High-resolution remote sensing suggests greater vegetation resilience in maize-based intercropping systems than monocropping

Abstract

Introduction:

Intercropping is widely promoted as a strategy to enhance crop resilience, but spatial evidence for its vegetation- stress benefits in smallholder systems based on remote sensing data remains limited.

Methods:

We used Sentinel-1 and Sentinel-2 data to map maize monocrop and maize intercrop in western Kenya, and Sentinel-2 vegetation indices to compare vegetation-stress conditions across these systems during the 2021 and 2023 long-rain seasons. We tracked monthly vegetation condition using normalized difference indices of vegetation, (VCI from NDVI, MCI from NDMI, and GCI from GNDVI) and were integrated into an ensemble stress index ENS. All these four metrics were subsequently incorporated into the Intercrop Advantage Score (IAS) - a remote-sensing proxy that combines the magnitude, uncertainty and seasonal consistency of intercrop-monocrop differences.

Results and Discussion:

Across both seasons, maize intercrop generally showed higher vegetation condition and lower stress than maize monocrop. The relatively small, but statistically significant differences, provide evidence for resilience from remotesensing, in the absence of yield data. The IAS provides an interpretable transferable framework for comparing crop-system vegetation-stress profiles using satellite data and has potential application for evaluating of a wider range of agricultural interventions.

1 Introduction

Crop resilience refers to the capacity of a crop or cropping system to resist and recover from stresses while maintaining productivity (Shilomboleni et al., 2024; Zampieri et al., 2020). Prior studies have quantified resilience using metrics like yield stability (Redhead et al., 2020) and soil moisture (Eeswaran et al., 2021). Direct assessment of crop-system resilience usually requires evidence on productivity, yield stability, post-stress recovery trajectories, or farmer-level outcomes. However, other practical metrics or indicators can be considered, such as remote sensing-based indicators have the potential to provide a scalable alternative, especially where ground truth data do not exist – as is the case for the vast majority of smallholder systems. Relevant indicators from remote sensing include plant stress responses (Wu and Liang, 2020; Shen et al., 2025) and use of vegetation indices to measure recovery from stress (Wang et al., 2025; Ji et al., 2025).

Here, we develop and evaluate a remote-sensing proxy for relative vegetation stress resilience, which we define as, the degree to which maize intercrop and maize monocrop systems maintain favourable canopy condition during the long-rain growing season. This distinction is important because vegetation, stress resistance, stress recovery and productivity represent related but different dimensions of resilience. Vegetation indices describe canopy state however stress resistance refers to the maintenance of favourable index values during stress. Indeed, NDVI and NDVI-derived indices have been widely used to detect drought onset and assess vegetation stress impacts (Kogan, 1995; Liu and Kogan, 1996; Jiao et al., 2016). For example, NDVI-based indices are used to measure vegetation health under stress: the Vegetation Condition Index (VCI) assess current vegetation health by comparing its current Normalized Difference Vegetation Index (NDVI) to historical minimum and maximum values for the same period, with low VCI values indicating drought stress and high VCI values indicating favourable vegetation condition (Kogan, 1990; Liu and Kogan, 1996).

Complementary vegetation indices can be used to assess crop condition under smallholder systems. The Normalized Difference Moisture Index (NDMI) captures water related (abiotic)stress (Lykhovyd and Sharii, 2024), because it is sensitive to canopy water content. Unlike NDVI, NDMI has shown superior sensitivity to fluctuations in vegetation water content (VWC), particularly under drought conditions. Jackson et al. (2004) used Landsat data over corn and soybean crops, demonstrated that while NDVI saturated, NDMI continued to reflect dynamic VWC changes with lower bias and error. Xu et al. (2020) showed that combining MODIS and Landsat enabled operational retrieval of NDMI-based daily VWC estimates at 30 m resolution. Similar conclusions on the value of NDMI have been obtained elsewhere (e.g., Zhou et al., 2022).

Green Normalized Difference Vegetation Index (GNDVI) is a chlorophyll-sensitive index used to monitor biotic stress and nutrient-related vigour loss. It is particularly useful for detecting early symptoms of nutrient deficiency or potential disease-related stress. For example, Jafari Jozani et al. (2022) used RapidEye and Landsat-8 to map Maize Lethal Necrosis in Kenya and identified GNDVI as key in distinguishing severely infected fields. Similarly, Radócz et al. (2023) reported significant GNDVI declines with increasing infection levels in sweet corn, highlighting its diagnostic value. Wójtowicz et al. (2021) and Abdulridha et al. (2023) also identified GNDVI as among the most informative indices for disease severity classification.

Intercropping is a mixed cropping strategy to maximize land use, improve soil fertility, and avoid risks related to climate. Recent studies in East Africa have found that mixed cropping can offer agronomic benefits. For example, intercropping in Rwanda improved soil properties and increased maize yields compared to monocropping (Hashakimana et al., 2024). However, different crops grown together frequently exhibit similar growth cycles and spectral signatures making it difficult for classifiers to distinguish crop types (Maleki et al., 2024). This difficulty is consistent with recent reviews of crop-type mapping, which show that remote-sensing crop classification has progressed rapidly but remains challenging in heterogeneous agriculture landscapes where crop type, crop mixture and management practices vary at fine spatial scales (Huang et al., 2026). Recent progress using high spatial and temporal resolution Sentinel products is successfully addressing the challenges related to classification of intercropped crop systems (e.g., Ibrahim et al., 2021; Liepa et al., 2024; Mahlayeye et al., 2024). Therefore, assessing the resilience benefit of intercropping has recently become achievable.

The current study was conducted in Western Kenya (Supplementary Information: Section 1 and Figure 1), a region characterized by a bimodal rainfall regime and highly diverse smallholder agricultural systems dominated by maize and mixed intercropping. The long rains extend from March to May, and the short rains lasts from October to December. Mean annual long-term rainfall equal to 1581 mm and Long-term average temperature ranges between 20.9 and 22.3 °C Musafiri et al. (2022). While intercropping is widely used in the study area, for its agronomic benefits in smallholder systems, there remains a lack or spatially explicit, quantitative methods using remote sensing-based approaches to assess whether it truly enhances stress resilience compared to monocrop system under real-world conditions. Moreover, although remote sensing has been increasingly used to monitor plant stress, few studies have systematically compared intercrop and monocrop systems using high-resolution vegetation indices. We address these gaps by quantifying and comparing a remote-sensing-based proxy for relative vegetation stress resilience in intercrop versus monocrop maize systems, rather than claiming a direct measurement of resilience.

Figure 1

This study contributes to remote sensing of smallholder crop systems in three ways. First, it applies a hierarchical crop-system classification methodology to separate maize monocrop and maize intercrop systems in a heterogeneous smallholder landscape. Second, it statistically compares the vegetation-stress profiles of these systems using Sentinel-2-derived normalized vegetation indices: VCI derived from NDVI, MCI derived from NDMI, and GCI derived from GNDVI. Third, it proposes the Intercrop Advantage Score (IAS), an interpretable remote-sensing proxy that combines magnitude, uncertainty and seasonal consistency to summarize relative stress resilience between crop systems. Whilst the practical significance of the IAS is out of scope for the current study, section 4 discusses this topic and presents some next steps in making the IAS a reliable and practical tool.

2 Materials and methods

2.1 Overview of the methodology

The iterative, multi-step, remote sensing framework used in this study is presented in Figure 1. Our methodology combines classification, stress monitoring, and statistical validation. First, we tested three classification methodologies to map the cropping systems in the study area, distinguishing maize intercrop from maize monocrop. The approach that provided the most satisfactory performance and spatial coherence is described in the main text, while the other classification approaches are presented in Supplementary Information (Section 2 and 3. Second, we monitored vegetation stress over time. The analysis focused on detecting vegetation stress during the long-rain growing season using Sentinel-2 data. NDVI, NDMI and GNDVI were normalized to produce the Vegetation Condition Index (VCI), the Moisture Condition Index (MCI) and the Greenness Condition Index (GCI), respectively (see Section 2.3.2). These three normalized indices capture complementary dimensions of crop stress. In addition, we developed a combined index (denoted ENS) that integrates these three indicators to provide a more complete measure of plant stress that include the general condition, biotic and abiotic stress. Meteorological and hydrological indicators provided essential context to interpret the vegetation stress patterns (see Section 2.3.3 for details on data and analysis). Statistical comparisons and sensitivity analyses, including pixel and field-level checks and 4 km spatial block bootstrapping, were used to evaluate the robustness of the observed differences (see Section 2.3.4 and 2.3.5). Finally, using these remote-sensing methods, a further analysis for remote-sensing-based proxy for relative vegetation stress resilience was performed, integrating a weighted stress scoring metric (StressScore), system difference measures (Δ), and bootstrap-based uncertainty framework leading to the development of the Intercrop Advantage Score (IAS) (see Section 2.3.6 and 2.3.7).

2.2 Materials

2.2.1 Remote sensing data

The analysis was based on the use of Sentinel-2 and Sentinel-1 imagery from Copernicus Open Access Hub and processed through Google Earth Engine (GEE). The datasets were filtered to the long-rain cropping season ranged between April to August for both 2021 and 2023.

Sentinel-2 (MSI Level-2A) provided atmospherically corrected surface reflectance data at 10–20 m spatial resolution, covering the visible, near-infrared (NIR) and shortwave infrared (SWIR) spectral wavelength. To ensure reliable pixel quality, cloud and shadow masking were applied using the Scene Classification Layer (SCL) provided with the Level-2A product. The following classes were excluded from analysis: 0 (no data), 1(saturated or defective pixels), 3 (cloud shadow), 8 (medium probability cloud), and 10 (thin cirrus). This filtering step removed areas affected by cloud contamination, sensor errors, or illumination artefacts, retaining only high-quality surface reflectance observations.

Sentinel-1 data were used to complement optical observations and provide information on crop structure for the crop classification for the four steps classification method (Supplementary Information). The C-band Synthetic Aperture Radar (SAR) images (Interferometric Wide Swath mode, dual polarization VV and VH) were acquired from the Sentinel-1 Ground Range Detected (GRD) collection. All images were pre-processed using the standard European Space Agency (ESA) workflow, which includes radiometric calibration, terrain correction and speckle filtering. The data were converted to sigma nought () backscatter coefficients in decibels (dB) and resampled to a 10m spatial resolution to match the Sentinel-2 grid. Both VV and VH bands, along with their VV/VH ratio, were included in the feature stack to capture textural differences associated with canopy.

2.2.2 Ground-truth reference samples

Ground-truth data were obtained from the Copernicus4GEOGLAM Kenya field-campaign database, available at: https://jeodpp.jrc.ec.europa.eu/ftp/jrc-opendata/Copernicus4GEOGLAM/Kenya/Field_Campaign/, which provides harmonized in-situ crop type observations collected through coordinated field surveys. The dataset includes georeferenced and delineated field-level polygons. Each polygon is annotated with crop type information, cropping system and additional metadata. Only polygons located within the western Kenya Area of Interest were retained. The resulting dataset includes a diverse set of labelled fields representing the main cropping systems: maize monocrop and maize-based intercropping systems. These field samples were used as training and validation data for all classification experiments (see Supplementary Information).

2.3 Methods

2.3.1 Multi-sensor four-step classification

The classification method was based on combining the use of Sentinel-2 and Sentinel-1 Ground Range Detected (GRD) data. Sentinel-2 scenes were cloud and shadow masked using the scene classification algorithm, while Sentinel-1 data acquired in interferometric wide (IW) provided dual-polarization backscatter (VV and VH) and their ratio.

A multi-sensor feature stack was constructed by combining spectral reflectance bands (visible and shortwave infrared), radar backscatter metrics and vegetation (NDVI, EVI and NDMI). All input datasets were resampled to 10m spatial grid defined by Sentinel-2, ensuring pixel-level correspondence between optical and radar features. Temporal alignment was achieved by restricting all inputs to the same seasonal window, long rain season starting in April until August. The resulting feature stack contained, for each pixel, a synchronized vector of spectral, radar, and index-based attributes used in supervised classification. This organization minimized spatial shifts and radiometric inconsistencies between sensors allowing robust model training and inference across the study area. Supervised classification was implemented using a Random Forest (RF). For the Random Forest crop-system classification, no formal hyperparameter-tuning procedure was performed. A fixed Random Forest configuration was used consistently across the classification steps and years, and classification reliability was evaluated using validation samples, confusion matrices, overall accuracy, producer’s accuracy, user’s accuracy and F1-score. RF was trained using ground truth observations obtained by GEOGLAM (Table 1). The workflow followed a hierarchical, four steps to progressively isolate cropland, maize, and intercropping systems.

Step 1 cropland versus natural vegetation: For this first step classification two classes were used: cropland and natural vegetation that include labels from ground truth data such as forest shrubland, and grassland. These two classes were used to train a binary Random Forest RF classifier, 80% of the data for the training and 20% for the validation. The output of this first step was a cropland map.

Step 2 maize all together (monocrop and intercrop) versus the rest of the crops: within the cropland map, pixels were reclassified as maize when field labels included monocrop maize and maize-based mixtures (maize-beans, maize-potatoes, maize-potato-beans, maize-others) or other crops. As in the previous step these two classes were used to train the RF algorithm using 80% of the data for the training and 20% for the validation. The resulting maize mask delineated maize (monocrop and intercropped maize with other crops) growing areas across the study area.

Step 3 Monocrop versus intercrop maize: pixels identified as maize from the previous step were further divided into monocrop (only maize) and intercropped (maize intercropped with other crops) classes. A separate RF model was trained on this subset using the same input features and strategy. The output map distinguished monocrop and intercrop maize systems. All these steps are summarized in Table 2.

Step 4 Refinement of the obtained map using Dynamic World: to minimize residual misclassification in non-cropland areas, the Step 3 output was refined using Dynamic World v1 probability layers. Pixels with non-cropland probabilities greater than 0.4 (built-up, shrub/scrub, flooded vegetation, grass, trees or water) were excluded. The resulting product provided a final three-class map distinguishing monocrop maize, intercrop maize and non-maize or no-data pixels.

2.3.2 Vegetation stress analysis: computation, normalization, and integration

To support the interpretation of vegetation stress, the GEOGLAM ground-truth polygons were used in two complementary ways. First, a small subset of the available GEOGLAM polygons from the 2023 long-rain season was used as an initial reference check of the raw NDVI, GNDVI and NDMI seasonal dynamics. This check, reported in the Supplementary Information (section 4, Figure 5), was used only to verify the temporal behaviour of the raw vegetation indices of some maize monocrop and maize intercrop samples. It was not used at the main spatial basis for the vegetation-stress maps, statistical comparisons of IAS analysis. Second, all available maize monocrop and maize intercrop GEOGLAM polygons were used later in the in the pixel- and field-level sensitivity analysis addressing pseudo-replication and spatial dependence (Section 2.3.5). The main vegetation-stress mapping, statistical comparison and IAS assessment were based on the classified crop-system map outputs, in which maize pixels across the study area were separated into maize monocrop and maize intercrop classes. Using these classification outputs, vegetation stress was evaluated with the Sentinel-2 derived vegetation indices. Each selected index captures a complementary aspect of crop condition: general vegetation condition, canopy moisture status and greenness-related vegetation condition.

Figure 2

The VCI was computed from the Normalized Difference Vegetation Index NDVI, which indicates relative vegetation vigour, and then compared to its historical range (Kogan, 1990; Liu and Kogan, 1996). The MCI, which indicates canopy moisture status, was derived from the Normalized Difference Moisture Index NDMI (Jackson et al., 2004), and was also compared to its historical range. The GCI was computed based on the GNDVI (Gitelson et al., 1996) and normalized relative to its historical ranges. The resulting index represents the chlorophyll concentration and canopy density and serves as proxy for nutrient status and pest-related stress. Taken together, these indices allow an evaluation of vegetation responses to both abiotic and biotic stresses.

For the year 2021, each index was normalized using multi-year from 2016 to 2020, as for 2023 the multi-year period from 2016 to 2022, to ensure inter-annual comparability. For each month of the long rainy season (from April to August) the long-term minimum () and maximum () values were derived across the baseline years, and the normalized index for the year t was computed as:

where represents the monthly mean of the given index. The normalization scaled all indices from 0 to 100, where higher values indicate better vegetation conditions relative to the historical record. Normalized indices were classified into four stress levels: extreme stress (0-10), severe stress (10-20), moderate stress (20-50) and healthy vegetation (>50). The stress classes were adapted from the VCI/vegetation-health framework, in which normalized vegetation condition values range from 0 to 100 and lower values indicate poorer vegetation condition. Kogan et al. (2003) showed that VCI values around 50 represent near-normal vegetation condition, values below 30 indicate strong drought/stress, and values above 60 indicate favourable vegetation condition. The same class structure was applied to VCI, GCI and MCI because all three indices were normalized to their historical range.

To obtain a single index (the ensemble stress index, ENS) that integrates biotic, abiotic stress and general vegetation conditions, the three indices were combined through a majority vote rule (2 out of 3) applied at the pixel level. For each pixel, agreement between indices was evaluated by comparing their assigned stress classes across all index pairs (VCI-MCI, VCI-GCI, and MCI-GCI). Where at least two indices agreed on the class, that class was assigned as the value of ENS for that pixel. Where there was no agreement between any two indices, the pixel is excluded from the analysis.

Cropping system maps were generated using the four-step classification method. This enabled vegetation stress classification to be conducted separately for monocropped and intercropped maize across each month. This analysis was performed using the three normalized indices (VCI, MCI, GCI) as well as the combined index ENS.

2.3.3 Climatic and hydrological contextualization of the results

To provide a contextual understanding of the meteorological and hydrological conditions, multiple complementary datasets were used. These datasets aimed to describe rainfall variability and soil moisture dynamics across the study area.

Monthly precipitation data from CHIRPS dataset (1981 to 2024) were used to quantify rainfall variability across the study area. A long-term monthly climatology was computed from the entire record and used to derive monthly precipitation anomalies, representing deviations from historical averages. Annual totals were also calculated and compared to the long-term mean to highlight interannual variability.

To complement the anomaly analysis, the Standardized Precipitation Index (SPI) was derived from the CHIRPS monthly precipitation data using the Google Climate Platform. The SPI was computed for the 1981–2023 reference period based on a log-logistic probability distribution and categorized according to established drought severity classes (from extremely dry to extremely wet). The index was calculated for the main cropping period (April-August) to ensure consistency with vegetation stress assessments. A workflow was implemented to extract SPI class distributions of the study area allowing evaluation of spatial differences in rainfall variability.

To characterize hydrological conditions, soil moisture datasets were incorporated from two complementary sources. Surface soil moisture (0-5cm) was obtained from NASA SMAP providing monthly mean anomaly values at 9km spatial resolution. Deeper soil moisture profiles were retrieved from the FLDAS NOAH model for the depth 0–10 cm at 0.1° resolution. Monthly values were extracted and averaged over the study are to obtain representative temporal trends in soil water content.

2.3.4 Statistical assessment of resilience

For the main vegetation-stress analysis, monthly distribution of the normalized stress indices (VCI, MCI, GCI and ENS) were extracted from maize monocrop and maize intercrop pixels for April-August in 2021 and 2023. The corresponding raw-index distributions (NDVI, NDMI and GNDVI) were also analysed as a supporting check. Both analyses are reported in the Supplementary Information.

For each year, month and index, monocrop and intercrop distributions were compared using the two-sample Kolmogorov-Smirnov (KS) test for both raw and normalized vegetation-index values, Cliff’s delta was also calculated as a non-parametric effect-size measure to describe the direction and magnitude of the monocrop-intercrop difference. Positive Cliff’s delta values indicate higher values for intercropped maize, while values close to zero indicate strong overlap between the two distributions. Detailed KS statistics, permutation p-values, effect sizes and formulas are provided in the Supplementary Information (section 5).

As these comparisons were initially performed at pixel level, additional sensitivity analyses were conducted to evaluate the influence of observational scale and spatial dependence. These analyses included field-level aggregation using GEOGLAM polygons, as described in the following section (Section 2.3.5).

2.3.5 Sensitivity analysis for pixel-level pseudo-replication and spatial dependence

To assess the potential influence of pixel-level-replication, we performed the vegetation-index comparison at both pixel level and field level using all the available GEOGLAM ground-truth polygons. For the pixel-level analysis, all pixel values within the ground-truth polygons were extracted and compared between maize monocrop and maize intercrop systems for each year, month and vegetation index. For the field-level analysis, pixel values within each polygon were first aggregated to a polygon-level mean and median, and monocrop and intercrop systems were compared using these polygon-level summaries as the observational units. The field-level analysis used 107 maize monocrop and 105 maize intercrop polygons in 2021, and 110 maize monocrop and 70 maize intercrop polygons in 2023.

To account for spatial dependence in the pixel-based analysis, we applied spatial block bootstrapping as a sensitivity test for both 2021 and 2023. We performed the analysis using a 4 km grid as size block, where classified maize pixels were assigned to maize monocrop or intercrop using the crop-system maps for each year. To justify the spatial block size more explicitly, we estimated empirical semivariograms from the maize-pixel vegetation-index values. Semivariance was calculated across increasing distance classes, and the approximate spatial autocorrelation range was identified as the distance at which the semivariogram approached its sill. Semivariance measures how dissimilar pixel values become as the distance between them increases. If pixels are spatially autocorrelated, semivariance is low at short distances and increases with distance until it approaches the sill. The distance at which the curve approaches the sill is considered as the approximate autocorrelation range. In our data, the annual median ranges were 3.65 km in 2021 and 3.9 km in 2023, and similar values were obtained with 75^th percentile range (Figure 3).

Figure 3

After the definition of the adequate block size, within each 4 km block, classified maize pixels were separated into monocrop and intercrop pixels using the corresponding crop-system map produced by the classification workflow. For each crop-system group inside the block, we calculated the mean and median vegetation-index value. We then calculated the block-level difference as intercrop minus monocrop. This produced one mean-based difference and one median-based difference for each block. A positive difference means that intercrop has a higher vegetation-index value than monocrop in that block, while a negative difference means that monocrop was higher. These block-level differences were then used for the bootstrap analysis. For each year, month and vegetation index, the observed result was calculated as the average of all block-level differences. To estimate uncertainty, the blocks, were randomly resampled 5000 times using the Bootstrap method. This produces a bootstrap distribution of 5000 possible average intercrop minus monocrop differences. The 95% confidence interval was calculated from the 2.5^th and 97.5^th percentiles. The bootstrap p-value was calculated using an empirical two-sided bootstrap/randomization -style formula:

where Pleft and Pright are the proportions of bootstrap estimates falling on the two sides of the null value. In this analysis, the tested entity was the intercrop minus monocrop difference. Therefore, the null hypothesis was that there was no difference between the crop systems.

Because the null hypothesis represents no difference, zero was used as the reference value. The two sides of zero were therefore:

• • Pleft = proportion of bootstrap estimates 0

• • Pright = proportion of bootstrap estimates 0

The empirical two-sided bootstrap p-value was then calculated as:

where is the bootstrapped average block-level difference. Smaller p-values mean that the bootstrapped differences remained consistently on one side of zero, giving stronger evidence that the average intercrop minus monocrop difference was not equal to zero.

2.3.6 Resilience indicator: intercrop advantage score

To summarize the vegetation-stress profile of each crop system, we calculated a StressScore for classified maize pixels separated into monocrop and intercrop classes using the crop-system maps produced by the classification workflow. The StressScore was computed for each crop system, vegetation index (VCI, MCI, GCI and ENS), month and year. It converts the four categorical vegetation-stress classes into a continuous and interpretable metric, allowing monocrop and intercrop systems to be compared consistently across indices and months. Each StressScore was derived from the fractional area of pixels in the four vegetation stress categories (defined earlier): Healthy (H), Moderate (M), Severe (S) and Extreme (X), weighted according to physiological logic and relevance as following:

The assigned weights scale the contribution of each category such that the resulting score lies between -1 and +1 (or -100 and +100 when expressed in percent). We used an ordinal and symmetric scoring system: health vegetation was assigned to +1, moderate

stress +0.5, severe stress -0.5 and extreme stress -1. This gives equal magnitude but opposite sign to the two end-member classes (healthy and extreme stress), and equal magnitude but opposite sign to the intermediate classes (moderate and severe stress). The purpose is to summarize the balance between favourable and stressed vegetation conditions in a simple and interpretable way.

For each vegetation index and month, the difference between cropping systems was then computed as following:

A positive Δ indicates healthier vegetation under intercropping, a negative Δ reflects monocrop advantage. Given that each StressScore is bounded within [-1, +1], so the Δ metric ranges between [-2, 2] or (-200 and +200 in percentage form). This Δ captures the relative vegetation health advantage between systems on biophysically interpretable scale for each month of the long rain season from April to August.

To mitigate month-specific variability and highlight overall patterns, Δ values were averaged across the cropping season (April to August) for each year and vegetation index. Two complementary descriptors were calculated:

• • Mean Δ: representing the average difference in vegetation condition between systems across the season.

• • Consistency: representing the number of months with , reflecting the persistence of the intercrop advantage.

Together, these metrics quantify both the magnitude and temporal stability of resilience differences between cropping systems.

Uncertainty in Δ estimates was quantified through bootstrap resampling. A new dataset (a bootstrap sample) was created by randomly selecting data points from the original dataset, with the important rule that each selected data point is returned to the pool (replaced) before the next selection. For each month, crop systems and index, the class counts (H, M, S, X) were repeatedly resampled under a multinomial model (B = 500), preserving total pixel counts. For each bootstrap replicate, new StressScore and Δ values were computed, yielding a distribution of possible Δ outcomes. From this, we derived:

• • 95% Confidence Intervals (CIs) using percentile bounds (2.5th and 97.5th)

• •  Bootstrap standard deviations SDs as measures of dispersion.

The bootstrap method allows uncertainty quantification without assuming statistical normality, ensuring that the multimodal distributions, which is typical of satellite-derived data, is taken into account. Two complementary uncertainty sources were then distinguished:

• • Seasonal uncertainty: variability of the mean Δ across the whole season, quantified by CI width.

• • Monthly variability: fluctuation of Δ between months, quantified by the monthly bootstrap SD

Those two complementary uncertainty sources were used to determine whether the observed Δ values reflect true physiological differences between the two-cropping system or were simply noise. This was achieved using an adaptive threshold , calculated for each vegetation index:

where the first term accounts for seasonal uncertainty, while the second term take into consideration the monthly short-term variability.

Taking the maximum ensures a conservative interpretation because Δ must exceed both noise components to be considered meaningful. It was calculated for each vegetation index, so that each vegetation index obtains a customized that reflects its natural variance structure instead of applying a fixed threshold to all indices.

Δ, and temporal consistency were integrated to assign each index-year result to an Intercrop Advantage Score (IAS) category. These thresholds were defined during the development of the IAS workflow, to provide a conservative interpretation of the intercrop-minus-monocrop difference. The threshold was calculated for each vegetation index as the maximum of two uncertainty components: half of the bootstrap confidence-interval width, representing uncertainty around the seasonal mean Δ, and twice the median monthly bootstrap standard deviation, representing month-to-month variability. This ensured that Δ had to exceed both seasonal uncertainty and short-term variability before being interpreted as an intercrop advantage. A strong intercrop advantage was assigned when 2 and consistency was ≥ 70%, meaning that the seasonal difference was at least twice the uncertainty threshold and 70% means it persisted high in at least 4 out of 5 of monthly observations (5 months the period of long-rain season). A moderate intercrop advantage was assigned when 2 and consistency was ≥ 60%, meaning that the difference exceeded the uncertainty threshold and persisted in at least 3 out of 5 of monthly observations. Results were classified as inconclusive when or consistency <60%. Negative Δ values were interpreted using the same thresholds for monocrop advantage. This method tries to ensure that detected differences in vegetation condition represent stable resilience benefits rather than temporary or noise-driven fluctuations.

Therefore, the combination of Δ, and temporal consistency was synthesized into a categorical intercrop Advantage Score IAS to summarize resilience outcomes:

• • Strong intercrop advantage: 2 and consistency 70%.

• • Moderate Intercrop advantage: 2 and consistency 60%.

• • Inconclusive: or consistency <60%.

• • Negative thresholds are interpreted likewise for monocrop advantage.

2.3.7 IAS sensitivity analysis

To assess whether the IAS results depended on selected parameter values, we recalculate the IAS under alternative threshold, weighting and decision-rule scenarios (Table 3). In total 540 parameter scenarios were tested for each year-index combination, combining 5 alternative stress-category thresholds, 12 StressScore weights, 3 strong-advantage rules and 3 monthly consistency thresholds. All scenarios used the same vegetation-index data and crop-system classification. For each scenario, the IAS class was recalculated and compared with the baseline result. This allowed us to evaluate whether the interpretation of intercrop advantage was stable under plausible alternative parameters choices, rather than depend on one fixed set of thresholds.

3 Results

3.1 Multi-sensor four-step classification

The classification approach combined Sentinel-1 and Sentinel-2 time series in a hierarchical four-step classification framework designed to progressively separate land-cover and cropping categories. This structure included: (step 1) cropland versus natural vegetation, (step 2) maize versus other crops, (step 3), monocrop maize versus intercropped maize, and (step 4) refinement using spatial filtering. In 2021, step 1 (cropland, non-cropland) achieved an OA equal to 0.97, while step 2 (maize, other crops) yielded 0.89, and the third step 3 discriminating monocrop and intercrop maize reached 0.78 OA. Results for 2023 were comparable, with step 1 OA = 0.91, step 2 OA equal to 0.81 and step 3 equal to 0.82. All the results are summarized in Table 4.

Spatially, the classified maps (Figure 4) show a realistic distribution of maize and intercropped pixels over zones of Kisumu County. The 2023 map reveals slightly greater intercrop coverage than 2021. The performance across for both classes and years demonstrates the capacity of the hierarchical steps and multi-sensor approach to generate more accurate classification results for the complicated crop systems in western Kenya. The combination of optical and radar data proved to be beneficial for minimizing misclassification that may be caused probably by cloud contamination and canopy structure differences. Sentinel-1 backscatter added sensitivity to crop architecture complementing Sentinel-2 indices. This data combination yielded more stable prediction than single-sensor models.

Figure 4

3.2 Vegetation stress results

After obtaining the classification results using the four-steps method (, vegetation stress classification was performed for both obtained classes (intercrop and monocrop) and for each month of the growing season of long rains (April to August) in both years 2021 and 2023, for all vegetation indices VCI, MCI, GCI and Ensemble ENS the vegetation index that integrate the three-vegetation index. To illustrate the main patterns, Figures 5, 2 present ENS for June, corresponding to the peak of the growing season, for 2021 and 2023. The maps display the four stress classes: Extreme, Severe, Moderate and Healthy across monocrop and intercropped maize systems.

Figure 5

Figure 6, and Figure 6 in Supplementary Information provide monthly bar plots of vegetation stress category distribution across monocropped and intercropped maize systems in 2021 and 2023. These visualizations clearly reveal the temporal dynamics of crop condition as captured by each index and their combination ENS. Across both years, the intercropped system shows a higher proportion of Healthy pixels and reduced Extreme/Severe categories especially during mid-season when crop stress risk is typically highest

Figure 6

3.3 Climatic and hydrological contextualization of the results

Meteorological and hydrological indicators provide essential context to interpret and understand the meteorological context of the obtained results. According to CHIRPS precipitation records Figure 7, 2021 received slightly lower mean rainfall than 2023; however, the soil moisture conditions at the beginning of the 2021 season were significantly higher. As shown in SMAP and FLDAS soil moisture time series (Figure 7), early 2021 followed an exceptionally wet 2020, which likely contributed to important residual soil moisture that can be observed in both shallow (0–5 cm) and deeper (0–10 cm) layers. This antecedent moisture helped sustain crop water availability during the early stages of the growing season. As result, vegetation stress levels remained low in 2021 despite the lower precipitation compared to 2023. On the other hand, 2023 was rainier throughout the season with moderate anomalies and slightly reduced soil moisture levels especially during the mid-season.

Figure 7

Seasonal SPI (April-August) provides a summary of rainfall conditions during the cropping season. Table 5 shows the proportion of the study area falling within the near-normal (-0.5 to 0.5) and slightly wet (0.5 to 0.7) of the SPI classes for 2021 and 2023,

These values confirm that: (i) 2021 was a very uniform season, with nearly the entire area experiencing near-normal rainfall. (ii) 2023 showed more hydrological variability, with over one-third of the area classified as slightly wet. Together with soil-moisture evidence, SPI reinforces that 2021 was buffered by high antecedent moisture, while 2023 reflected more variable rainfall contributing to more spatial variation in vegetation stress.

In summary, the vegetation stress patterns observed in 2021 and 2023 cannot be explained by rainfall alone. Instead, they are the consequence from combined influence of precipitation and antecedent soil moisture and probably crop management practices. The high residual moisture entering the 2021 season helped sustain crop growth despite rainfall deficits (compared to 2023), producing low and spatially uniform stress.

3.4 Sensitivity analysis of pixel-level statistical results and spatial dependence

The main pixel-level KS analysis was first applied to classified maize pixels separated into maize monocrop and maize intercrop using the annual crop-system maps. These tests showed that monocrop and intercrop vegetation-index distributions were frequently different across years, months and indices, with intercrop generally showing higher vegetation-index values. However, the associated Cliff’s delta values were small, indicating substantial overlap between the two crop-system distributions. The full KS statistics, permutation p-values, Cliff’s delta values and monthly distributional results are provided in Supplementary Information Figures 7–12, Tables 10, 11. Because these tests were based on pixel-level observations, they were followed by field-level and spatial block-bootstrap sensitivity analysed to evaluate whether the direction of the result remained stable when observational scale and spatial dependence were considered.

Because neighbouring pixels may have similar vegetation-index values, the pixel-level KS tests were evaluated alongside more conservative sensitivity analyses. Table 6 summarizes the pixel-level, field-level and spatial block-bootstrap results. The detailed supporting results are provided in Supplementary Information Tables 12–20: sample sizes in Table 12, raw pixel-pixel results and their average differences in Supplementary Tables S13, S14, normalized pixel-level results and their average differences in Supplementary Tables S15, 16, field-level polygon results and their average differences in Supplementary Tables S17–19 and spatial block-bootstrap results in Supplementary Table S20.

In this section, the term comparison refers to one year-month-index combination unless otherwise stated. Therefore, 60 comparisons correspond to six vegetation indices (NDVI, NDMI, GNDVI, VCI, GCI and MCI) analysed over five months (April-August) and two years (6 indices x 5 months x 2 years = 60). For the spatial block bootstrap, each year-month-index combination was evaluated twice, using the block mean and the block median, resulting in 120 bootstrap comparisons in total (6 indices x 5 months x 2 years x 2 block summaries (mean and median)), or 60 comparisons per year.

At pixel level, using all valid pixels extracted from GEOGLAM ground-truth polygons, intercrop values were higher than monocrop values in 52 of 60 comparisons based on group means and in 48 of 60 comparisons based on group medians. The KS test was significant in 59 of 60 comparisons, with a median absolute Cliff’s delta of 0.088, indicating consistent but small distributional differences. When, each GEOGLAM polygon was treated as the observational unit, statistical significance decreased substantially: the polygon-mean analysis was significant in 10 of 60 comparisons, and the polygon-median, with intercrop higher in 43 of 60 comparisons based on group means. This indicates that the pixel-level tests likely overstate statical significance, but the main direction of the monocrop-intercrop contrast remains stable.

The spatial block-bootstrap analysis provided an additional check on spatial dependence in the mapped pixel-level classification results (Table 10). Using the 4 km grid, intercrop values remained higher than monocrop values in 95 of 120 comparisons. The 95% block-bootstrap confidence interval excluded zero in 78 of 120 comparisons. The used expression, “excluded zero” means that the 95% confidence interval around the intercrop-minus-monocrop difference did not include the value zero, which represents no difference between the two crop systems. When the interval was entirely above zero, the result indicated a consistent intercrop advantage after block-level resampling. When the interval included zero, the result can be considered as uncertain. Therefore at 4 km block size, intercrop values remained higher, while the magnitude of the effect was small and pixel-level p-values should not be interpreted as fully independent evidence, the combined pixel-, field- and block-level analyses support a consistent tendency for maize intercrop to show higher vegetation-index values than maize monocrop maize systems.

Overall, these sensitivity analyses show that accounting for spatial dependence reduced the strength of statistical significance but did not reverse the main direction of the monocrop-intercrop contrast. Therefore, the classified pixel-level maps were retained for the full vegetation-stress and IAS assessment, while the field-level and 4 km block-bootstrap analyses were used as conservative robustness checks.

3.5 Analysis of the resilience indicator: intercrop advantage score IAS process

The resilience indicator was summarized for each vegetation index (ENS, VCI, GCI, MCI) and season (2021 and 2023) using the seasonal StressScore difference (Δ), its 95% confidence interval, the median monthly standard deviation, the index-specific uncertainty threshold (τ_i) and the consistency. These results were presented in Table 7. Positive Δ indicate higher vegetation health for intercropping, while the uncertainty threshold (τ_i) defines the minimum difference required to be considered biophysically meaningful. Consistency (%) represents how often the intercrop advantage persisted across the months of the season.

Across all vegetation indices, the sign of Δ was constantly positive, confirming that intercropping maintained better canopy condition than monocropping in both years. However, the magnitude of the advantage differed noticeably between 2021 and 2023. In 2021, the intercrop advantage was present but moderate, VCI and ENS recorded the largest differences Δ, exceeding twice their uncertainty thresholds (2τ_i) with 80% consistency, signifying a strong advantage. GCI, Δ = 0.0122, also indicated a positive effect but lower consistency 60%, corresponding to a moderate advantage. The weakest signal was observed for MCI Δ = 0.0122 higher than 2τ but exhibited only 40% consistency, which means that it is classified as inconclusive or mixed. In 2023, the contrast between cropping systems intensified noticeably. All four indices showed strong and statistically robust advantages, with Δ values ranging between 0.023 and 0.0385, approximately two to three times higher than in 2021 and all of them twice higher than their uncertainty thresholds (2τ_i). These increases were accompanied by near-perfect temporal consistency (80-100%) and relative uncertainty τ_i ranged between 0.00109 and 0.0018. The ensemble index ENS and VCI followed similar patterns and exhibited the highest combination of Δ and consistency. GCI and MCI also showed strong performance with consistency equal to 80%. Overall, these results showed that intercropped maize system outperformed monocrops in vegetation conditions and resilience. The combination of Δ, τ and consistency within the results interpretation indicates that these differences are not random noise but reflect consistent physiological benefits of intercropping.

3.6 Sensitivity analysis of the intercrop advantage score

The sensitivity analysis (Table 8) showed that the IAS interpretation was highly stable for 2023, where VCI, GCI, MCI and the ensemble index (ENS) were classified as strong intercrop advantage in 100% of the tested scenarios. In 2021, VCI, was also fully stable, and IAS remained classified as strong intercrop advantage in 88.9% of scenarios. In contrast, the 2021 GCI and MCI results were less stable and were mainly classified as inconclusive. Importantly, none of the tested parameter scenarios produced a monocrop advantage. These results indicate that the IAS framework was robust for identifying intercrop advantage in 2023 and for VCI and ENS in 2021, while also showing where the interpretation should remain cautious.

4 Discussion

A number of discussion points emerge from the study: i. assessment of the overall robustness of the results; ii. relationship of the results to previous work; iii. future work suggested by the current study. We address each of these in turn.

Classification uncertainty may influence the vegetation stress and IAS results because all subsequent analyses depend on the correct separation of maize monocrop and maize intercrop pixels. Misclassification between the two crop-system classes would mix the vegetation-index distributions used to estimate stress categories. Such errors could also affect the magnitude of the Intercrop Advantage Score by either diluting of exaggerating the observed difference between monocrop and intercrop systems. Because confusion occurred in both directions, random classification error would most likely reduce the apparent contrast between systems. We therefore interpret the stress and IAS results conditional on the reliability of the crop-system classification and report classification accuracy, class-specific metrics and confusion matrices to make this uncertainty explicit.

Integrating vegetation indices from multiple sensors can be considered as strategy in remote sensing for agricultural stress monitoring, particularly in drought assessment. Several remote-sensing approaches have been proposed for crop-condition and stress monitoring, ranging from high-resolution UAV-based observation for crop growing status (Zheng et al., 2021) to multi-sensor combinations, such us, MODIS, Landsat and Sentinel products that enhance both temporal coverage and spatial detail (Li et al., 2025a; Liepa et al., 2024). However, these multi-sensor approaches often involve methodological challenges: cross-sensor harmonization requires careful calibration due to differences in spectral response, spatial resolution and revisit time (Sun et al., 2021). These complexities can introduce inconsistencies, particularly in fragmented smallholder landscapes. That is why, in this study, we deliberately avoided multi-sensor fusion and instead focused on a unified, biophysically grounded approach using only Sentinel-2 surface reflectance data. We combined three vegetation indices: VCI (derived from NDVI), MCI (from NDMI) and GCI (from GNDVI), each targeting a distinct biophysical response: general canopy condition, water stress, and chlorophyll/nutrient-related stress, respectively. By integrating these into the composite ENS index, we aimed to comprehensively capture abiotic and biotic stress. This approach ensured consistent spatial resolution (10m) reducing uncertainty associated with data fusion and placing the emphasis specifically on vegetation biophysical responses to stress.

To evaluate resilience to vegetation-stress of those two assessed crop-systems, we introduced the Intercrop Advantage Score (IAS), a composite framework that goes beyond simple comparisons. While past studies have compared cropping systems resilience using yield anomalies (Li et al., 2019), our method is based on a standardized score that incorporates the magnitude (Δ), temporal consistency and statistical uncertainty of vegetation condition differences.

Recent work in East Africa reported agronomic benefits of mixed cropping systems, including improved soil properties and higher maize yield compared with monocropping (Hashakimana et al., 2024). More broadly, our findings are consistent with agronomic literature showing that intercropping can improve crop-system performance through resource complementarity and higher resource-use efficiency (Brooker et al., 2015; Mao et al., 2015), increase yield stability and reduce production risk (Raseduzzaman and Jensen, 2017), and strengthen resilience under moisture variability in African farming systems (Adam et al., 2025). Our results also align with remote-sensing studies showing that vegetation indices can capture crop condition and stress dynamics, including VCI for vegetation condition and drought stress, MCI for canopy water-related stress, and GCI for greenness or chlorophyll-related stress. However, our study differs from yield-based resilience studies because it does not directly measure productivity, biomass or farmer outcomes. Instead, IAS should be interpreted as a remote-sensing-based proxy for relative vegetation-stress resilience. This distinction is important in heterogeneous smallholder landscapes, where mixed crops, small field sizes, trees, bare soil and household compounds can complicate both crop classification and field-level interpretation. Therefore, while our results support a consistent tendency for maize intercrop systems to show more favourable vegetation-stress profiles than maize monocrop systems, the practical agronomic significance of these differences remains to be evaluated with yield, biomass and farmer-level data. Also, the differences between maize monocrop and intercropped maize were persistent across months, indices and seasons, but distributional and statistical tests showed that their numerical magnitude was often small.

A further caveat includes the fact that the analysis was limited to the 2021 and 2023 long-rain seasons. Whilst the consistency of the intercrop advantage across the two analysed seasons, together with agreement with previous agronomic intercropping studies, suggests that similar patterns may plausibly occur in other seasons, a key question for future work is the generalisability of these findings. Overall, the consistency of results across indices and tested years suggests that the observed intercropping advantage reflects underlying biophysical processes rather than site-specific observation. This indicates that the IAS framework may be transferable to similar crop systems and study areas. Although IAS used normalized vegetation indices and therefore has the potential to be applied in other regions, its crop-system classification and decision thresholds should be locally checked before cross-regional use. Recent domain-adaptation and test-time adaption studies in remote sensing show that model performance can change when classifiers are transferred across regions because spectral conditions can change when classifiers are transferred across regions because spectral conditions, class distribution and landscape structure differ between domains (Li et al., 2025b; Liang et al., 2025). In fragmented smallholder landscapes, transferability and field-level validation are also constrained by the difficulty of deriving complete agricultural parcel boundaries from remote-sensing imagery, a limitation highlighted in recent reviews of agricultural parcel and boundary delineation (Zheng et al., 2026). We therefore present IAS as a transferable framework that requires local calibration and validation, rather than as a fixed model that can be applied unchanged in all agroecological zones and seasons.

In sum, future studies combining remote sensing indicators with ground-based observations, including yield data and farmer-reported outcomes, will be essential to establish the practical significance of IAS values. Field studies and/or farmer-perceptions could be used to evaluation the practical significance of the observed IAS differences. Such integration would relate the IAS categories or thresholds (e.g. strong, moderate, inconclusive) against real-world impacts and thereby strengthening its utility as decision-support tool for agricultural management and policy.

5 Conclusion

This study has demonstrated how remote sensing can be used to show that maize-based intercropping systems showed a more favourable vegetation-stress profile than maize monocrop in smallholder landscapes in western Kenya during 2021 and 2023 long-rain seasons. Across multiple vegetation indices, intercropped maize generally maintained healthier canopy conditions and lower stress levels than monocropped maize. By combining crop systems classification output with multi-index vegetation stress monitoring, the analysis demonstrates that intercropped fields maintain healthier canopy conditions during the long rain growing season. The IAS framework provides a structured methodology to detect and assess differences in vegetation stress between cropping systems. The IAS builds on remote sensing derived stress signals and takes into consideration seasonal variability through bootstrap-based uncertainty thresholds. The study also shows the importance of using classification approaches adapted to complex smallholder landscapes.

Overall, The IAS provides a solid first step towards a transferable, decision-relevant framework for quantifying resilience using satellite data. Our aspiration is for it to be developed into a valuable tool for agricultural monitoring, climate-smart intervention assessment and policy-relevant evaluation of intercropping strategies in smallholder farming systems. The approach can be extended to other regions and cropping systems where resilience under climate stress is a concern.

References

• 1

  AbdulridhaJ.MinA.RouseM. N.KianianS.IslerV.YangC. (2023). Evaluation of stem rust disease in wheat fields by drone hyperspectral imaging. Sensors23, 4154. doi: 10.3390/s23084154

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

• 2

  AdamA. M.GillerK. E.RusinamhodziL.RascheF.KoomsonE.MarohnC.et al. (2025). Enhancing the resilience of intercropping systems to changing moisture conditions in Africa through the integration of grain legumes: a meta-analysis. Field Crops Res.321, 109663. doi: 10.1016/j.fcr.2024.109663

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

• 3

  BrookerR. W.BennettA. E.CongW.-F.DaniellT. J.GeorgeT. S.HallettP. D.et al. (2015). Improving intercropping: a synthesis of research in agronomy, plant physiology and ecology. New Phytol.206, 107–117. doi: 10.1111/nph.13132

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

• 4

  EeswaranR.NejadhashemiA. P.CoutoF.SaraviB. (2021). Evaluating the applicability of soil moisture-based metrics for gauging the resiliency of rainfed agricultural systems in the Midwestern United States. Soil Tillage Res.205, 104818. doi: 10.1016/j.still.2020.104818

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

• 5

  GitelsonA. A.KaufmanY. J.MerzlyakM. N. (1996). Use of a green channel in remote sensing of global vegetation from EOS-MODIS. Remote Sens. Environ.58, 289–298. doi: 10.1016/S0034-4257(96)00072-7

  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  HashakimanaL.TessemaT.NiyitangaF.WatabajiM. D.GeleteT. B.HirwaH. (2024). Monocropping vs mixed cropping systems under a changing climate: smallholder farmers' perceptions and farm profitability in Eastern Rwanda. Environ. Sustain. Indic.24, 100527. doi: 10.1016/j.indic.2024.100527

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

• 7

  HuangJ.XiaoG.LiX.ZhengJ.ZengY.SuW.et al. (2026). Progress and perspectives of crop type mapping with remote sensing: a review. IEEE Geosci. Remote Sens. Mag.14, 129–162. doi: 10.1109/MGRS.2025.3648119

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

• 8

  IbrahimE. S.RufinP.NillL.KamaliB.NendelC.HostertP. (2021). Mapping crop types and cropping systems in Nigeria with Sentinel-2 imagery. Remote Sens.13, 3523. doi: 10.3390/rs13173523

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

• 9

  JacksonT. J.ChenD.CoshM.LiF.AndersonM.WalthallC.et al. (2004). Vegetation water content mapping using Landsat data derived normalized difference water index for corn and soybeans. Remote Sens. Environ.92, 475–482. doi: 10.1016/j.rse.2003.10.021

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

• 10

  Jafari JozaniH.ThielM.Abdel-RahmanE. M.RichardK.LandmannT.SubramanianS.et al. (2022). Investigation of Maize Lethal Necrosis (MLN) severity and cropping systems mapping in agro-ecological maize systems in Bomet, Kenya utilizing RapidEye and Landsat-8 imagery. Geol. Ecol. Landsc.6, 125–140. doi: 10.1080/24749508.2020.1761195

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

• 11

  JiJ.XuM.WangS.CaoC.ZhangX.TianF.et al. (2025). Analysis of spatial pattern of vegetation resilience and influencing factors in Hubei Province based on long time series remote sensing data. Environ. Sustain. Indic.27, 100742. doi: 10.1016/j.indic.2025.100742

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

• 12

  JiaoW.ZhangL.ChangQ.FuD.CenY.TongQ. (2016). Evaluating an enhanced Vegetation Condition Index (VCI) based on VIUPD for drought monitoring in the continental United States. Remote Sens.8, 224. doi: 10.3390/rs8030224

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

• 13

  KoganF. N. (1990). Remote sensing of weather impacts on vegetation in non-homogeneous areas. Int. J. Remote Sens.11, 1405–1419. doi: 10.1080/01431169008955102

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

• 14

  KoganF. N. (1995). Droughts of the late 1980s in the United States as derived from NOAA polar-orbiting satellite data. Bull. Am. Meteorol. Soc76, 655–668. doi: 10.1175/1520-0477(1995)076<0655:DOTLIT>2.0.CO;2

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  KoganF. N.GitelsonA. A.ZakarinE.SpivakL.LebedL. (2003). AVHRR-based spectral vegetation index for quantitative assessment of vegetation state and productivity: calibration and validation. Photogramm. Eng. Remote Sens.69, 899–906. doi: 10.14358/PERS.69.8.899

  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  LiM.PetersonC. A.TautgesN. E.ScowK. M.GaudinA. C. M. (2019). Yields and resilience outcomes of organic, cover crop, and conventional practices in a Mediterranean climate. Sci. Rep.9, 12283. doi: 10.1038/s41598-019-48747-4

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

• 17

  LiM.WangP.TanseyK.SunY.GuoF.ZhouJ. (2025a). Improved field-scale drought monitoring using MODIS and Sentinel-2 data for Vegetation Temperature Condition Index generation through a fusion framework. Comput. Electron. Agric.234, 110256. doi: 10.1016/j.compag.2025.110256

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

• 18

  LiQ.ZhangY.ZhengJ.ZhangY.HuangJ.FuH. (2025b). Boosting universal domain adaptation in remote sensing with dual-classifiers consistency discrimination and cross-domain feature mixup. IEEE Trans. Geosci. Remote Sens.63, 1–15. doi: 10.1109/TGRS.2025.3571747

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

• 19

  LiangY.CaoS.ZhengJ.ZhangX.HuangJ.FuH. (2025). Low saturation confidence distribution-based test-time adaptation for cross-domain remote sensing image classification. Int. J. Appl. Earth Obs. Geoinf.139, 104463. doi: 10.1016/j.jag.2025.104463

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

• 20

  LiepaA.ThielM.TaubenböckH.Steffan-DewenterI.AbuI.-O.Singh DhillonM.et al. (2024). Harmonized NDVI time-series from Landsat and Sentinel-2 reveal phenological patterns of diverse, small-scale cropping systems in East Africa. Remote Sens. Appl. Soc Environ.35, 101230. doi: 10.1016/j.rsase.2024.101230

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

• 21

  LiuW. T.KoganF. N. (1996). Monitoring regional drought using the Vegetation Condition Index. Int. J. Remote Sens.17, 2761–2782. doi: 10.1080/01431169608949106

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

• 22

  LykhovydP. V.ShariiV. O. (2024). Normalised difference moisture index in water stress assessment of maize crops. Agrology7, 21–26. doi: 10.32819/202403

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

• 23

  MahlayeyeM.DarvishzadehR.NelsonA. (2024). Characterising maize and intercropped maize spectral signatures for cropping pattern classification. Int. J. Appl. Earth Obs. Geoinf.128, 103699. doi: 10.1016/j.jag.2024.103699

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

• 24

  MalekiS.BaghdadiN.BazziH.DantasC. F.IencoD.NasrallahY.et al. (2024). Machine learning-based summer crops mapping using Sentinel-1 and Sentinel-2 images. Remote Sens.16, 4548. doi: 10.3390/rs16234548

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

• 25

  MaoL.ZhangL.ZhangS.EversJ. B.van der WerfW.WangJ.et al. (2015). Resource use efficiency, ecological intensification and sustainability of intercropping systems. J. Integr. Agric.14, 1542–1550. doi: 10.1016/S2095-3119(15)61039-5

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  MusafiriC. M.KiboiM.MachariaJ.Ng'etichO. K.KosgeiD. K.MuliangaB.et al. (2022). Smallholders' adaptation to climate change in Western Kenya: considering socioeconomic, institutional and biophysical determinants. Environ. Chall.7, 100489. doi: 10.1016/j.envc.2022.100489

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

• 27

  RadóczL.SzabóA.TamásA.IllésÁ.BojtorC.RagánP.et al. (2023). Investigation of the detectability of corn smut fungus (Ustilago maydis DC. Corda) infection based on UAV multispectral technology. Agronomy13, 1499. doi: 10.3390/agronomy13061499

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

• 28

  RaseduzzamanM.JensenE. S. (2017). Does intercropping enhance yield stability in arable crop production? A meta-analysis. Eur. J. Agron.91, 25–33. doi: 10.1016/j.eja.2017.09.009

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

• 29

  RedheadJ. W.OliverT. H.WoodcockB. A.PywellR. F. (2020). The influence of landscape composition and configuration on crop yield resilience. J. Appl. Ecol.57, 2180–2190. doi: 10.1111/1365-2664.13722

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

• 30

  ShenJ.WangH.TarequeH. (2025). Toward resilience in broadacre agriculture: a methodological review of remote sensing in crop productivity, phenology, and environmental stress detection. Remote Sens.17, 3886. doi: 10.3390/rs17233886

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

• 31

  ShilomboleniH.EpsteinG.MansinghA. (2024). Building resilience in Africa's smallholder farming systems: contributions from agricultural development interventions - a scoping review. Ecol. Soc29, 22. doi: 10.5751/ES-15373-290322

  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  SunL.GaoF.XieD.AndersonM.ChenR.YangY.et al. (2021). Reconstructing daily 30 m NDVI over complex agricultural landscapes using a crop reference curve approach. Remote Sens. Environ.253, 112156. doi: 10.1016/j.rse.2020.112156

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

• 33

  WangS.ZhangS.ZhouY.ZhaoX.LiR.LinX.et al. (2025). Vegetation disturbance and recovery assessment through the synergistic effects of albedo and vegetation index: evidence from China's arid regions. Ecol. Indic.178, 113856. doi: 10.1016/j.ecolind.2025.113856

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

• 34

  WójtowiczA.PiekarczykJ.CzerneckiB.RatajkiewiczH. (2021). A random forest model for the classification of wheat and rye leaf rust symptoms based on pure spectra at leaf scale. J. Photochem. Photobiol. B. Biol.223, 112278. doi: 10.1016/j.jphotobiol.2021.112278

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

• 35

  WuJ.LiangS. (2020). Assessing terrestrial ecosystem resilience using satellite Leaf Area Index. Remote Sens.12, 595. doi: 10.3390/rs12040595

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

• 36

  XuC.QuJ. J.HaoX.CoshM. H.ZhuZ.GutenbergL. (2020). Monitoring crop water content for corn and soybean fields through data fusion of MODIS and Landsat measurements in Iowa. Agric. Water Manage.227, 105844. doi: 10.1016/j.agwat.2019.105844

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

• 37

  ZampieriM.WeissteinerC. J.GrizzettiB.ToretiA.van den BergM.DentenerF. (2020). Estimating resilience of crop production systems: from theory to practice. Sci. Total Environ.735, 139378. doi: 10.1016/j.scitotenv.2020.139378

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

• 38

  ZhengJ.FuH.LiW.WuW.YuL.YuanS.et al. (2021). Growing status observation for oil palm trees using unmanned aerial vehicle (UAV) images. ISPRS J. Photogramm. Remote Sens.173, 95–121. doi: 10.1016/j.isprsjprs.2021.01.008

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

• 39

  ZhengJ.YeZ.WenY.HuangJ.ZhangZ.LiQ.et al. (2026). A comprehensive review of agricultural parcel and boundary delineation from remote sensing images: recent progress and future perspectives. IEEE Geosci. Remote Sens. Mag.14, 206–237. doi: 10.1109/MGRS.2026.3658493

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

• 40

  ZhouH.ZhouG.SongX.HeQ. (2022). Dynamic characteristics of canopy and vegetation water content during an entire maize growing season in relation to spectral-based indices. Remote Sens.14, 584. doi: 10.3390/rs14030584

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