A machine-learning model has grouped satellite observations of maize into four stress stages that the researchers say are easier to interpret than a single raw vegetation reading. The model, called EigenCL, produced clusters labelled Healthy, Mild, Moderate and Severe from changes in Sentinel-2 NDRE over time, according to an arXiv preprint.
The work is a computational model evaluation, not a test of farm management. It found associations between the satellite patterns and soil-moisture measurements, as well as links with county or seasonal yield anomalies, but it did not test whether using the system would improve irrigation, scouting, fertilizer choices or harvest results.
Reading the shape of crop stress
EigenCL was designed to learn the shape of an NDRE trajectory rather than rely only on how high or low the index was at one moment. It used the principal eigenvector of a similarity matrix built from NDRE time series as the main stress-trajectory signal guiding the learning process. In plain terms, the method compared how patches changed across the season and used the dominant shared pattern to organize them.
The researchers evaluated the model against K-Means, SimCLR, ProtoCLR and a version of the system without eigenvector guidance. The main Iowa training set contained 10,000 patches, while the listed Nebraska transfer-validation set contained 3,500 patches. The patches were sampled from maize fields, with additional 1,500-patch Iowa sets used for soil-moisture validation in 2020, 2022 and 2023.
On the Iowa data, EigenCL's reported clustering scores were a Silhouette Score of 0.748, a Davies-Bouldin Index of 0.350 and a Calinski-Harabasz Index of 49,624.06. Taken together, those were the strongest reported results among the compared models. The analysis reported no uncertainty intervals or inferential comparison test for these clustering metrics.
The sharpest warning appeared mid-season
The four labels were not presented as arbitrary numerical groups. The paper describes them as physiologically coherent, with the Severe group showing a sharp NDRE decline beginning in mid-July and coinciding with the tasseling-silking period. That timing gives the model's clusters a seasonal interpretation and supports the paper's proposed use of interpretable stress diagnostics.
A further check found that the principal eigenvector captured 76.3% of the variance. When lower-variance components replaced it, the paper reported poorer clustering and classification performance, although it did not give exact alternative performance values. This sensitivity result supports the authors' choice of the dominant trajectory signal, while leaving the size of the difference unclear.
The learned representations also supported two downstream classifiers. With the embeddings held fixed, k-nearest neighbors reached 89.1% accuracy and an F1 score of 0.87, compared with 85.2% accuracy and an F1 score of 0.82 for logistic regression. No confidence intervals were reported for these figures.
The clusters were separated by their reported average NDRE values as well. A one-way analysis of variance gave F = 37,950.39 with p less than 0.0001, and the reported Tukey HSD pairwise p-values were also below 0.0001. The paper notes that the displayed pairwise table is incomplete relative to its narrative.
A promising transfer test, with uneven agreement
The pretrained model was applied directly to Nebraska without fine-tuning. The authors reported that it retained stress-aware clustering and an interpretable NDRE pattern there, although clustering metrics declined modestly. Exact Nebraska scores and uncertainty intervals were not supplied, so the transfer result is qualitative rather than a complete external benchmark.
The soil-moisture comparison showed positive year-specific associations, but the strongest timing varied. In 2020, the reported Spearman correlation was 0.525 at a six-day lag, with a 95% confidence interval of 0.319 to 0.667 and p = 0.002. In 2022, it was 0.721 at a 14-day lag, with a 95% confidence interval of 0.519 to 0.831 and p = 0.002. In 2023, it was 0.608 on the same day, with a 95% confidence interval of 0.444 to 0.734 and p = 0.002.
A separate agreement measure, the adjusted Rand index, was higher at the reported best alignment lag than with same-day matching in each year. The values were 0.120 versus 0.013 in 2020, 0.155 versus 0.014 in 2022, and 0.371 versus -0.006 in 2023. The study did not report uncertainty intervals for these comparisons.
The authors also reported that counties or seasons with larger proportions of Severe clusters had larger negative yield anomalies than Healthy-dominated patterns. That is a directional association, not a calculated yield penalty for an individual cluster: the study did not quantify how much yield a field would lose after receiving a given label.
Useful maps are still a proposal
The paper proposes turning the four stress labels into field heatmaps for scouting, irrigation or fertilizer decisions, and into regional stress-risk indices. Those outputs are described as decision-support uses, not as tested products. The proposed rules and alerts were not evaluated prospectively for their effects on management or yield.
The evidence remains bounded by the data. The analysis covered maize in Iowa and Nebraska and did not evaluate broader crop types, agroecological settings, sensor types or longer historical periods. It also left field-scale calibration of cluster labels to yield penalties unresolved, and the Nebraska transfer assessment lacked a complete set of external metrics.
The findings suggest that a trajectory-based satellite model can organize crop stress into interpretable patterns, while stopping short of showing that it changes farm decisions or improves harvests. The datasets are stated to be openly available on Kaggle, and the EigenCL training and inference code is available from the corresponding author upon reasonable request. The authors declare no conflicts of interest related to the study.
Paper data and sources
Original title: Embedding NDRE Trajectories into Contrastive Learning for Label-Free, Physiology-Aware Crop-Stress Staging and DSS Outputs
Authors: Shafqaat Ahmad
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text