Multi-source data fusion for estimating potato transpiration under water stress using machine learning models

Accurate estimation of crop transpiration (T) is critical for optimizing irrigation and enhancing water use efficiency. This study developed a multi-source data fusion framework to estimate daily cumulative transpiration in potato under varying water stress by integrating canopy indices from images with meteorological measurements. The Crop Water Stress Index (CWSI) and Relative Leaf Area Index (RLAI) were extracted using semantic segmentation and image registration. These indices, combined with air temperature, humidity, CO₂ concentration, light intensity measured via a LoRa wireless sensor network, as well as normalized time indicators (hour and minute, scaled to a 0–1 range) corresponding to each observation, served as inputs for Random Forest Regression (RFR), Back-Propagation Neural Network (BPNN), and Long Short-Term Memory (LSTM) models. Six datasets collected over two years (2022 and 2024) with three irrigation treatments were analyzed. Compared to using meteorological variables alone, incorporating CWSI and RLAI significantly enhanced model performance, increasing R² by 1.77–18.44 %, 3.44–11.87 %, and 0.44–18.42 % for RFR, BPNN, and LSTM respectively. In stable environmental conditions of 2022, RFR achieved the best accuracy (R² = 0.8851–0.9654, RMSE = 2.60–9.63 g, MAE = 1.83–6.06 g, RPD = 2.96–5.49). Under more variable conditions in 2024, LSTM outperformed other models (R² = 0.9187–0.9898, RMSE = 14.36–21.02 g, MAE = 10.92–14.62 g, RPD = 3.64–10.54). These findings suggest that RFR is preferable for stable environments, while LSTM is better suited to fluctuating conditions. Integrating CWSI and RLAI with meteorological data improves daily cumulative transpiration estimation, offering a robust foundation for precision irrigation management in potato production.