Field-Level Crop Stress Detection from Temporal Satellite Imagery and Local Weather Records

Abstract

Crop stress is often identified only after visible damage has developed, limiting the time available for irrigation or disease management. This study presents a field-level crop stress detection method that combines temporal multispectral satellite imagery with local weather observations. Vegetation indices, spectral changes, rainfall, temperature, and soil moisture estimates are modeled over consecutive observation periods to distinguish persistent crop stress from short-term spectral variation caused by weather or imaging conditions. The evaluation dataset contains 4,218 agricultural fields covering approximately 31,600 hectares across two growing seasons, with 1,037 field-level stress events confirmed through agronomic inspection records. The proposed method detects stress events with an F1-score of 88.6% and AUROC of 0.941, compared with 79.3% F1 for a single-date image classifier and 83.7% for a temporal satellite-only model. For 62.4% of correctly detected cases, the system identifies abnormal conditions at least seven days before the corresponding field inspection report. Removing weather information decreases F1 by 4.8 percentage points, while removing temporal spectral features decreases it by 7.1 points. These results indicate that combining temporal remote sensing and local environmental observations can improve early identification of field-level crop stress.