A Modular Orchestrated Pipeline for Soil Moisture Forecasting and Irrigation Decision Support: An Operational Evaluation

Abstract

Accurate soil moisture information is central to efficient irrigation scheduling, yet most operational monitoring systems report only current conditions in real time and stop short of forecasting future soil moisture or translating it into actionable irrigation decisions within a single evaluated, end-to-end system. To address this gap, this work proposes a modular orchestrated approach that forecasts soil moisture content and derives irrigation advisories from those forecasts. It was tested using two US Department of Agriculture datasets: Cook Farm in Washington State for dryland wheat (totaling 3,346 instances) and SCAN Bushland in Texas for irrigated wheat (a total of 2,192 instances). Data ingestion, transformation, forecasting, and generation of advisory recommendations are performed by four modules (Data Ingestion, Forecasting, Advisory, and Orchestrator) to produce a machine-readable JSON output with retry logic and execution logging capabilities using gridMET precipitation and reference evapotranspiration data as input. With regard to Cook Farm, adding gridMET weather features improves the Gradient Boosting holdout RMSE from 0.00260 (soil and temperature only) to 0.00222, an improvement of about 14.6%. Across the two sites the relative value of the learned model depends on the soil-moisture dynamics: Gradient Boosting improves modestly over a strong persistence baseline at Cook Farm (holdout RMSE 0.00222 versus 0.00244), whereas at the noisier irrigation-driven Bushland series persistence (RMSE 0.01083) outperforms Gradient Boosting (RMSE 0.01978). This indicates that the added meteorological features improve on a strong persistence baseline at one site but not the other; under a Diebold-Mariano test, however, neither the Gradient Boosting advantage at Cook Farm nor the persistence advantage at Bushland is statistically significant (p = 0.29 and p = 0.08), so the learned model is best described as competitive with persistence rather than decisively better. The central contribution is accordingly the fully implemented, evaluated, and portable decision-support pipeline, in which the forecasting model is one replaceable and transparently benchmarked component. A 398-step chronological replay on Cook Farm yielded an advisory accuracy of 0.9975 with a median latency of 14.01 ms per iteration. Applying the same pipeline to a second, contrasting site without code changes demonstrates portability across climates, soils, and irrigation regimes, while also showing that the learned model’s advantage over persistence is site-dependent.