Real-Time Crop Stress Monitoring and Early Warning System for Paddy and Maize Using Multi-Temporal Sentinel-2 Data and Deep Learning in Semi-Arid Regions

Abstract

Semi-arid regions with a high potential for rice and maize cultivation have become some of the most actively farmed areas. They now face the challenge of achieving food security despite the threats of crop water stress, nutrient loss, and environmental changes. In this paper, we develop a real-time crop stress monitoring and early warning system that utilizes multi-temporal Sentinel-2 images and deep learning models in Mahabubabad district, Telangana, India. Different types of crop stresses such as water stress, nutrient deficiency, and phenological anomalies were detected and classified using a hybrid CNN-LSTM architecture with an attention mechanism. The methodology was based on 874 field polygons with extensive in-situ data collection during 2023-24, incorporating multi-temporal spectral indices (NDVI, EVI, NDWI, REP), weather variables, and soil characteristics. The total classification accuracy reached 89.4% for paddy and 87.2% for maize over all stress types, showing that stress detection from satellite images is quite reliable. Water stress was the category that was detected most accurately (92.1% for paddy and 89.8% for maize), followed by nutrient stress (88.7% and 86.3%) and phenological stress (85.2% and 83.9%). The warning system made it possible to identify the problem 15-25 days before there were visible symptoms, making it possible for the farm management to respond in time. Activities of the farm that were most vulnerable to detection were air and water temperatures, precipitation, and crop growth stages for water stress 45-60 days after sowing, 30-45 days for nutrient stress, and during the reproductive phase for phenological stress. The system could be extended for industrial crop stress monitoring across the semi-arid agricultural systems which might lead to precision agriculture and climate-resilient farming practices.