Decision-oriented root-zone soil water tension forecasting and calibrated threshold-exceedance alerting for smart irrigation in peanut production

Advance information on root-zone soil water tension (SWT) may help irrigation managers identify plots approaching a management threshold. This study evaluated 168-h SWT forecasts and threshold-exceedance alerts using four peanut growing seasons (2022–2025). A smoothed operational root-zone SWT (SOFT) target was constructed from measurements at 10, 30, and 50 cm using a trailing 3-h median. Daily environmental predictors were taken from the preceding completed day. Models were trained on 2022–2023, validated on an earlier 2024 block, calibrated on a later 2024 block after removal of target-overlapping windows, and tested on 32,897 hourly forecast-origin windows from 2025. Four neural forecasters were evaluated across five random seeds. Persistence achieved mean absolute error of 10.29 kPa and root mean square error of 17.01 kPa, below the corresponding mean errors of all neural forecasters. PatchTST produced the highest forecast-derived average precision (AP) among the neural forecasters, at 0.768 ± 0.011 for 45 kPa and 0.714 ± 0.017 for 70 kPa, compared with deterministic persistence values of 0.746 and 0.693. The TCN–TFT hybrid achieved AP of 0.622 ± 0.134 and 0.569 ± 0.143. Across three direct-TCN training series, each evaluated with seeds 42–46, mean AP ranged from 0.847 to 0.862 at 45 kPa and from 0.546 to 0.588 at 70 kPa. These numerical comparisons were interpreted descriptively. The results support separate evaluation of trajectory error, threshold-exceedance discrimination, transition detection, and probability calibration. Irrigation-policy performance, physiological crop stress, yield improvement, and water savings were not evaluated.