Research on inversion and prediction of root region soil water content in kiwifruit based on hyperparameter tuning by transformer-DsaGRU

The fruit expansion period of kiwifruit is highly sensitive to water demand, and the precise monitoring and prediction of root zone soil water content (RSWC) are crucial for ensuring fruit quality and yield. This study develops a Transformer–DsaGRU model that fuses UAV multispectral indices, meteorological variables and a rainfall-threshold-based dynamic step-size adjustment (Dsa) to forecast plant-scale RSWC. Field campaigns were conducted in a kiwifruit orchard in Meixian County, Shaanxi Province, China, over three growing seasons (2023–2025), yielding 3288 plant-day samples from 54 vines. After feature selection via Pearson correlation (p < 0.05) and mutual information, 12 core variables were retained and their contributions quantified using SHAP analysis. With a base input window of 10 days and Dsa-driven extension to 12 days under daily rainfall > 0.2 mm, the Transformer–DsaGRU achieved determination coefficients (R²) of 0.78 and 0.75 and mean absolute errors (MAE) of 1.11 and 1.06 % v/v for 1- and 2-day-ahead RSWC forecasts, respectively, outperforming RNN, LSTM, GRU and TCN baselines. Independent extrapolation using the 2025 data showed that the model maintains acceptable skill across phenological stages, indicating robust generalization to an unseen year. Nevertheless, the present work is constrained by the limited spatial extent of a single orchard, the focus on a specific growth phase, future studies will extend to more sites, phenological phases and soil–crop variables to support wider irrigation applications.