation and configuration files. Hyperparameter tuning was conducted using grid search to identify optimal settings for each dataset. The training strategy incorporated early stopping based on validation performance to prevent overfitting.

We provide full details of the hyperparameter configurations used in our experiments. The learning rate was initially set to 0.001 and decayed using a cosine annealing strategy, which gradually reduces the learning rate during training to ensure stable convergence. The Adam optimizer was adopted with parameters β₁ = 0.9 and β₂ = 0.999, and the weight decay was set to 1 × 10⁻⁵ to mitigate overfitting. The batch size was fixed at 64 for all experiments in order to balance memory efficiency and training stability. For the feature extraction backbone, we employed a pre-trained ResNet-50 model, in which the final fully connected layer was replaced by a projection head with a hidden dimension of 512. During fine-tuning, the first two convolutional blocks were frozen to preserve general visual representations learned from large-scale datasets. A dropout rate of 0.3 was applied to the fully connected layers to further reduce overfitting. In the temporal modeling component, a Transformer architecture was utilized, consisting of four encoder layers with eight attention heads and a feedforward network dimension of 2048. Sine-based positional encoding was adopted to model temporal order, and a dropout rate of 0.1 was applied to both the attention and feedforward sublayers. The input sequence length was fixed to 12 time steps, corresponding to 12 temporal snapshots. For the graph-based fusion module, a K-nearest neighbor graph with K = 8 was constructed based on feature similarity. Two graph convolutional layers were used, with hidden dimensions set to 256 and 128, respectively. ReLU was employed as the activation function, and layer normalization was applied after each graph convolutional layer to improve training stability. During training, early stopping was applied when the validation loss failed to improve for 10 consecutive epochs. All experiments were repeated three times, and the average performance was reported. All hyperparameters were selected using grid search based on validation performance for each dataset.

Standard classification metrics such as accuracy, precision, recall, and F1 score are widely used in machine learning. We used these metrics to compare our model with existing baseline models and determine the core performance of our model on datasets with labeled ground truth values. We recognize that ecological risk prediction involves not only correct classification but also minimizing false negatives and underreporting of high-risk areas, as well as understanding the uncertainty of the model. In this context, misclassification can lead to underestimating threats such as soil degradation, crop failure, or pest and disease outbreaks, thus affecting timely intervention. To quantify and analyze this, we expanded the relevant metrics and incorporated uncertainty-aware information into the model's output.

4.4 Comparison with SOTA methods

The experimental results presented in Tables 2, 3 demonstrate the superior performance of our proposed method compared to state-of-the-art (SOTA) approaches across multiple benchmarks. In Table 2, our method consistently achieves higher accuracy across all datasets, including the Agricultural Climate Impact, the Soil Quality Monitoring, and the Crop Yield Prediction. This improvement can be attributed to the novel architecture design, which effectively captures spatiotemporal features and leverages advanced data augmentation techniques to enhance generalization. For instance, the integration of multi-scale feature extraction enables our model to better handle diverse motion patterns, which is particularly beneficial for datasets like the Crop Yield Prediction that contain a wide range of action categories. The use of a robust optimization strategy, including adaptive learning rate schedules and gradient clipping, ensures stable convergence during training, minimizing overfitting and improving performance. The results in Table 2 further highlight the scalability of our approach, as it maintains high accuracy even when applied to large-scale datasets such as the Agricultural Climate Impact, where other methods often struggle due to computational constraints or insufficient feature representation.

Table 3 provides a detailed comparison of our method against SOTA techniques in terms of computational efficiency and inference speed. Our approach demonstrates a significant reduction in computational overhead while maintaining superior accuracy, which is critical for real-world applications requiring real-time processing. This efficiency is achieved through the incorporation of lightweight modules and optimized network architectures, such as depthwise separable convolutions and attention mechanisms, which reduce the number of parameters without compromising feature extraction capabilities. The use of mixed precision training and hardware-specific optimizations, such as GPU acceleration, contributes to faster training and inference times. The results in Table 3 also underline the robustness of our method in handling noisy or incomplete data, a common challenge in practical scenarios. By employing advanced regularization techniques and data augmentation strategies, our model effectively mitigates the impact of data imperfections, outperforming competing methods in terms of both accuracy and reliability.

The performance gains observed in Tables 2, 3 can be attributed to several key innovations in our methodology. First, the use of a hybrid loss function that combines cross-entropy loss with a contrastive loss component enhances the model's ability to distinguish subtle differences between similar classes, which is particularly advantageous for fine-grained action recognition tasks. Second, the incorporation of domain adaptation techniques allows our model to generalize effectively across different datasets, addressing the issue of domain shift that often hampers the performance of SOTA methods. The comprehensive evaluation strategy employed in our experiments ensures the reliability of the reported results, as we utilize multiple metrics, including precision, recall, and F1-score, to provide a holistic assessment of model performance. These advancements collectively contribute to the significant improvements observed in our method, establishing it as a new benchmark for action recognition tasks.

Table 4 presents a quantitative performance comparison on the Agricultural Climate Impact Dataset, where all results are reported as mean values with standard deviations over three repeated experiments. The Random Forest baseline achieves an accuracy of 81.76%, while the Support Vector Machine reaches 82.31%, indicating limited capability in capturing complex climate patterns. Deep learning baselines show incremental improvements, with the one-dimensional CNN obtaining an accuracy of 83.68% and the simple RNN reaching 84.02%. In contrast, the proposed ARFN combined with the Adaptive Fusion Strategy achieves a substantially higher accuracy of 89.34%, along with a precision of 88.76%, a recall of 88.92%, and an F1-score of 88.84%. Compared with the strongest baseline, the proposed method improves accuracy by more than five percentage points while maintaining lower variance across all metrics, demonstrating its effectiveness and robustness in modeling the agricultural climate impact data.

Table 5 reports the quantitative performance of the proposed model under different resource-constrained scenarios on the Agricultural Climate Impact Dataset. When all data sources and model components are available, the full ARFN combined with the Adaptive Fusion Strategy achieves the highest performance, with an accuracy of 89.34% and an F1-score of 88.84%. When only satellite and weather data are used, the model maintains an accuracy of 86.12%, indicating that acceptable performance can still be achieved with reduced data inputs. Removing the graph convolutional module results in an accuracy of 85.78%, while excluding the temporal modeling component further reduces accuracy to 84.67%, reflecting the contribution of spatial and temporal information to overall performance. In addition, the quantized model with INT8 inference attains an accuracy of 87.56% and an F1-score of 87.11%, demonstrating that computational efficiency can be improved with only a limited performance loss. These results collectively suggest that the proposed method remains robust and effective under various resource limitations.

Table 6 presents a comparison of different models using the domain-specific Crop Loss Risk Index (CLRI), where lower values indicate better performance. The Random Forest and SVM baselines yield relatively high mean CLRI values of 0.176 and 0.165, respectively, accompanied by larger variability and higher prediction errors in the top 10% high-risk regions. The one-dimensional CNN achieves a lower mean CLRI of 0.142 with reduced variance, indicating improved risk estimation capability. In contrast, the proposed ARFN combined with the Adaptive Fusion Strategy attains the lowest mean CLRI of 0.092 and the smallest standard deviation of 0.036, while also reducing the top 10% high-risk error to 5.4%. The quantized version of the proposed model maintains comparable performance with a mean CLRI of 0.098 and a top high-risk error of 6.1%. These results demonstrate that the proposed approach provides more accurate and stable crop loss risk estimation, particularly in high-risk scenarios.

While the improved performance of our report such as higher accuracy, F1 scores, and lower misclassification rates demonstrates the technical superiority of our approach, its ecological implications are equally significant. In agro-ecosystems, risk misclassification can lead to delayed false negatives in interventions or false positives in resource mismatches. For example, underestimating soil degradation or failing to detect climate-induced crop stress can result in irreversible yield losses, land abandonment, or increased risk of pests and diseases. The improved accuracy of our model in identifying high-risk areas directly reduces the likelihood of such outcomes. The temporal modeling component enhances the ability to detect both static risks and ecological threats that emerge over time. By capturing early signs of risk evolution, such as a gradual decline in vegetation indices or a sudden surge in abnormal precipitation, the model enables proactive rather than reactive responses. This temporal sensitivity is particularly important for agricultural decision-makers who must plan irrigation, pesticide use, or planting strategies in advance. Integrating heterogeneous data sources such as climate, soil, crop health, and human activities ensures that risk assessments reflect system-wide interactions. This multidimensional perspective allows policymakers and farm managers to prioritize interventions based on a comprehensive risk level rather than isolated indicators. For example, if a region is identified as a high-risk area due to a combination of factors such as poor soil quality, drought, and human disturbance, land remediation, sustainable agriculture incentives, or adaptive technologies can be implemented in that region. The performance improvement of this model is not only reflected in the algorithm's refinement but also in its more reliable and context-aware risk diagnosis. These diagnostic results lay the foundation for evidence-based ecological management, contributing to the long-term sustainable development of agriculture and food security.

Table 7 presents a quantitative comparison of multiple models on agriculture related tasks, highlighting clear performance differences across key indicators. The proposed ARFN model achieves strong results with drought detection accuracy of 89.4 percent, pest outbreak F1 score of 0.912, and soil degradation AUC of 0.903, exceeding all baseline methods. When the adaptive fusion strategy is incorporated, the ARFN with AFS delivers the highest performance, reaching 91.1 percent accuracy in drought detection, an F1 score of 0.927 for pest outbreak identification, and an AUC of 0.916 for soil degradation assessment. These gains are obtained while maintaining competitive inference time close to other deep learning models, demonstrating that the proposed approach offers a favorable balance between predictive accuracy and computational efficiency.

4.5 Ablation study

To evaluate the contribution of individual components in our proposed method, we conducted a comprehensive ablation study. The results of these experiments are summarized in Tables 8, 9. Each module was systematically removed or modified to assess its impact on the performance. This section provides a detailed analysis of the findings and highlights the significance of each component in achieving state-of-the-art results.

The first set of experiments, as shown in Table 8, focuses on the core architectural components of our method. The baseline model, which excludes the Multimodal Encoder Architecture, demonstrates significantly lower performance across all evaluation metrics. This result underscores the importance of effectively representing heterogeneous data sources in a unified latent space. When the Graphical Propagation Layer is incorporated, we observe a substantial improvement in accuracy, indicating its critical role in capturing interdependencies between data sources. The inclusion of the Temporal Risk Prediction Module leads to additional gains, particularly in scenarios involving sequential data. This module effectively models temporal patterns, enhancing the model's ability to generalize to varied data distributions. The final configuration, which combines all components, achieves the highest performance, validating the synergistic effect of these modules.

Table 9 presents the results of experiments designed to evaluate the impact of training strategies and optimization techniques. The use of data augmentation techniques contributes to improved robustness and generalization, as evidenced by the performance gains in the augmented model. The choice of optimizer plays a crucial role; models trained with Adam optimizer outperform those trained with SGD, particularly in terms of convergence speed and stability. The learning rate scheduling strategy further enhances performance by dynamically adjusting the learning rate during training, preventing overfitting and ensuring steady progress. Regularization techniques, including dropout and weight decay, significantly reduce overfitting, as reflected in the improved validation accuracy. These findings highlight the importance of adopting a comprehensive training pipeline to maximize the potential of the proposed method.

4.6 Limitations

While the proposed ARFN and AFS frameworks demonstrate good performance on multiple datasets, some limitations need to be recognized to guide future improvements. The computational complexity of the full model is high, especially when incorporating graph convolutional layers, recurrent or attention-based temporal modules, and high-resolution multi-source inputs, posing a challenge for real-time or resource-constrained applications. In scenarios requiring low-latency decision-making, even with GPU acceleration and optimization techniques, inference speed may still become a bottleneck. Model performance is highly dependent on the availability and quality of labeled multi-source data. Missing modalities, noisy inputs, or timestamp mismatches between different data sources significantly degrade performance. Although AFS introduces dynamic weighting and partial compensation mechanisms, it cannot completely eliminate the impact of data sparsity or data corruption. In real-world agricultural systems, sensor failures or incomplete satellite coverage are common, which can reduce the robustness of predictions. Despite achieving high average performance metrics, we observed several failure cases during error analysis. These failure cases typically occurred under extreme weather conditions, sudden ecological disturbances, or in areas with atypical land use patterns that were not representative of the training data. For example, rapidly urbanizing or deforestation areas pose challenges to graph-based components that rely on stable inter-node relationships. Similarly, when the temporal continuity of input data is disrupted, the temporal module sometimes fails to capture abrupt changes in risk. Interpretability remains a concern. Although ARFNs include structured components such as Graph Convolutional Networks (GCNs) and attention mechanisms that can partially reveal the importance of features, the entire model remains a complex black box. Future work should consider integrating more interpretable sub-modules or post-hoc interpretation techniques to enhance the model's credibility in high-risk decision-making environments.

5 Conclusions and future work

This study proposes a novel multi-source data fusion risk assessment method specifically tailored for the agricultural ecological environment, aiming to address key challenges in heterogeneous data integration and dynamic modeling of complex environments. Our proposed framework introduces two key innovations: the Agricultural Risk Fusion Network (ARFN), which integrates multimodal feature extraction, graph-based representation learning, and temporal modeling to achieve accurate ecological risk prediction; and the Adaptive Fusion Strategy (AFS), which dynamically adjusts data source weights based on context and temporal relevance, ensuring the model's robustness and adaptability across various agricultural scenarios. Experimental results on multiple benchmark datasets demonstrate that our method outperforms state-of-the-art methods in terms of accuracy, interpretability, and computational efficiency. Ablation experiments confirm the individual contributions of multimodal encoding, graph-based propagation, and temporal prediction, validating the effectiveness of the architecture design. Notably, the model effectively captures the multidimensional interactions between climate, soil, vegetation, and anthropogenic factors, providing a comprehensive understanding of ecological risk. Beyond performance improvements, the framework also enhances data integrity and context awareness, making it suitable for practical applications such as precision agriculture and environmental monitoring. These contributions have collectively driven the development of the field of agricultural ecological risk assessment, providing a scalable, interpretable, and adaptive modeling approach.

The risk assessment framework based on multi-source data fusion proposed in this paper has great potential for practical applications in agricultural management, environmental monitoring, and sustainable development planning. Utilizing multiple data modalities such as satellite imagery, soil measurement data, and meteorological information, the model can perform comprehensive ecological risk prediction, thereby supporting timely, data-driven decision-making in agro-ecosystems. One of the main advantages of this method is its adaptability: through an adaptive fusion strategy (AFS), the model can dynamically adapt to different environments and handle incomplete or noisy data sources, which is a common challenge in practical applications. Another advantage of the model is its scalability and modular design. Feature extraction, graph-based fusion, and temporal modeling of each component can be adapted to different datasets or integrated into existing monitoring processes with minimal structural reconstruction. This makes the method particularly important for precision agriculture platforms, government-level environmental monitoring systems, and climate risk early warning platforms. Some challenges remain to be overcome in practical applications. The model has high computational complexity, especially when processing high-frequency temporal data or large-scale spatial inputs, which may limit its feasibility in edge computing scenarios or regions with limited computing infrastructure. The reliance on high-quality multi-source data may pose integration challenges in underdeveloped regions with inconsistent data collection. Although the model incorporates some interpretable components, it still largely operates as a black box, which may reduce stakeholder trust in critical decision-making environments. To overcome these limitations, future work should focus on lightweight model compression techniques, integration with mobile sensing platforms, and the development of interpretable AI components suitable for environmental science. The proposed framework demonstrates good practical application potential and lays the foundation for building intelligent, sustainable, and adaptive ecological risk monitoring systems in agriculture and other fields.

References

• 1

  AlhamidM. F.RawashdehM.DongH.HossainM. A.SaddikA. E. (2016). Exploring latent preferences for context-aware personalized recommendation systems. IEEE Trans. Human-Mach. Syst. 46, 66–77. doi: 10.1109/THMS.2015.2509965

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  Cabrera-SánchezJ.-P.de LunaI. R.Carvajal-TrujilloE.Villarejo-RamoS. A. F. (2020). Online recommendation systems: factors influencing use in e-commerce. Sustainability12:8888. doi: 10.3390/su12218888

  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  ChaneyA.StewartB. M.EngelhardtB. (2017). “How algorithmic confounding in recommendation systems increases homogeneity and decreases utility.” in ACM Conference on Recommender Systems (New York, NY: Association for Computing Machinery), 224–232. doi: 10.1145/3240323.3240370

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  ChenZ.GanW.WuJ.HuK.LinH. (2025). Data scarcity in recommendation systems: a survey. Trans. Recomm. Syst.3, 1–31. doi: 10.1145/3639063

  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  CuiW.MeiY.LiuS.ZhangX. (2023). Health risk assessment of heavy metal pollution and its sources in agricultural soils near Hongfeng lake in the mining area of Guizhou province, China. Front. Public Health.11:1276925. doi: 10.3389/fpubh.2023.1276925

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  de OliveiraD.TadeiR.TriquesM. C.GoulartB.MontagnerC.EspíndolaE.et al. (2024). Fipronil risk assessment in tropical soils: relevant concentrations cause high risk to the edaphic community. Water, Air Soil Pollut. 235:123. doi: 10.1007/s11270-024-06925-9

  • CrossRef
  • Google Scholar

• 7

  DhelimS.AungN.BourasM. A.NingH.CambriaE. (2022). A survey on personality-aware recommendation systems. Artif. Intellig. Rev.55, 2409–2454. doi: 10.1007/s10462-021-10063-7

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  ElkahkyA.SongY.HeX. (2015). “A multi-view deep learning approach for cross domain user modeling in recommendation systems,” in The Web Conference (Geneva: International World Wide Web Conferences Steering Committee), 278–288. doi: 10.1145/2736277.2741667

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  FiroozshahianA.CoburnJ.LevensteinR.NattojiR.KamathA.WuO.et al. (2023). “MTIA: First generation silicon targeting meta's recommendation systems,” in International Symposium on Computer Architecture. (New York, NY: ACM), 1–13. doi: 10.1145/3579371.3589348

  • CrossRef
  • Google Scholar

• 10

  ForouzandehS.RostamiM.BerahmandK. (2022). A hybrid method for recommendation systems based on tourism with an evolutionary algorithm and topsis model. Fuzzy Inform. Eng. 14:3. doi: 10.1080/16168658.2021.2019430

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  FuZ.XianY.ZhangY.ZhangY. (2020). “Tutorial on conversational recommendation systems,” in ACM conference on recommender systems (New York, NY: Association for Computing Machinery), 751–753. doi: 10.1145/3383313.3411548

  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  GinartA. A.NaumovM.MudigereD.YangJ.ZouJ. Y. (2021). “Mixed dimension embeddings with application to memory-efficient recommendation systems,” in International Symposium on Information Theory (Melbourne: IEEE). doi: 10.1109/ISIT45174.2021.9517710

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  Huynh-LyT.-N.NguyenH.-H.Thai-NgheN. (2016). “Methods for building course recommendation systems,” in International Conference on Knowledge and Systems Engineering.

  • Google Scholar
  • View reference in article

• 14

  KhanB. M.ManshaA.KhanF.BashirS. (2017). “Collaborative filtering based online recommendation systems: a survey,” in International Congress on Information and Communication Technology (Karachi: IEEE). doi: 10.1109/ICICT.2017.8320176

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  KhanalS. S.PrasadP.AlsadoonA.MaagA. (2019). A systematic review: machine learning based recommendation systems for e-learning. Educ. Inform. Technol.: Off. J. IFIP Tech. Committee Educ. 25, 2635–2664. doi: 10.1007/s10639-019-10063-9

  • CrossRef
  • Google Scholar

• 16

  KimM. C.ChenC. (2015). A scientometric review of emerging trends and new developments in recommendation systems. Scientometrics104, 239–263. doi: 10.1007/s11192-015-1595-5

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  KreutzC. K.SchenkelR. (2022). Scientific paper recommendation systems: a literature review of recent publications. Int. J. Digit. Libraries23, 335–369. doi: 10.1007/s00799-022-00339-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  PeakeG.WangJ. (2018). “Explanation mining: post hoc interpretability of latent factor models for recommendation systems,” in KDD '18: proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining (New York, NY: Association for Computing Machinery), 2060–2069. doi: 10.1145/3219819.3220072

  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  RamlatchanA.YangM.LIUQ.LiM.WangJ.LiY. (2018). A survey of matrix completion methods for recommendation systems. Big Data Mining Anal.1, 308–323. doi: 10.26599/BDMA.2018.9020008

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  SchedlM. (2019). Deep learning in music recommendation systems. Front. Appl. Mathem. Statist.5:44. doi: 10.3389/fams.2019.00044

  • CrossRef
  • Google Scholar

• 21

  SharmaA.HofmanJ. M.WattsD. (2015). “Estimating the causal impact of recommendation systems from observational data,” in ACM Conference on Economics and Computation (New York, NY: Association for Computing Machinery), 453–470. doi: 10.1145/2764468.2764488

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  ShiH.MudigereD.NaumovM.YangJ. (2020). “Compositional embeddings using complementary partitions for memory-efficient recommendation systems,” in KDD '20 (Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining) (New York, NY: Association for Computing Machinery), 165–175. doi: 10.1145/3394486.3403059

  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  ShuaiC.PangC. (2025). Comprehensive assessment of urban ecological environment based on multisource data integration analysis. Ecol. Chem. Eng. 32, 501–514. doi: 10.2478/eces-2025-0026

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  SonJ.KimS. B. (2017). Content-based filtering for recommendation systems using multiattribute networks. Expert Syst. Appl.89, 404–412. doi: 10.1016/j.eswa.2017.08.008

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  SukarjoH. C. O Zu'amahH. (2022). Ecological risk assessment and spatial distribution of heavy metals of agricultural soils in Nganjuk Regency, Indonesia. IOP Conf. Series: Earth Environm. 1114:012022. doi: 10.1088/1755-1315/1114/1/012022

  • CrossRef
  • Google Scholar

• 26

  SunZ.XuY.LiuY.HeW.JiangY.WuF.et al. (2022). A survey on federated recommendation systems. IEEE Trans. Neural Netw. Learn. Syst. 36, 6–20. doi: 10.1109/TNNLS.2024.3354924

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  TangE. (2019). “A quantum-inspired classical algorithm for recommendation systems,” in Electronic Colloquium on Computational Complexity (New York, NY: Association for Computing Machinery), 217–228. doi: 10.1145/3313276.3316310

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  TrichopoulosG. C.KonstantakisM.AlexandridisG.CaridakisG. (2023). Large language models as recommendation systems in museums. Electronics12:3829. doi: 10.3390/electronics12183829

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  VartakM.HuangS.SiddiquiT.MaddenS.ParameswaranA. G. (2016). Towards visualization recommendation systems. SGMD. 45, 34–39. doi: 10.1145/3092931.3092937

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  WangQ.GaoM.ZhangH. (2022). Agroecological efficiency evaluation based on multi-source remote sensing data in a typical county of the Tibetan Plateau. Land11:561. doi: 10.3390/land11040561

  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  WangS.ZhengZ.WuZ.LyuM. R.YangF. (2015). Reputation measurement and malicious feedback rating prevention in web service recommendation systems. IEEE Trans. Serv. Comp.8, 755–767. doi: 10.1109/TSC.2014.2320262

  • CrossRef
  • Google Scholar

• 32

  WeiG.WuQ.ZhouM. (2021). A hybrid probabilistic multiobjective evolutionary algorithm for commercial recommendation systems. IEEE Trans. Comp. Soc. Syst. 8, 1163–1171. doi: 10.1109/TCSS.2021.3055823

  • CrossRef
  • Google Scholar

• 33

  WuJ.-G.ZhaoJ.-P.XuX.-J.LiuY.HuL.-X.WenZ.et al. (2025). Pollution characteristics and ecological risk assessment of biocides in surface water of Guangdong province. Huan Jing Ke Xue.46, 6296–6305. doi: 10.13227/j.hjkx.202410034

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 34

  YangB.HanC.LiY.ZuoL.YuZ. (2022). “Improving conversational recommendation systems' quality with context-aware item meta information,” in NAACL-HLT (Seattle, WA: Association for Computational Linguistics), 38–48. doi: 10.18653/v1/2022.findings-naacl.4

  • CrossRef
  • Google Scholar

• 35

  YangL.TanB.ZhengV.ChenK.YangQ. (2020). “Federated recommendation systems,” Federated Learning: Privacy and Incentive, eds. Q. Yang, L. Fan, and H. Yu, (Cham: Springer), 225–239. doi: 10.1007/978-3-030-63076-8_16

  • CrossRef
  • Google Scholar

• 36

  ZhangZ.PatraB. G.YaseenA.ZhuJ.SabharwalR.RobertsK.et al. (2023). Scholarly recommendation systems: a literature survey. Knowl. Inform. Syst.65, 4433–4478. doi: 10.1007/s10115-023-01901-x

  • CrossRef
  • Google Scholar

• 37

  ZhaoF.GuoY. (2017). “Learning discriminative recommendation systems with side information,” in Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17) (Melbourne: International Joint Conferences on Artificial Intelligence), 3469–3475.

  • Google Scholar

• 38

  ZhaoJ.LiuL.WangY.TangK.HuoM.ZhaoY. (2023). Evaluation of sustainable development of the urban ecological environment and its coupling relationship with human activities based on multi-source data. Sustainability15:4340. doi: 10.3390/su15054340

  • CrossRef
  • Google Scholar