Cross-station transferability of machine learning and deep learning models for multi-horizon temperature forecasting in Bangladesh agricultural zones

ABSTRACT

Four linked panels with a take-home ribbon. (1) Station diagram: Mymensingh (north-central training station, daily temperature SD 4.27 degree symbol C) linked to Rajshahi (dry western, 5.03 degree symbol C) and Chittagong (coastal south-east, 3.39 degree symbol C); solid arrows show direct transfer (trained at Mymensingh, tested at Rajshahi and Chittagong) and dashed arrows show reverse transfer (trained at Rajshahi or Chittagong, tested at Mymensingh); a small bar chart ranks native variability (Rajshahi greater than Mymensingh greater than Chittagong). (2) Six ML and five DL models over four horizons, with the best model at Mymensingh per horizon: CNN_LSTM at 1 day (RMSE 1.15 degree symbol C, R2 0.93) and 7 days (1.60 degree symbol C, 0.87), CNN_GRU at 15 and 30 days (1.63 degree symbol C, 0.86). (3) Grouped bar chart of mean R2 change across eleven models for the four transfer directions; Mymensingh-to-Chittagong shows the largest loss (up to 28% at 30 days) and Rajshahi-to-Mymensingh the smallest (under 3%). (4) Key findings: RMSE rose 6.1 to 68.2% under transfer to Rajshahi; at Chittagong RMSE fell 3.1% at 1 day through a scale effect while R2 dropped 21.5-28.1% at 7-30 days; Rajshahi-to-Mymensingh RMSE changed only -1.4 to +6.8%, with RFR leading at 7-30 days; Chittagong-to-Mymensingh RMSE rose 11.4-19.6%; XGB collapsed at 30 days at every station (RMSE 2.42 degree symbol C at Mymensingh). Ribbon: transfer difficulty follows native temperature variability, not training-station identity, and transferability should be judged by R2 rather than RMSE alone.

Accurate air temperature forecasting supports agricultural planning in climate-vulnerable regions, yet spatial transferability of machine learning and deep learning models remains underexplored. This study compares eleven architectures (MLR, RFR, GBR, XGB, LGBMR, SVR, LSTM, CNN, GRU, CNN_LSTM, CNN_GRU) across four horizons (1, 7, 15, 30 days) using 25-year records (2000–2024) from three climatically distinct Bangladesh stations. Models trained at Mymensingh were validated without retraining at Rajshahi and Chittagong, then tested in reverse, training independently at each station and validating at Mymensingh. CNN_LSTM led at the 1- and 7-day horizons (RMSE = 1.15 ± 0.05 °C and 1.60 ± 0.04 °C) and CNN_GRU at 15 and 30 days (RMSE = 1.63 ± 0.05 °C and 1.63 ± 0.06 °C), all with R2 above 0.85. Transfer produced 6–68% RMSE degradation at Rajshahi, and a smaller RMSE change but larger R2 loss at Chittagong, reaching 50.7% for MLR at 30 days. Reverse transfer confirmed that difficulty tracks native variability rather than training-station identity, with milder changes and RFR outperforming the deep hybrid models at Rajshahi. XGB, though efficient, collapsed at the 30-day horizon at all stations. A seven-metric, PCA-based composite ranking confirmed these patterns, guiding horizon- and location-specific model selection, supporting climate-resilient agricultural planning in data-sparse regions.