Embedding NDRE Trajectories into Contrastive Learning for Label-Free, Physiology-Aware Crop-Stress Staging and DSS Outputs

Shafqaat Ahmad

Source-linked AI summary

Embedding NDRE Trajectories into Contrastive Learning for Label-Free, Physiology-Aware Crop-Stress Staging and DSS Outputs

Shafqaat Ahmad

arXiv:2608.25888v1Aug 26, 2026cs.CV

TL;DR

Crop-stress progression remains difficult to capture with conventional vegetation-index thresholds and image-based clustering, limiting interpretable decision support. EigenCL uses physiology-guided contrastive learning on Sentinel-2 NDRE trajectories and produces four coherent stress stages that generalize across Iowa and Nebraska while aligning with agronomic validation data.

Problem
Conventional vegetation-index thresholds and image-based clustering have limited ability to capture stress progression and interpretability for agricultural decision support.
Method
EigenCL anchors label-free contrastive similarity to eigenvector structure in NDRE temporal trajectories, emphasizing chlorophyll dynamics for cross-region stress staging.
Results
EigenCL generalized from Iowa 2020 to Nebraska 2023 without retraining, producing four physiologically coherent clusters validated against soil moisture, drought maps, and county yield records.
Takeaways & Limitations
EigenCL provides interpretable stress maps, alerts, and regional indices that support field-scale and broader crop-stress decision monitoring.
Takeaways & Limitations
The study covers only maize in Iowa and Nebraska, while field-scale quantification of yield penalties remains open.

Abstract

from arXiv · show

Timely detection of crop stress is critical for sustaining yields under increasing drought frequency, yet conventional vegetation index thresholds or image-based clustering often fail to capture stress progression, limiting their value for farm decision-making. To address this gap, we present EigenCL, a physiology-guided contrastive learning framework that stages crop stress from Sentinel-2 NDRE trajectories, with the goal of providing interpretable and transferable stress diagnostics for decision support systems (DSS). EigenCL was trained on 10,000 maize NDRE patches from drought-affected Iowa fields in 2020 and tested on Nebraska fields in 2023 without retraining, with validation incorporating soil-moisture records, U.S. Drought Monitor maps, and county-level yield statistics. The model produced four physiologically coherent stress clusters (Healthy, Mild, Moderate, Severe), significantly outperforming baselines including K-Means, SimCLR, ProtoCLR, and an ablation model (Silhouette = 0.748, DBI = 0.35, CHI = 49,624). Clusters aligned with maize growth stages, with severe stress peaking around tasseling-silking (VT-R1), a stage known to drive yield loss; moreover, EigenCL clusters correlated with soil moisture at 0-14-day lags (rho up to 0.72) and matched yield anomalies in drought-affected counties. By embedding NDRE trajectory dynamics into contrastive learning, EigenCL enables early stress alerts and interpretable DSS outputs (e.g., heatmaps, scouting priorities, regional risk indices), extending beyond single-date NDRE thresholds and supporting scalable monitoring for climate-smart agronomy.

1. Introduction

The introduction motivates timely, physiologically meaningful crop-stress monitoring and presents EigenCL, an unlabeled, trajectory-based contrastive framework designed for interpretable stress clustering and decision support. It evaluates cross-year, cross-region generalization and validates clusters using soil moisture, drought maps, and county yield records.

Motivation: Up to 50% yield reductions can result from maize stress during tasseling–silking (VT–R1), making timely monitoring important for scouting, irrigation scheduling, and risk assessment.
Problem: NDRE captures canopy chlorophyll and responds earlier than NDVI, but threshold-based or static values often fail to reflect stress progression.
Contribution: EigenCL anchors contrastive similarity to the eigenvector structure of NDRE temporal trajectories, emphasizing chlorophyll dynamics rather than static appearance.
Evaluation: EigenCL is trained on Iowa 2020 NDRE patches and tested directly on Nebraska 2023 without retraining to assess cross-year, cross-region generalization.
Evaluation: Clusters are validated against soil moisture, U.S. Drought Monitor maps, and county yield records to assess whether stress trajectories are meaningful rather than noise.
Decision Support: The resulting stress detection is positioned to support decision support systems for field-level and regional-scale monitoring.Applications include scouting, irrigation, nutrient planning, extension dashboards, crop insurance, and policy evaluation.

2. Materials and Methods

The study uses Sentinel-2 NDRE trajectories from drought-affected maize fields to construct a physiology-aware contrastive learning framework. EigenCL derives biological similarity from temporal stress patterns, trains on Iowa 2020, and evaluates transferability in Nebraska 2023 and validation datasets.

Datasets: Iowa 2020 supplied the main training dataset, while Nebraska 2023 tested cross-region transferability and Iowa 2020, 2022, and 2023 supported soil-moisture validation.The dataset summary identifies Iowa 2020 as the main training set, Nebraska 2023 as the transferability test, and the Iowa subsets as validation data.
Data acquisition: 10,000 Iowa 2020 and 3,500 Nebraska 2023 NDRE patches were sampled from 100 × 100-pixel Sentinel-2 images at five-day intervals during July–September.Sentinel-2 imagery had 10 m resolution, and NDRE trajectories captured temporal chlorophyll dynamics rather than static values.
EigenCL framework: EigenCL replaces visual augmentation and cosine-based similarity with biologically grounded similarity derived from NDRE temporal patterns.RBF similarity matrices represented five-date NDRE trajectories, and principal-eigenvector weights served as continuous proxies for stress severity.
Cross-region evaluation: 3,500 Nebraska 2023 patches were evaluated with the pretrained Iowa model without fine-tuning, using Iowa-derived embeddings and cluster thresholds.Nebraska experienced milder drought and later thermal-stress onset, providing an external generalizability evaluation under the same five-date protocol.
Optimization and evaluation: The final configuration was λ = 4.0, τ = 0.075, σ = 0.5, and m = 0.2, selected through grid search using clustering and physiological interpretability criteria.Selection considered Silhouette Score, Davies–Bouldin Index, Calinski–Harabasz Index, contrastive-loss convergence, and NDRE-based centroid stability.

3. Results and Discussion · 3.1 Physiological and Agronomic Interpretation of Clusters

EigenCL produced physiologically interpretable maize stress clusters by organizing NDRE trajectories into separated, monotonic stress gradients that reflected temporal dynamics rather than static spectral magnitude. The clusters aligned with maize growth stages, including severe-stress declines around tasseling-silking and stress windows associated with yield-relevant outcomes.

3. Results and Discussion: 0.748 Silhouette, 0.350 DBI, and 49,624.06 CHI showed that EigenCL outperformed the evaluated baselines.The comparison used an ablation model, K-Means, and SimCLR with a shared ResNet-50 backbone.
3.1 Physiological and Agronomic Interpretation of Clusters: NDRE correlated with the guiding eigenvector at r = 0.95, supporting NDRE as the biologically relevant overlay for interpreting EigenCL embeddings.EigenCL clusters formed a clear monotonic gradient from low to high stress with internally consistent NDRE values.
3.1 Physiological and Agronomic Interpretation of Clusters: SimCLR produced weak stress structure, with 0.416 Silhouette Score, 0.724 DBI, and 2,251.86 CHI, alongside high NDRE-overlay entropy.The ablation model showed fractured clusters, whereas K-Means was compact but biologically inconsistent because its clusters contained heterogeneous NDRE values.
3.1 Physiological and Agronomic Interpretation of Clusters: Cluster 1 remained high in NDRE while Cluster 3 declined sharply, demonstrating that EigenCL captured trajectory shape and stress dynamics.The temporal separation indicated that clustering was not driven by static NDRE magnitude.
3.1 Physiological and Agronomic Interpretation of Clusters: Severe-stress NDRE declines began in mid-July, coinciding with maize tasseling and silking (VT–R1).Rapid vegetative growth generally occurs at V6–V8 in late June, while grain fill occurs at R3–R5 in August.
3.1 Physiological and Agronomic Interpretation of Clusters: Moderate-stress declines aligned with late vegetative stress onset, while gradual declines into grain fill were consistent with reduced kernel weight under late-season stress.These correspondences positioned EigenCL as capturing physiologically meaningful stress windows linked to yield outcomes.

3.2 Sensitivity to Eigenvector Selection · 3.3 Downstream Utility

EigenCL depends on the principal eigenvector as its sole supervisory signal, while its frozen embeddings support strong downstream stress classification with both nonlinear and linear classifiers.

3.2 Sensitivity to Eigenvector Selection: Replacing the principal eigenvector with lower-variance components clearly reduced clustering and classification performance.This sensitivity result supports retaining the principal eigenvector during training.
3.2 Sensitivity to Eigenvector Selection: 76.3% of variance was captured by the principal eigenvector, which was strongly correlated with NDRE.These properties motivated its use as the biologically meaningful supervisory signal.
3.2 Sensitivity to Eigenvector Selection: EigenCL retained the principal eigenvector as its sole supervisory signal.The choice followed the observed performance drop from substituting lower-variance components.
3.3 Downstream Utility: 89.1% accuracy was achieved by k-Nearest Neighbors on frozen EigenCL embeddings using a 70/30 split.The classifier was trained without fine-tuning and obtained F1: 0.87.
3.3 Downstream Utility: 85.2% accuracy was achieved by Logistic Regression on the same frozen embeddings and split.The classifier was trained without fine-tuning and obtained F1: 0.82.
3.3 Downstream Utility: The embeddings encoded patterns relevant to stress classification for both nonlinear and linear separability.This conclusion was drawn from the k-Nearest Neighbors and Logistic Regression results.

3.4 Statistical Confirmation of Cluster Separation

Statistical tests confirmed that EigenCL clusters represent distinct physiological regimes, with significant NDRE differences across and between all clusters. The pretrained model also generalized to Nebraska without fine-tuning, retaining stress-aware clustering and NDRE interpretability despite modest metric declines.

Statistical confirmation: F = 37,950.39 with p < 0.0001 confirmed significant NDRE mean variation across EigenCL clusters.A one-way ANOVA supported separation among the detected clusters.
Statistical confirmation: p < 0.0001 for every Tukey HSD pairwise comparison validated each cluster as a distinct physiological regime.The largest NDRE difference was – 0.398 between Clusters 0 and 2, while the smallest was – 0.076 between Clusters 1 and 3.
Cross-region generalization: Without fine-tuning, EigenCL was directly applied to Nebraska datasets and retained stress-aware clustering and NDRE interpretability.The transfer evaluation involved environmental differences between regions.
Cross-region generalization: Despite a modest decline in clustering metrics, robust trends supported EigenCL’s generalization across time and geography.The Nebraska results were described as consistent with the model’s ability to generalize across regional variability.

3.5 Validation with soil moisture, drought maps, and yields · 3.5 Lagged NDRE–soil moisture correlation

Validation used U.S. Drought Monitor maps and county-level yields to identify drought-stressed Iowa seasons, while lagged analyses showed NDRE tracked soil moisture within approximately two weeks. Peak correlations varied by year, reaching rho = 0.721 in 2022.

3.5 Validation with soil moisture, drought maps, and yields: A ground-truth dataset identified drought-stressed Iowa zones across the 2020, 2022, and 2023 seasons using U.S. Drought Monitor maps.Sampling was anchored to three Iowa Environmental Mesonet stations.
3.5 Validation with soil moisture, drought maps, and yields: Buena Vista in 2020, Monona in 2022, and Buchanan in 2023 showed noticeably lower county-level yields than surrounding seasons.Yield data came from USDA Quick Stats and helped confirm these years as stress years.
3.5 Lagged NDRE–soil moisture correlation: Lagged agreement correlated NDRE at day t with 12-inch soil moisture measured 0–14 days earlier using Spearman correlations, bootstrap 95% confidence intervals, and permutation p-values.The best lag was defined as the lag with the largest absolute Spearman correlation.
3.5 Lagged NDRE–soil moisture correlation: rho = 0.525 peaked at a 6-day lag in 2020, with 95% CI [0.319, 0.667] and p = 0.002.The corresponding Pearson correlation was r = 0.514.
3.5 Lagged NDRE–soil moisture correlation: rho = 0.721 peaked at a 14-day lag in 2022, with 95% CI [0.519, 0.831] and p = 0.002.The corresponding Pearson correlation was r = 0.608.
3.5 Lagged NDRE–soil moisture correlation: rho = 0.608 peaked at zero lag in 2023, with 95% CI [0.444, 0.734] and p = 0.002.The corresponding Pearson correlation was r = 0.524.
3.5 Lagged NDRE–soil moisture correlation: Across years, NDRE tracked soil moisture within approximately zero to two weeks, while the exact lag varied by year.Peak lags were 6 days in 2020, 14 days in 2022, and 0 days in 2023.

3.6 Classification agreement (ARI) between NDRE clusters and soil-derived · 3.7 EigenCL as a Decision Support Tool · 3.8 Using Clusters for DSS Rules

EigenCL clusters showed strongest soil-label agreement after year-specific lagging, especially in 2023, and translated into field- and regional-scale DSS outputs. Cluster-based rules specify escalating responses from monitoring to urgent scouting and irrigation, while weighted aggregation tracks county-level weekly risk.

3.6 Classification agreement (ARI) between NDRE clusters and soil-derived: Soil-derived stress classes served as ground truth, with soil shifted by lag d and labels assigned using per-year quartiles of 12-inch VWC.Lower moisture indicated higher stress, and ARI was computed between lagged soil labels and EigenCL NDRE labels alongside a same-day baseline.
3.6 Classification agreement (ARI) between NDRE clusters and soil-derived: 2023 achieved the highest agreement at 2 days, with ARI = 0.371 versus lag 0 = −0.006.The best lags were 6 days in 2020, with ARI = 0.120 versus lag 0 = 0.013, and 8 days in 2022, with ARI = 0.155 versus lag 0 = 0.014.
3.7 EigenCL as a Decision Support Tool: At field scale, Healthy, Mild, Moderate, and Severe clusters can be displayed as heatmaps to identify emerging and severe stress.These maps support decisions about scouting, irrigation, and fertilizer adjustment.
3.7 EigenCL as a Decision Support Tool: At regional scale, cluster-area shares can be summarized into a stress risk index for extension services, cooperatives, or other decision systems.The regional aggregation complements field-level visualization with an area-based summary.
3.8 Using Clusters for DSS Rules: Severe stress affecting ≥ 20% of field patches triggers urgent scouting and, where feasible, supplemental irrigation within 24–48 h.The rule converts the Severe cluster into an immediate operational alert.
3.8 Using Clusters for DSS Rules: Moderate stress, or Severe+Moderate ≥40%, prioritizes scouting, soil-moisture and valve checks, and consideration of variable-rate input plans.This rule escalates action when moderate or combined moderate-to-severe coverage is substantial.
3.8 Using Clusters for DSS Rules: Mild stress at 20–40% with limited spread warrants weekly monitoring and follow-up, whereas a majority Healthy area requires routine monitoring only.These lower-severity rules avoid immediate intervention while retaining surveillance.
3.8 Using Clusters for DSS Rules: A 0–3 weighted score assigns Healthy=0, Mild=1, Moderate=2, and Severe=3, averaged over fields or pixels to track county-level weekly stress risk.The rules therefore produce operational signals at both farm and regional scales.

3.9 Linking Stress Clusters to Yield and Management Outcomes

EigenCL stress-cluster prevalence relates to county- and season-level maize yield anomalies, with Severe-dominated areas showing larger negative anomalies. The clusters also provide actionable DSS outputs for field management and regional monitoring.

Yield outcomes: Counties and seasons with more Severe clusters showed larger negative yield anomalies than those dominated by Healthy clusters.The relationship is consistent with maize physiology, in which stress around tasseling-silking (R1) strongly reduces yield.
Management outcomes: Early shifts into Moderate or Severe clusters can support irrigation responses during R1 where available and inform insurance decisions.The passage identifies early detection as a basis for timely management and risk-response actions.
DSS integration: EigenCL outputs support field-level scouting or irrigation rules and regional stress indices for extension dashboards, crop insurance, and policy monitoring.The DSS workflow converts four clusters—Healthy, Mild, Moderate, and Severe—into operational decision-support products.

4. Limitations and Future Work

EigenCL showed robustness across seasons and regions but remains limited in crop, geographic, and historical coverage, while field-scale yield-penalty quantification remains unresolved. Future work should broaden the framework and integrate additional signals.

Current limitations: EigenCL demonstrated robustness across seasons and regions, but the study covered only maize in Iowa and Nebraska.It did not yet cover a wider range of crops, agroecological zones, or longer historical periods.
Current limitations: Field-scale quantification of yield penalties remains an open step despite links between stress clusters, soil moisture, drought maps, and county-level yield anomalies.The existing validation established associations but not precise field-scale yield-penalty estimates.
Future work: Future work should extend EigenCL beyond the current study scope and integrate additional signals.The supplied passage introduces additional signals as a future direction but does not specify them.

5. Conclusion

EigenCL stages maize crop stress from NDRE trajectories using physiology-aware learning, producing four agronomically relevant clusters across Iowa and Nebraska. Its sensor-agnostic design supports scalable decision-support applications, while broader validation across crops, regions, and sensors remains a future priority.

Core contribution: EigenCL stages crop stress from NDRE trajectories as indicators of canopy chlorophyll dynamics.The framework embeds physiological processes into data-driven modeling for stress staging.
Validation: Four stress clusters aligned with soil-moisture patterns, drought monitor maps, and yield anomalies across Iowa (2020) and Nebraska (2023).These alignments supported the clusters’ physiological coherence and agronomic relevance.
Scalability: EigenCL is sensor-agnostic and adaptable to UAV and proximal data streams, supporting scalability across farms and landscapes.This adaptability extends the framework beyond NDRE inputs from the evaluated field settings.
Future validation: Validation across other crops, agroecological zones, and multi-sensor datasets is needed to consolidate EigenCL as a decision-support foundation.The proposed direction links remote sensing with sustainable and climate-smart agronomy.

Appendix A. Hyperparameter Grid Search Summary

Appendix A summarizes the hyperparameter tuning performed for the EigenCL framework.

Table A1 summarizes the hyperparameter tuning for EigenCL.
The appendix presents a consolidated summary of EigenCL’s hyperparameter grid search.
The reported tuning results are organized as Table A1.