Beyond Persistence: A Chronological Benchmark of Machine Learning for Short-Term Soil Moisture Forecasting Across Station-Year Datasets

Abstract

Short-term soil moisture forecasting can support agricultural monitoring and irrigation decisionsupport systems, but the incremental value of machine learning (ML) should be evaluated against strong temporal baselines. This study asks whether ML provides predictive value beyond persistence for shortterm soil-moisture forecasting across station-year datasets and multiple horizons. A publicly available AgriDataValue environmental dataset containing soil moisture, weather, and soil-temperature observations was evaluated across two monitoring stations and four station-year datasets. The raw dataset contained 22,111 observations; quality control retained 22,050 observations (99.724%). Ridge regression, Random Forest, and histogram-based gradient boosting (HGB) were compared with a persistence forecast at 1-, 6-, 12-, and 24-hour horizons. Samples were constructed only within contiguous temporal sequences, and targets were retained only when the future timestamp occurred at the exact requested elapsed time. Chronological expanding-window validation used an initial 50% training window followed by two sequential test folds. MAE, RMSE, and R² were reported, while temporal block bootstrap resampling with 2,000 replicates quantified uncertainty in ML-minus-persistence absolute-error differences. Persistence achieved the lowest MAE in all eight station-horizon combinations. Ridge was the strongest evaluated ML model but remained worse than persistence at every horizon. The mean relative MAE increase of the best ML model across the eight station-horizon combinations was 117.68%, with a median increase of 103.59%. All primary Ridge-minus-persistence 95% bootstrap confidence intervals were strictly above zero. Robustness analyses included seasonal persistence, a controlled LSTM baseline, grouped feature ablation, training-window sensitivity, bootstrap block-length sensitivity, and sequence-level ADF/KPSS stationarity analysis. These analyses did not alter the primary conclusion. Under the evaluated dataset, feature configuration, model families, horizons, and chronological validation protocol, the tested ML approaches did not demonstrate incremental predictive value beyond persistence. The result is specific to the investigated setting and is not evidence that ML is universally ineffective for soil-moisture forecasting.