Sustainable field-scale detection of walnut pests and diseases using deep transfer learning under changing-climatic conditions

Abstract

Introduction:

Pests and diseases are major problems in horticulture, intensifying synergistic interactions between pests and host pathogens, increase management costs, and posing severe economic risk. Automatic pest and disease detection is essential to minimize pathogen intensification, yield losses, and prevent spread to adjacenttrees.

Methods:

This study presents a sustainable field-scale detection of walnut pests and diseases using deep transfer learning approach. Containing 11,221 original walnut pest images representing 17 distinct classes, and 2,850 original diseases images across three classes, developed from farmers’ fields in and around the Kupwara district of Jammu and Kashmir, India, reflecting real-world field variability in illumination, background clutter, orientation, scale, occlusion, and image acquisition conditions. Preprocessing steps, including high-resolution image selection,data cleaning to remove duplicate and near duplicate images, stratified data split before augmentation is performed to enhance feature fidelity, improve model generalization, mitigate overfitting, and ensure robust model performance. Among the rigorously evaluated advanced deep transfer learning models InceptionV3, ResNet50, and ViTs model including self-supervised learning DINOv2, and contrastive learning MoCov3 across accuracy, precision, recall, F1-score, ROC-AUC and confusion matrix, and compared with current existing studies.

Results and Discussion:

The proposed customized InceptionV3 achieves superior performance for walnut pest and disease detection, attaining training accuracy of 99.87%, validation accuracy of 98.62%, testing accuracy of 98.05%, precision, recall, and f1-score of 98.07% for walnut disease detection the proposed model achieved training accuracy of 99.77%, validation accuracy of 98.68%, testing accuracy of 98.25%, with strong precision, recall, and f1-score of 98.67%. The proposed model can be integrated and deployed in real-world agricultural applications embedding AIoT for resource constraint devices for integrated pest and disease management (IPDM) in smart horticulture. This study also significantly contributes to advancing the UN’s Sustainable Development Goals (SDGs), notably SDG 1: No Poverty, SDG 2: Zero Hunger, SDG 3: Good Health and Well-being, SDG 12:Responsible Consumption and Production, SDG 13: Climate Action, and SDG 15:Life on Land by enabling early detection of walnut pests and diseases, supporting timely, precise, and resource-efficient interventions.

1 Introduction

Pests and diseases are major biotic stress in horticulture production, threatening the life and livelihood of millions of small-scale farmers, global food security, agricultural sustenance, and development under rapidly changing climatic conditions, exacerbating and transmitting pathogens and pest pressure, altering tree physiological and metabolic processes, posing significant harvest strain, and economic burden (Ojiambo, 2025; Yeshiwas et al., 2026). FAO reports that over the past 60 years (1960 to 2020), estimated pest and disease invasion costs $4.52 trillion USD. Annually, India accounts for approximately $18.5 billion USD (Rajan and Singh, 2025). The Indian Council of Agricultural Research (ICAR) estimated ₹90,000 crores in annual yield losses in horticultural production due to severe pest and disease outbreaks, which wiped out 35-40% of total yield, costing approximately 2 billion USD (Kumar et al., 2026). Globally, India is the second-largest producer of horticulture crops (fruits and vegetables) after China, with a total annual production of around 367.472 million tons (Yadav et al., 2026). Horticulture contributes 33% to India’s agricultural GVA and uses only 13% of the gross cropped area (Ashraf et al., 2026). Walnut (Juglans regia L.) is a vital temperate nut crop in India, primarily cultivated in Jammu and Kashmir (responsible for ~95–98% of production), followed by Himachal Pradesh, Uttarakhand, and Arunachal Pradesh (Yaseen et al., 2025). Covering over 100,000 hectares of crop land (Sharma et al., 2025b). In recent years, walnut production in India has severely declined from 3,292 MT in 2015–2016 to just 638.07 MT in 2023–2024, owing to climate change (Khursheed and Taufiq, 2020). Continuously extended to reduce walnut productivity at an annual growth rate of 4.5%, with productivity gaps between 20% to 40% or more compared to global production, resulting in an estimated annual loss of 80–150 million USD (Liu et al., 2026). At the regional level, Jammu and Kashmir experience even higher vulnerability. Pest species infest leaves, sensitive stems, blooms, emerging buds, and developing fruits, causing antagonistic relationships between a host organism and pathogen (bacteria, virus, fungus, or parasite) (Madrid et al., 2026; Saxena et al., 2026). These interactions define how pathogens invade, survive, and cause disease. Fungal diseases alone can cause 71.8–97.3% walnut productivity losses, while bacterial infections lead to 50% fruit drop in the pre-harvesting process (Mir et al., 2025). Phytosanitary threats in walnut production are a double-edged challenge: they not only compromise yield, food security, and farmer income but also intensify pesticide use, contributing to environmental pollution and climatic risk (Shukla et al., 2025). Disease management in walnut orchards largely relies on traditional practices that are costly, labor-intensive, and prone to human error, often involving extensive pesticide application (Singh et al., 2026). Meeting the food demands of a rapidly growing global population is projected to require approximately a 70% increase in agricultural and horticultural production (De Mastro et al., 2025). However, factors such as urbanization and climate change, particularly in India, limit horticultural expansion, increasing the need for sustainable and efficient crop management strategies to support horticultural production. Currently, the annual cost of pesticides and chemical treatments in walnut orchards in India typically ranges from USD 150 to over USD 300 per acre for mature in conventionally managed systems (Sharma et al., 2026). More than 98% of pesticides sprayed in walnut orchards fail to reach their intended targets thus contaminating the environment. Such pesticide contamination and environmental exposure directly or indirectly endanger habitat diversity, disturb ecological balance, and harm human life through cultivation practices, ingestion, or inhalation via food, water, or air, highlighting the need for intelligent pest and disease detection in a sustainable way using images processing and resource-effective deep transfer learning approach. The broader consequences of the Pest–pathogen complexes and pesticide overuse in horticulture production are summarized in Figure 1.

Figure 1

In recent years, promising cutting-edge modern technologies such as precision monitoring, intelligent pest and disease detection, and decision-support systems through artificial intelligence, especially deep learning and image processing, have transformed horticulture and applied fruit science, and can be applied for precision prevention and early detection of active walnut pests and diseases as a solution for sustainable walnut production.

This study is motivated by the pressing need to address walnut pest–pathogen stress to minimize economic losses, improve farmer livelihoods, fruit quality, productivity, and marketability while minimizing need of manual inspection and labor dependency; mitigate subjective monitoring; and prevent the early spread of pathogens in neighboring tree to minimizing pesticide applications, promoting climate-resilient horticulture aligning with sustainable development goals (SDGs) integrating expert knowledge, advanced field level image analysis, and deep learning techniques leveraging transfer learning approach. Early prevention and timely intervention of pests and diseases in horticulture significantly reduces excessive pesticide exposure (Manono et al., 2026). Such approaches are cost-effective, environmentally sustainable, and economically viable, thereby enhancing horticultural productivity, strengthening food security, and promoting inclusive growth while advancing sustainable horticulture and pomological research (Sharma et al., 2025a).

The workflow of this study is presented in Figure 2, which includes independent walnut pest and disease image datasets developed from the farmer’s field using a high-resolution digital camera annotated by expert plant pathologists and epidemiologists team and leveraging advanced preprocessing step and standardization. Among the trained and tested models, InceptionV3 performs far better than the existing approaches, trained on the training set to learn complex image patterns and underlying structures to monitor training performance, validated on the validation set to model generalization, and hyperparameter optimization to provide an objective performance, rather than simply performing well in the learning process and rigorously tested on unseen testing set to measure accuracy, precision, recall, f1-score, roc-auc and confusion rate and validated using stratified k fold validation, grad-cam analysis and verified by fields experts (Ahad et al., 2026). The suggested model can perform inference and predict unseen images for pest and disease inputs separately, has diverse applications in smart horticulture, and is applied to fruit science under complex environmental and illumination settings.

Figure 2

1.1 Problem statement

Pest and disease complexes are the most significant biotic constraints to horticultural productivity under changing climate, causing substantial economic losses and threatening global food security (Mir et al., 2025; Sharma et al., 2026; Singh et al., 2026). According to the Food and Agriculture Organization (AFO), pest and disease invasions caused an estimated USD 4.52 trillion in global losses between 1960 and 2020, India alone incurs approximately USD 18.5 billion annually due to phytosanitary damage (Ojiambo, 2025; Kumar et al., 2026; Yeshiwas et al., 2026). The Indian Council of Agricultural Research also reports that 35–40% of horticultural production in India is lost each year because of pest and disease stress (Yadav et al., 2026). Among temperate nut crops, walnut (Juglans regia L.) production has shown a significant decline, particularly in Jammu and Kashmir, which contributes nearly 95–98% of India’s walnut output. Recent reports indicate a sharp decline in walnut production from 3,292 MT in 2015–2016 to 638.07 MT in 2023–2025, largely attributed to severe pest infestations and pathogenic diseases (Hassan and Ahmad, 2017; Karadeniz et al., 2023). Fungal infections alone can cause productivity losses of 71.8–97.3%, while bacterial diseases may lead to up to 50% fruit drop, resulting in estimated annual economic losses of USD 80–150 million (Khursheed and Taufique, 2020). Currently, the annual cost of pesticides and chemical treatments in walnut orchards in India typically ranges from USD 150 to over USD 300 per acre for mature, conventionally managed systems (Liu et al., 2026). More than 98% of pesticides sprayed in walnut production fail to reach their intended targets and contaminate the environment. Such pesticide contamination and exposure directly or indirectly endanger habitat diversity, disturb ecological balance, and harm human life through cultivation practices, ingestion, or inhalation via food, water, air, or dust (Saxena et al., 2026). These alarming statistics underscore the urgent need for sustainable field-Scale detection of walnut pests and diseases to support sustainable walnut production and minimize economic and environmental impacts.

1.2 Study contribution

The contributions of this study are briefly described as follows:

• This study proposed a robust deep learning–based method leveraging the transfer learning technique for field-level pests and diseases detection in walnut, facilitating timely data-driven prevention and early intervention for farmers.

• The novel, large-scale, high-resolution walnut pest and disease image datasets are developed from farmer’s fields containing 17 diverse pest classes and 3 disease classes, supporting in-depth analysis and improved predictive model generalization, having diverse applications in horticulture research and development.

• Development of a precise and intelligent walnut pest and disease detection model as a solution to support walnut farmers in resource-constrained and remote horticultural regions.

• Transfer learning with fine-tuning is leveraged for model optimization and knowledge transfer, specifically for walnut pests and diseases detection, to improve detection accuracy.

• Regularization and generalization techniques incorporating dense layers, dropout, and global average pooling layers were implemented to enhance the proposed model’s performance in the real-world environment.

• The proposed model is rigorously evaluated and compared with existing approaches, demonstrating improved performance, robustness, and strong potential for real-world field applications and deployment under varying environmental and illumination settings.

The remainder of this study is structured as follows: Section 2 delineate related work and recent advancement in pest and disease detection; section 3 elucidates the materials and methods used in this study; Section 4 delineates results and discussion; validate this study using k fold cross validation and compared with state of art methods and verified through grad-cam analysis and expert plant pathologist and entomologist team suggesting recommendations for sustainable horticulture and fruit productivity highlighting study key contribution to SDGs 2030, Improvement and future work; Section 5 bridge the conclusion.

2 Related work

The related pest and disease detection work in diverse horticulture crops explored promising cutting-edge computational techniques, including Artificial Intelligence (AI), Deep Learning, Transfer Learning, VITs, IoT, and Integrated Edge Computing, highlighting their broader applications in pest and disease detection for sustainable horticulture (Gunasekaran et al., 2026; Qwaid et al., 2026; Xu et al., 2026). Cucumber disease was detected using advanced image processing and convolutional neural networks; VGG-16 and VGG-M attained 91.08% accuracy (Mia et al., 2021). Chili pest and disease detection compares and evaluates classical and deep learning approaches. Six classical and deep learning models were tested. Among them, ConvNeXt demonstrates superior performance to existing methods, achieving an accuracy of 92.10% (Ahmad Loti et al., 2021). Cashew disease was detected using DenseNet-201, leveraging the transfer learning technique, scoring the highest accuracy of 95.75% (Sudha and Kumaran, 2023).The Mamba-YOLO-ML model proposed for mulberry disease detection incorporates SCFEM, DSCFEM, and SPPM modules, scoring an accuracy of 93.63% (Yuan et al., 2025). Seven fine-tuned transfer learning models were trained on 9,000 images representing common solanaceous foliar diseases. The custom Solanaceous-Net CNN model achieves an identification accuracy of 88.00% (Qin et al., 2025). VG-GNBNet deep learning model developed for areca nut disease recognition, integrating a two-stage feature extraction module to extract spatial features, and matrix factorization to improve identification accuracy (Hegde and Kumar, 2025). A CNN model using transfer learning was proposed for citrus disease identification, leveraging RPN and pre-trained CoAtNet, which attained testing accuracy of 94.37% and an average precision of 93.8% (Sankaran et al., 2024). Apricot pest and disease detection is accomplished by the APNet model, incorporating global average pooling, a dense layer of 256, with a 50% dropout rate, which greatly reduces the model’s complexity, avoids overfitting, and improves performance, attaining a 95.97% success rate (Li et al., 2025). To detect mango pests and diseases, a YOLOv8 model was developed, incorporating the ATA module, with a score accuracy of 92.89% (Wang and Wang, 2024). The VGG-19 model proposed for apple disease management was trained on image samples collected from controlled and open environmental conditions and tested in a vineyard, farmyard, and farmland. Evaluations showed a significant reduction in open-set accuracy from 66.3% to 42.1% on a closed set to 34.8% open set (Rohith et al., 2025). Banana, fig, and potato leaf disease were detected using VGG19 with the highest accuracy of 93.81% (Shafik et al., 2024). The EPDNet model suggested for tomato leaf disease detection, adding trainable dense layers and a dropout mechanism, achieved an accuracy of 94.01% (Wu et al., 2026). Related pest and disease identification proposes resource-efficient deep learning utilizing knowledge transfer methodologies, thus enabling early and precise pest and disease detection in multiple crops, improving real-time performance. These novel approaches provide actionable and sustainable solutions that significantly enhance IPM and align with the objectives of the proposed study for walnut pest and disease detection for sustainable pomology.

3 Materials and methods

3.1 Walnut pest and disease dataset development

The diverse walnut pests and diseases datasets were developed from farmer’s fields in and around the Kupwara province of Jammu and Kashmir, India, from diverse tree/orchard/location. The spatial extent of study region is presented in Figure 3 and different orchards where on-site walnut pest and disease images were collected with the help of expert pathologists and epidemiologists, are illustrated in Figures 4, 5. A substantial collection of entomological and pathological specimens was collected during two consecutive summer seasons (2024 and 2025). High-resolution images were scrupulously snapped with the iPhone 13 Pro digital camera under natural field settings without artificial lighting. The camera specification used for walnut pests and diseases image datasets acquisition is presented in Table 1. The acquired walnut pest dataset contains high-resolution images of 17 diverse pest categories presented in Table 2, and the diseases dataset contains 3 categories presented in Table 3, verified and validated by expert plant pathologists and entomologists. The walnut pest and disease detection dataset applications in integrated domains are presented in Table 4 for advancing agriculture and horticulture research and development. The examples of the walnut pest and disease dataset sample and categories are presented in Figures 6, 7, demonstrating representative examples of pest and disease presence captured from farmer’s fields.

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

3.2 Walnut pest and disease dataset pre-processing

Imperfections, missing visual cues, and repetition in optical patterns are common issues in field collected images dataset that can undermine model generalization and performance. Preprocessing is an important step in agricultural image analysis, transforms raw, noisy sensor data into high-quality, actionable, consistent images that reduce noise from environmental factors such as lighting and geometric distortions, enabling precise, reliable, and automated pest and disease detection (Simhadri et al., 2025). To mitigate the risk of data leakage, the duplicate and near-duplicate image screening process was carefully conducted before data splitting (Khan et al., 2025). Visually identical images and highly similar samples captured from the same frame sequence, tree, location, or orchard environment were removed through manual expert inspection and image similarity checks (Devarajan et al., 2026). This ensured that no duplicated or near-duplicated images appeared across the training, validation, and testing sets. Only relevant images were retained in the walnut pest and disease datasets and subsequently downsampled to 224×224 pixels and scaled to the 0–1 range. Furthermore, dataset splitting was performed before augmentation to avoid potential data leakage (Sambana et al., 2025). Walnut pest and disease datasets were stratified split by class, training (80%), validation (10%), and testing (10%) subsets, ensuring that the class distribution remained consistent across all splits. Data augmentation was applied to the training subset only to increase the diversity of training samples, meticulously using rescaling, rotation, shifting, shearing, zooming, reversal, and fill mode operations, presented in Table 5, resulting in approximately 1000 augmented training samples per class in both pest and disease training sets, presented in Tables 3 and 2. The validation and testing sets of pest and disease datasets contained non-augmented images. For better understanding, the preprocessing workflow is diligently presented in Algorithm 1. The number of original images in separate pest and disease datasets per class, training set before augmentation, number of images in augmented training set, training set after augmentation, validation set, and testing set samples without augmentation are presented in Table 3 for the disease dataset and Table 2 for the walnut pest dataset. The image preprocessing steps significantly enhance dataset diversity, mitigate class imbalance, prevent data leakage, overfitting, and improve walnut pest and disease detection accuracy and generalization to complex visual features (Kaur et al., 2025; Mamun et al., 2026).

Algorithm 1

3.3 Proposed walnut pest and disease detection method

The walnut pest and disease detection pipeline is presented in Figure 2. Image preprocessing, augmentation, inceptionV3, and ResNet50 model development, leveraging transfer learning, fine-tuning processes, and rigorous performance evaluation, are core components of this section, utilized for early walnut pest and disease detection, early intervention, and management practices to empower farmers and other stakeholders in sustainable horticulture. The proposed AI-powered smart walnut pest and disease detection method significantly prevents yield loss, increases fruit productivity, sustains food security, and timely prevents pest and disease pressure in the economically important walnut crop.

3.3.1 Development of deep transfer learning InceptionV3, ResNet50 model for walnut pest and disease detection

The proposed InceptionV3 and ResNet50 models were developed precisely for early walnut pest and disease detection. The key steps of the developed process include careful model selection, optimization, performance evaluation, and benchmarking with state-of-the-art methods. The proposed architectures have shown excellent feature extraction, representation learning abilities, and precision in agriculture image detection tasks (Anwar et al., 2026). Key hyperparameters of the proposed InceptionV3 and ResNet 50 models were separately optimized for pest and disease detection, leveraging the advanced knowledge transfer approach to extract and repurpose learned features for the novel walnut pest and disease image dataset (Li et al., 2026). This optimized approach significantly reduces computational complexity and training time compared to developing models from scratch and improves model generalization and performance. The custom layer design incorporates a global average pooling 2D layer, dense layer, and dropout mechanism, as presented in Figure 8. The key parameters, optimization configuration and search space or values for InceptionV3 and ResNet50 models independently for pest and disease detection are presented in Algorithm 2 and Table 6.

Figure 8

3.3.2 Development of deep transfer learning InceptionV3 and ResNet50 model for walnut pest and disease detection

The InceptionV3 model is developed by integrating task-specific layers. The input samples were downsampled to 224 × 224 px. The pretrained model’s ImageNet weights remained static to preserve effective generic features (Inamdar et al., 2026). The model architecture is enhanced by adding a SoftMax output layer, a ReLU-activated 256 dense layer, global average pooling, and a 50% dropout rate to mitigate overfitting. The customized InceptionV3 model architecture for walnut pest and disease detection is presented in Figure 9. The ResNet50 model is enhanced by adding a new classification head, the convolutional base frozen to preserve its generalized feature representations. The appended top head uses a global average pooling layer, a fully connected layer of 512 ReLU-activated neurons, 25% dropout rate to mitigate overfitting, and finally a task-specific dense layer with SoftMax transfer function for pest and disease detection. The customized ResNet50 architecture for walnut pest and disease detection is presented in Figure 10.

Figure 9

Figure 10

3.3.3 Development of self-supervised learning DINOv2, and contrastive learning MoCov3 ViTs model for walnut pest and disease detection

The DINOv2 and MoCo v3 Vision Transformer (ViT) models were developed to evaluate the possibility of self-supervised representation learning for the early detection and classification of walnut pests and diseases. Self-supervised pretraining, transfer learning-based fine-tuning, hyperparameter optimization, performance analysis, and comparative benchmarking against conventional convolutional neural network (CNN) architectures were all facets of the model development methodology. DINOv2 leverages a self-distillation method to learn rich and semantically relevant visual representations without requiring user annotations during the pretraining stage (Li et al., 2026). The architecture of proposed DINOv2 model is presented in Figure 11. In contrast, the self-supervised MoCo v3 ViT model utilizes a momentum contrastive learning approach designed specifically for Vision Transformers, which enables the model to train discriminative feature representations by maximizing matching between positive imagery pairs and decreasing similarity with negative samples (Devarajan et al., 2026). These self-supervised learning techniques facilitate the extraction of dependable and transferable features that are highly relevant for walnut pest and disease detection and classification in smart horticulture (Sankaran et al., 2024). Transfer learning was used to exploit information from large-scale imagery datasets and repurpose it for domain-specific feature extraction in order to adapt the pretrained models to the walnut pest and disease classification problem (Wu et al., 2026). The architecture of proposed MoCov3 model is presented in Figure 12. This method improved model generalization and classification performance while substantially reducing computational costs and training time needed (Yuan et al., 2025). Therefore, to enhance unbiased learning and reduce overfitting, task-specific classification layers consisting of a Global Average Pooling (GAP) layer, fully linked dense layers, and dropout regularization were incorporated into the proposed backbones models presented in Algorithm 2. A thorough evaluation of the efficacy of self-supervised Vision Transformers for walnut pest and disease identification is provided by comparing DINOv2 and MoCo v3 with the CNN-based InceptionV3 and ResNet50 models.

Figure 11

Figure 12

Input: Walnut pest dataset ((D_p)), walnut disease dataset ((D_d)), candidate models ((M= InceptionV3, ResNet50, DINOv2, MoCo V3)Output: Optimal model ((M_{best})) for walnut pest and disease detection1. Acquire and annotate two independent datasets: ● Walnut Pest Dataset ((D_p)) ● Walnut Disease Dataset ((D_d))2. Perform image quality assessment and remove duplicate, corrupted, and low-quality samples.3. Preprocess all images through resizing, normalization, and class-label encoding.4. Apply data augmentation to training set only to increase sample diversity and improve model generalization.5. Initialize the experimental environment and fix the random seed for reproducibility.6. For each dataset (D/in {D_p, D_d}) do6.1 Partition (D) into training (80%), validation (10%), and testing (10%) subsets using stratified sampling technique.6.2 Perform stratified 5-fold cross-validation on the training subset.6.3 For each model (M_i/in M) doa. Load pretrained backbone weights.b. Apply transfer learning by freezing low-level layers and fine-tuning high-level layers.c. Construct the classification head consisting of:● Global Average Pooling layer,● Dense layer,● Dropout layer,● Softmax output layer.d. Perform hyperparameter optimization using grid search.e. Train the model using the training dataset.f. Monitor validation performance and save the best-performing model checkpoint.g. Evaluate the optimized model using:● Accuracy, Precision, Recall, F1-score, ROC-AUC, confusion matrix.h. Generate the confusion matrix to analyze class-wise prediction performance.6.4 End For6.5 Evaluate all optimized models on the independent testing dataset.6.6 Perform comparative benchmarking of InceptionV3, ResNet50, DINOv2, and MoCo v3 based on Accuracy, Precision, Recall, F1-score, ROC-AUC, and confusion matrix analysis.6.7 Conduct Grad-CAM analysis to identify discriminative image regions contributing to model predictions.6.8 Rank all candidate models according to their overall predictive performance and generalization capability.6.9 Select the highest-performing model as (M_{best}).7. End For8. Compare the optimal models obtained for walnut pest and walnut disease detection.9. Identify InceptionV3 as the final model based on superior Accuracy, Precision, Recall, F1-score, ROC-AUC, confusion matrix performance, and interpretability analysis.10. Report the final classification results, and Grad-CAM visualizations.Return: (M_{best} = InceptionV3)

3.3.4 Hyperparameter tuning

Optimal parameter tuning is an important step in deep learning pest and disease detection approaches to achieve optimized performance (Li et al., 2026). In this study, the most important hyperparameters were carefully selected, refined, and fine-tuned to optimize and enhance the accuracy and robustness of pest and disease detection, mitigating overfitting issues and maintaining consistent performance during model testing (Ahmed et al., 2025). The proposed deep transfer learning and self-supervised learning ViTs model training configuration and hyperparameter selection strategy are presented in Table 6. The experimental process begins with the specification of the deep learning framework, hardware configuration, and experimental tasks. A random seed is initialized to ensure reproducibility. Subsequently, the data set is organized using an appropriate partitioning and cross-validation strategy. The model architecture is defined by selecting the backbone network, weight initialization approach, and the fine-tuning strategy for pretrained layers. In addition, dense layers and activation functions are configured to construct the final classification network. To enhance generalization and mitigate overfitting, regularization techniques such as dropout and L2 weight regularization are incorporated. The training configuration is established by defining the loss function, optimizer, learning rate, learning rate scheduler, batch size, and number of epochs. Furthermore, hyperparameters are optimized using a dedicated optimization strategy across multiple experimental trials. Finally, the best-performing model is selected based on an appropriate model selection metric, and its performance is assessed using standard evaluation metrics.

3.3.5 Experimental setup

All experiments were conducted on a high-performance computing NVIDIA GTX 1650 GPU, 64-bit Core i5 processor, 12 GB of RAM, 1 TB SSD, and 2 TB HDD system. This configuration provided efficient computing resources and ample storage for the precise and smooth execution of experiments. The specific hardware and software configuration used in experiments is presented in Table 7.

3.3.6 Performance evaluation metrics

Accuracy: Accurate predictions among all the predictions the detection model performs are measured as accuracy.

Precision: It measures the ratio of TP to all positive predictions made by the detection model.

Recall/Sensitivity: It measures the ratio of TP to all actual positives, including TP and FN made by the detection model.

F1 Score: The harmonic-mean of the two measures effectively fuses precision and recall into a single useful metric that can quantify a balanced and clear measure of detection performance.

Confusion Matrix: Visualizes the model’s performance by showing correct vs. incorrect predictions, detecting where the model gets confused. InceptionV3 and ResNet50 model performance is evaluated through the confusion matrix of nxn rows and columns, where rows denote actual values and columns denote predicted values (Li et al., 2025) as provided in Table 8.

4 Results and discussion

The results and findings of the proposed study for walnut pest and disease detection are discussed in this section. InceptionV3 achieves 99.87% training accuracy and 98.05% testing accuracy for walnut pest detection. The model also successfully detects walnut diseases, achieving training accuracy of 99.77% and testing accuracy of 98.25%, strongly demonstrating InceptionV3’s pest and disease detection ability and its potential applications for timely crop management in smart horticulture, improving on compared and the existing study. The model’s superior performance is owing to its sophisticated architectural design and integration of compound scalability and multi-stage methods optimization. The structural customization improves the analysis of subtle, fine-grained patterns and model generalization. Furthermore, the transfer-learning approach used in this study, leveraging pre-trained weights from a large-scale ImageNet dataset, further improves model convergence (Verma et al., 2026). The learning and validation curve continuously improves, and the loss curve drops across epochs, showing effective feature learning, consistent model stability, and improved accuracy. The training and validation accuracy and loss curve and overall performance of InceptionV3, ResNet50, DINOv2, MoCov3 for walnut pest and disease detection are presented in Tables 9, 10, Figures 13, 14. The walnut pest and disease detection model overall precision, recall and f1 score measure presented in Tables 11, 12 and class specific performance across precision (P), recall (R), f1 score (F1) and ROC-AUC curve (AUC) evaluations is present in Tables 13, 14. From each of these performance measures, the InceptionV3 model outperformed the ResNet50. DINOv2 and MoCov3 models for pest and disease detection.

Figure 13

Figure 14

The confusion matrix is an important performance evaluation and visualization tool. It measures class-wise detection accuracy by mapping the relationship between actual and predicted class labels. Provides a rigorous analysis to correctly identify instances as TP or TN, and false instances as FP or FN. This evaluation measures the distribution of errors and class-wise model predictive behavior. Confusion matrices evaluation of the proposed InceptionV3, ResNet50, DINOv2, MoCov3 for walnut pest detection are presented in Figures 15–18 and for disease detection confusion matrix evaluation of the proposed models is presented in Figure 19.

Figure 15

Figure 16

Figure 17

Figure 18

Figure 19

4.1 Proposed model robust validation using stratified k-fold cross dataset split technique

K-fold cross-validation is an effective resampling technique for evaluating the performance of deep learning models. It involves dividing the dataset into equal-sized subsets, known as folds. The deep learning model is trained multiple times, where the number of folds determines the number of training iterations, and in each iteration, one-fold is used for validation while the remaining folds are used for training (Gupta et al., 2026). Initially, the complete pest and disease datasets were partitioned using a stratified approach to prevent data leakage, with 80% allocated for training, 10% for validation, and the remaining 10% reserved as an independent test set for performance evaluation. The workflow of the k-fold cross-validation is presented in Figure 20. After computing all cross-validation iterations, the evaluation metrics obtained from each fold were aggregated to evaluate the mean performance and variability across k-folds (Lilhore et al., 2026). The mean accuracy across the folds is computed using Equation 5, and the standard deviation (SD) is computed to measure the model’s flexibility across folds using Equation 6.

Figure 20

The experimental 5 fold-validation results of the proposed InceptionV3, ResNet50, DINOv2, MoCov3 for walnut pest detection from Tables 15–22 highlight that the proposed InceptionV3 model is precise and robust for accurate pest and disease detection in precision horticulture. Owing to its strong detection performance and stable generalization across cross-validation folds, the model has significant potential for deployment in smartphone applications, enabling early detection and timely intervention for crop protection.

4.2 Grad-CAM heatmap analysis to mitigate Clever Hans effect

To increase the transparency and interpretability of the proposed best performing InceptionV3 model, Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to highlight the image regions that significantly influenced the model’s detection decisions to address the potential risk of the Clever Hans effect rather than detect the actual visual characteristics of walnut pest and disease samples. As observed in Figures 21, 22, the resulting Grad-CAM heatmaps are mostly concentrated on biologically significant symptomatic regions, comprising insect sites, feeding damage, lesion boundaries, necrotic tissues, chlorotic patches, and other disease-specific symptoms. The substantial localization of activations across affected image regions indicates that the model has successfully learned discriminative features that strongly correlate with pest and disease symptoms, visualizations revealed negligible activation in irrelevant background areas, including soil surfaces, surrounding vegetation, orchard scenery, and lighting-related artifacts. The analysis reveals that the proposed technique does not primarily rely on contextual information for prognosis, which greatly allays worries about background bias and deceptive correlations. The consistency of attention maps over a range of test samples further demonstrates the robustness and generalizability of the learnt feature representations under various imaging conditions. Importantly, the highlighted areas closely correspond to symptom areas that specialists in agriculture would consider diagnostically relevant, indicating that the model’s predictions are based on important entomological and pathological characteristics.

Figure 21

Figure 22

Expert plant pathologists and entomologists independently scrutinized the Grad-CAM visualizations to validate that the model consistently addressed disease lesions, pest regions, and distinctive symptomatic patterns pertinent to the evaluation of walnut pest and disease detection. The proposed InceptionV3 model learns significant biological properties rather than erroneous background correlations, as evidenced by a strong correlation observed between the highlighted activation regions and expert knowledge. As a result, the Grad-CAM analysis offers compelling proof of the model’s biological validity, interpretability, and reliability while significantly lowering worries about the Clever Hans effect and boosting confidence in its useful application for farmer and other stockholders.

4.3 Comparative evaluation of the proposed method with existing studies

We rigorously compare the effectiveness of the proposed InceptionV3, ResNet50, DINOv2, MoCov3 models with state-of-the-art scientific studies across diverse plant pest and disease detection approaches in this section. The proposed methodology demonstrates improved performance over existing methods, achieving 98.25% accuracy for walnut disease detection and 98.05% for walnut pest detection. A similar study for tomato leaf disease detection was suggested, adding multiple trainable dense layers and a dropout mechanism, achieving 94.01% detection accuracy (Shafik et al., 2024). The two-stage feature extraction approach leveraging transfer learning VG-GNBNet for areca nut disease detection scores the highest accuracy of 96.85% (Hegde and Kumar, 2025). Rice plant pests and diseases were detected in farmers’ fields using the EfficientNetB0 model with transfer learning, achieving a highest accuracy of 96.07% (Bijlwan et al., 2025). Anthracnose diseases in the cashew plant were addressed using pretrained DenseNet-201 models and a transfer learning approach, achieving 95.75% accuracy (Sudha and Kumaran, 2023). The comparative analysis of the proposed study with state-of-the-art pest and disease detection methods is summarized in Table 23.

4.4 Recommendations for sustainable horticulture and fruit productivity

To improve long-term fruit productivity and sustainable horticulture, ensuring food security, mitigating economic losses, and prompt pest and disease management are important. This study successfully addresses walnut pests and diseases for sustainable horticulture and fruit production. For sustainable crop protection, deployment of the proposed InceptionV3 and ResNet50 models on handheld and resource-constrained mobile applications and AIoT-enabled wireless sensor systems, further empowering farmers and marginalized communities in rural infrastructure, improving applied fruit science, horticulture, and phytoprotection programs, in a sustainable way. Further, model pruning and class-balance techniques can help researchers and developers in enhancing model robustness and minimizing misclassification in real-time settings. The relevance and contribution of our study to advancing several SDGs are presented in Table 24.

4.5 Improvements and future scope

Our research team has developed a robust and intelligent AI-based solution for prompt pest and disease detection in walnut plants, working together with walnut farmers in the remote area of Kupwara, Jammu and Kashmir, India. The efficient and accurate solvency of the developed model for walnut pest and disease detection can assist farmers through data-driven decision-making ability for timely interventions. The robust and improved results of the study reveal that AI can substantially improve walnut production and has the potential to transform horticulture and fruit-growing science. Still, some attention needs to be addressed in future work.

• Data scarcity remains a significant constraint in data-driven horticulture despite continuous improvements. This study includes 17 walnut pests and 3 disease classes and does not include intercrop or detailed metadata (e.g., tree/location identifiers). Future research should focus on developing high-resolution, large-scale, easily accessible, and standardized image datasets encompassing extensions across geographical regions, especially remote locations, intercrop systems incorporating detailed metadata of tree/location identifiers for farming ease, through expert collaboration leveraging synthetic data generation, and self-supervised learning approaches, which will further enhance model proficiency, adaptability, robustness, and study applications.

• Class imbalance in image analysis remained a considerable concern; use of oversampling, undersampling, or class weighting approaches further improve detection performance in horticulture.

• Integrating the proposed AI model with interpretability tools like t-SNE, IntGrad, LIME & SHAP techniques through AIoT (Artificial Intelligence of Things) and wireless sensor technology incorporating climatic and meteorological parameters represents an important direction for future work. Such multimodal approaches may further enhance model robustness, adaptability to fluctuating environmental conditions, and the capability for early pest and disease forecasting under changing climate scenarios, through a convenient mobile application or expert system, can improve remote pest and disease management, minimize economic losses, while support broader sustainability goals.

5 Conclusion

This study significantly addresses walnut pest and disease problems using deep learning and knowledge transfer techniques for sustainable agriculture, aligning well with several SDGs. The proposed InceptionV3, ResNet50, DINOv2 and MoCoV3 models were trained, validated, and tested using separate, well-curated, and diverse walnut pest and disease datasets. The walnut pest dataset includes 17 pest classes, while the walnut disease dataset comprises 3 classes. Data preprocessing involved selecting high-resolution images and generating synthetic samples to address class imbalance in both datasets. Suggested models were optimized using key hyperparameters, incorporating learning rate, batch size, dropout rate, and optimizer type, leveraging the transfer learning technique, yielding the best performance. The InceptionV3 model for pest detection outperformed ResNet50,DINOv2 and MoCoV3 achieving 99.87% training accuracy and 98.05% test accuracy, and achieves 99.87% training accuracy and 98.25% testing accuracy for disease detection, with a very low generalization gap conformed by expert plant pathologist and entomologist with Grad-cam visualization technique. This AI-powered, precise pest and disease detection method significantly contributes to the following SDGs: SDG 1 (No Poverty), SDG 2 (Zero Hunger), SDG 3 (Good Health and Well-being), SDG 12 (Responsible Consumption and Production), SDG 13 (Climate Action), and SDG 15 (Life on Land). Future improvements may focus on integrating ensemble or multimodal learning approaches for pest and disease management in walnut using diverse intercrops, mobile application, pathogen, or pesticide recommendation systems. Furthermore, investigating advanced pre-trained architectures such as EfficientNetB3, DenseNet201, and InceptionV4 could enhance model performance. Class imbalance remains a significant challenge in plant image analysis; therefore, techniques such as oversampling, undersampling, and class-weighting strategies may further improve detection accuracy. For large-scale pest and disease surveillance, real-time deployment of the proposed models through drone platforms and AIoT-enabled wireless systems, integrating the mobile application, offers substantial potential for next-generation smart horticulture and automated farming systems.

References

• 1

  AhadM. T.ChowdhuryN. H.AhmedA.NoorA. A.SongB.LiY. (2026). A lightweight convolutional neural network for real-time monitoring of smart mango orchard systems. Sci. Rep. doi: 10.1038/s41598-026-40758-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 2

  Ahmad LotiN. N.Mohd NoorM. R.ChangS. (2021). Integrated analysis of machine learning and deep learning in chili pest and disease identification. J. Sci. Food Agric.101, 3582–3594. doi: 10.1002/jsfa.10987

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 3

  AhmedK.DubeyM. K.PandeyD. K. (2025). Application of different transfer learning models to classify destructive insects in walnut fruit. Appl. Fruit Sci.67, 364. doi: 10.1007/s10341-025-01596-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 4

  AnwarZ.MasoodS.SrivastavaA. (2026). Empowering agriculture from pixels to diagnosis: A review of computer vision techniques for plant disease detection. Arch. Comput. Methods Eng. doi: 10.1007/s11831-026-10520-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 5

  AshrafM. A.NawazS.BilalU.AbubakarM.AnamS.AzamM.et al. (2026). Photoperiod and temperature: Key abiotic drivers of horticultural crop physiology in a changing climate. Biol. (Bratisl).81, 50. doi: 10.1007/s11756-025-02073-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 6

  BijlwanA.RanjanR.SinghM.PurohitR.NathA. J.ChakravartyS. (2025). Predicting crop disease severity using real-time weather variability through machine learning algorithms. Scientific Reports. 15 (1), 34767.

  • Google Scholar

• 7

  De MastroF.BrunettiG.FarragK.ZangH. (2025). Sustainable horticulture: Advancements and challenges in organic fertilizer applications. Horticulturae11, 307. doi: 10.3390/horticulturae11030307

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 8

  DevarajanD.AllafiR.ObayyaM.NemriN. (2026). AI based real time disease diagnosis in plants using deep learning driven CNNs. Sci. Rep.16, 4587. doi: 10.1038/s41598-025-34681-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 9

  GunasekaranH.RajkumarS.Kirubhadharsini B.L. (2026). Lightweight deep learning for tomato disease detection: Trends, challenges, and edge AI perspectives. Front. Plant Sci.16. doi: 10.3389/fpls.2025.1737208

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 10

  GuptaA.ChugA.SinghA. P. (2026). Feature extraction in sensor plant disease datasets using reformed membership functions independent of class variables. Sci. Rep.16, 4115. doi: 10.1038/s41598-025-33569-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 11

  HassanM.AhmadK. (2017). Anthracnose disease of walnut- a review. Int. J. Environment Agric. Biotechnol.2, 2319–2327. doi: 10.22161/ijeab/2.5.6

  • CrossRef
  • Google Scholar

• 12

  HegdeS. N.KumarH. S. V. (2025). “ Karnataka Malenadu region arecanut leafspot disease severity assessment using machine learning”, in: 2025 3rd International Conference on Smart Systems for Applications in Electrical Sciences (ICSSES) (Piscataway, NJ: IEEE), 1–5. doi: 10.1109/ICSSES64899.2025.11009869

  • CrossRef
  • Google Scholar

• 13

  InamdarN.ManaguliM.KotiR.JakatiJ.P. H.S.KulkarniP. (2026). Precision cotton disease detection via transformer models applied to leaf imagery. Front. Artif. Intell.8. doi: 10.3389/frai.2025.1743264

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 14

  KaradenizA. T.ÇelikY.BaşaranE. (2023). Classification of walnut varieties obtained from walnut leaf images by the recommended residual block based CNN model. Eur. Food Res. Technol.249, 727–738. doi: 10.1007/s00217-022-04168-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 15

  KaurR.MittalU.WadhawanA.AlmogrenA.SinglaJ.BharanyS.et al. (2025).  YOLO-LeafNet: A robust deep learning framework for multispecies plant disease detection with data augmentation. Sci. Rep.15, 28513. doi: 10.1038/s41598-025-14021-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 16

  KhanM. B.TamkinS.AraJ.AlamM.BhuiyanH. (2025). CropsDisNet: An AI-based platform for disease detection and advancing on-farm privacy solutions. Data (Basel).10, 25. doi: 10.3390/data10020025

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 17

  KhursheedV.TaufiqueM. (2020). Spatial analysis of horticulture efficiency and fruit production concentration in Kashmir Valley. GeoJournal85, 1635–1643. doi: 10.1007/s10708-019-10043-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 18

  KumarD.ChhoidubC.SharmaM.RaniS. (2026). Growth, trends and future projections in India’s horticulture sector: Insights from a time series analysis. Discover Agric.4, 47. doi: 10.1007/s44279-026-00490-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 19

  LiM.TaoZ.YanW.LinS.FengK.ZhangZ.et al. (2025). Apnet: Lightweight network for apricot tree disease and pest detection in real-world complex backgrounds. Plant Methods21, 4. doi: 10.1186/s13007-025-01324-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 20

  LiX.BianW.YangB.LiY.WangS.QinN.et al. (2026). Soybean leaf disease recognition methods based on hyperparameter transfer and progressive fine-tuning of large models. Agronomy16, 218. doi: 10.3390/agronomy16020218

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 21

  LilhoreU. K.KavithaP.G.S.R.S.SimaiyaS.GhithE. S.et al. (2026). Early detection of potato leaf diseases with PotatoNet-X: A deep learning model combining ResNet, DenseNet, and attention mechanisms. Potato Res.69, 45. doi: 10.1007/s11540-025-09957-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 22

  LiuY.ZhangW.CaoJ.GanM.FanX. (2026). Development of a rapid visual detection method for walnut anthracnose based on RPA-LFD. Plant Dis. doi: 10.1094/PDIS-11-25-2412-RE

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 23

  MadridA.BravoG.Castillo‐NovalesD.MolinaV. N.SalinasA.BesoaínX.et al. (2026). Antibacterial agents to control of pathogenic <scp> Xanthomonas arboricola </scp> pv. <scp> juglandis </scp> on walnut fruits and plants by polygodial derivatives formulations. Pest Manage. Sci.82, 2300–2312. doi: 10.1002/ps.70367

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 24

  MamunA. A.ZhangM.Ahmedt-AristizabalD.HayderZ.AwrangjebM. (2026). ConMamba: Contrastive vision Mamba for plant disease detection. Pattern Recognit.176, 113177. doi: 10.1016/j.patcog.2026.113177

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 25

  ManonoB. O.MwamiB.MutaviS.NziluF. (2026). Precision farming with smart sensors: Current state, challenges and future outlook. Sensors26, 882. doi: 10.3390/s26030882

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 26

  MiaM. J.MariaS. K.TakiS. S.BiswasA. A. (2021). Cucumber disease recognition using machine learning and transfer learning. Bull. Electrical Eng. Inf.10, 3432–3443. doi: 10.11591/eei.v10i6.3096

  • CrossRef
  • Google Scholar

• 27

  MirJ. I.IrfanM.RashidU.JavidS.RashidM.SharmaO. C.et al. (2025). Comparative pomological and biochemical characterization of indigenous walnut (Juglans regia) genotypes from temperate region. Appl. Fruit Sci.67, 119. doi: 10.1007/s10341-025-01372-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 28

  OjiamboP. S. (2025). Insect pest and plant disease management in horticultural crop production: Recent insights provide opportunities for improved control. Front. Hortic.4. doi: 10.3389/fhort.2025.1605035

  • CrossRef
  • Google Scholar

• 29

  QinR.WangY.LiuX.YuH. (2025). Advancing precision agriculture with deep learning enhanced SIS-YOLOv8 for Solanaceae crop monitoring. Front. Plant Sci.15. doi: 10.3389/fpls.2024.1485903

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 30

  QwaidM. A.SarkerM. T.ShawonS. M.ZubairH. T. (2026). AI-enabled smart farming framework for sustainable date palm cultivation in arid regions using machine learning and IoT integration. Sci. Rep.16, 5125. doi: 10.1038/s41598-026-36106-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 31

  RajanR.SinghD. R. (2025). Smart Horticulture Production (New York: Apple Academic Press). doi: 10.1201/9781998511679

  • CrossRef
  • Google Scholar

• 32

  RohithD.SaurabhP.BisenD. (2025). An integrated approach to apple leaf disease detection: Leveraging convolutional neural networks for accurate diagnosis. Multimed. Tools Appl.84, 40307–40342. doi: 10.1007/s11042-025-20735-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 33

  SambanaB.NnadiH. S.WajidM. A.FideliaN. O.Camacho-ZuñigaC.AjuzieH. D.et al. (2025). An efficient plant disease detection using transfer learning approach. Sci. Rep.15, 19082. doi: 10.1038/s41598-025-02271-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 34

  SankaranS.SubbiahD.ChokkalingamB. S. (2024). CitrusDiseaseNet: An integrated approach for automated citrus disease detection using deep learning and kernel extreme learning machine. Earth Sci. Inform.17, 3053–3070. doi: 10.1007/s12145-024-01320-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 35

  SaxenaA.FaiqM.GhatrehsamaniS.ZahraS. R. (2026). Integrating artificial intelligence and machine learning for sustainable development in agriculture and allied sectors of the temperate Himalayas. AgriEngineering8, 35. doi: 10.3390/agriengineering8010035

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 36

  ShafikW.TufailA.De Silva LiyanageC.ApongR. A. A. H. M. (2024). Using transfer learning-based plant disease classification and detection for sustainable agriculture. BMC Plant Biol.24, 136. doi: 10.1186/s12870-024-04825-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 37

  SharmaS.SharmaU.SharmaS.SharmaS.LikhitaJ.ThakurS.et al. (2025a). Strategies for sustainable land management through climate‐resilient horticulture: A systematic review. Land Degrad. Dev.36, 6225–6236. doi: 10.1002/ldr.70082

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 38

  SharmaS.SinghN.BhagatS.KumarR.SharmaA. (2026). “ Entomopathogens for pest management for sustainable agriculture,” in Biotechnology Innovations for a Sustainable Future ( Springer Nature Singapore, Singapore), 337–360. doi: 10.1007/978-981-97-9859-9_15

  • CrossRef
  • Google Scholar

• 39

  SharmaV.ChandelA.GuptaR. K.SharmaA.GuptaV.MishraN.et al. (2025b). Statistical insights into almond and walnut cultivation trends in Himachal Pradesh. Appl. Fruit Sci.67, 344. doi: 10.1007/s10341-025-01592-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 40

  ShuklaS.SinghR.KatochS.ShuklaS. K. (2025). “ Advanced stress mitigation techniques in horticulture: Enhancing crop resilience to biotic and abiotic challenges,” in Sustainable Agriculture Management in Semi-Arid Climates ( Springer Nature Switzerland, Cham), 45–69. doi: 10.1007/978-3-031-94066-8_3

  • CrossRef
  • Google Scholar

• 41

  SimhadriC. G.KondaveetiH. K.VatsavayiV. K.MitraA.AnanthachariP. (2025). Deep learning for rice leaf disease detection: A systematic literature review on emerging trends, methodologies and techniques. Inf. Process. Agric.12, 151–168. doi: 10.1016/j.inpa.2024.04.006

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 42

  SinghN. S.KumariA.TripathiV.BhardwajA.VermaR. B.FredericksR.et al. (2026). Post-harvest losses in horticultural crops: Causes and management. Eur. J. Nutr. Food. Saf.18, 195–213. doi: 10.9734/ejnfs/2026/v18i21967

  • CrossRef
  • Google Scholar

• 43

  SudhaP.KumaranP. (2023). Early detection and control of anthracnose disease in cashew leaves to improve crop yield using image processing and machine learning techniques. Signal. Image Video Process.17, 3323–3330. doi: 10.1007/s11760-023-02552-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 44

  VermaM.SharmaR.KotwalA.ManhasJ.SharmaV. (2026). Deep transfer learning-based ensemble framework for identifying plant-parasitic nematodes from microscopic images. Multimed. Tools Appl.85, 265. doi: 10.1007/s11042-026-21408-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 45

  WangJ.WangJ. (2024). A lightweight YOLOv8 based on attention mechanism for mango pest and disease detection. J. Real-Time Image Process.21, 136. doi: 10.1007/s11554-024-01505-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 46

  WuZ.PengD.HeM.HeC.HuY. (2026). EPDNet: A method for identifying tomato leaf diseases with uneven samples. Comput. Electron. Agric.240, 111145. doi: 10.1016/j.compag.2025.111145

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 47

  XuL.WangW.CaiZ.HuangW.ZhaiF.ZhaoJ. (2026). CFNet-YOLOv12: A novel deep learning framework for apple leaf disease detection. Hortic. Plant J. doi: 10.1016/j.hpj.2025.09.017

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 48

  YadavA.JajoriaD. K.KumariA.SharmaV.SodhaparmarH.MeenaP. K.et al. (2026). Next-generation precision farming technologies to enhance sustainable crop production and environmental sustainability. Bionature46, 64–81. doi: 10.56557/bn/2026/v46i12092

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 49

  YaseenM.JabeenA.HussainS. Z.GautamS.NazirN.BashirS.et al. (2025). Standardization of gamma irradiation doses for microbial decontamination and quality preservation in walnut kernels grown in temperate climate of western himalayas. Front. Food Sci. Technol.5. doi: 10.3389/frfst.2025.1669738

  • CrossRef
  • Google Scholar

• 50

  YeshiwasY.TadeleE.MohammedA.HuX. (2026). Impacts of climate change on horticultural systems with focus on socio-economic implications, production and postharvest challenges, and adaptive pathways. Discover Environ.4, 55. doi: 10.1007/s44274-026-00539-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 51

  YuanC.LiS.WangK.LiuQ.LiW.ZhaoW.et al. (2025). Mamba-YOLO-ML: A state-space model-based approach for mulberry leaf disease detection. Plants14, 2084. doi: 10.3390/plants14132084

  • Pubmed Abstract
  • CrossRef
  • Google Scholar