Published Paper Details:

Crop stress is one of the foremost threats to global food security, causing yield losses estimated at 20-40% annually. Traditional monitoring relies on costly sensor arrays or time-intensive manual inspection, limiting accessibility for small-scale farmers. This paper proposes a low-cost, camera-only crop stress detection system built around a fixed ESP32-CAM RGB module. A multi-task convolutional neural network (CNN) based on MobileNetV2 simultaneously classifies crop health (Healthy/Stressed), estimates a continuous stress severity score, and assesses soil surface condition (Dry/Moist/Wet). Temporal analysis via an exponentially weighted moving average (EWMA) over a rolling N-frame buffer enables detection of progressive stress onset before visible leaf damage appears, reducing false-positive alerts by 31% compared to single-frame classification. A hybrid learning pipeline combines supervised training on a labeled base dataset with semi-supervised pseudo-label fine-tuning for unseen crop types. Grad-CAM visualisations highlight stress-causative regions, improving model transparency and farmer trust. The system achieved 93.4% binary classification accuracy, 89.7% soil condition classification accuracy, and a stress severity MAE of 0.06 on held-out test data.