A Low-Resource Machine-Learning Framework for Cold-Stress Early Warning in Aquaculture Nursery Ponds Using Manual Temperature Readings

Cold stress is a recurring risk in tropical and subtropical aquaculture nursery ponds, yet warning tools remain limited where continuous automated sensors are impractical. This study developed a low-resource cold-stress early warning framework using four years (2022-2025) of 6-hourly manual air and pond-water temperature readings from a Nile tilapia (Oreochromis niloticus) nursery pond in Cumilla, Bangladesh. Models were fitted on 2022-2023, validated on 2024 for threshold selection, and tested on 2025 as an independent year. Cold stress (daily mean water temperature 35{degrees}C) on only 4 days. Air-water coupling was strong overall (r = 0.976) but weakened in winter (r = 0.776) and further within the 18-22{degrees}C boundary zone where cold-stress classification is most sensitive. Solar radiation only marginally increased boundary-zone classification AUC from 0.782 to 0.789. In 6-hour regression, the same-hour-yesterday baseline (MAE = 1.117{degrees}C) nearly matched Extreme Gradient Boosting (XGBoost) with MAE of 1.116{degrees}C, cold-zone bias +0.36{degrees}C, and train-test gap 0.02{degrees}C; Random Forest (RF) and Long Short-Term Memory (LSTM) had MAEs of 1.195{degrees}C and 1.244{degrees}C, respectively. For cold-stress classification, Multiple Linear Regression (MLR) gave the highest F1 (0.755), while XGBoost provided the more protective operating point, detecting 95 of 108 cold-stress readings at 6-hour lead time (sensitivity = 0.880, F1 = 0.739). XGBoost warning skill extended to 12, 18, and 24-hour lead times, with F1 scores of 0.722, 0.646, and 0.704, respectively. The framework converts routine manual thermometer readings into short-lead cold-stress alerts for nursery management decisions. Multi-pond validation is needed before deployment.