| 06 Aug 2026
From Sensors to Irrigation Decisions: An Automated Data–Model Pipeline for Real-Time Soil Moisture Forecasting in Agriculture
Abstract. Increasing water scarcity and climate variability demand reliable, site-specific soil moisture information to guide irrigation and crop management decisions at the plot scale. Land surface models simulate the interactions between land, vegetation, and the atmosphere, and provide detailed information on hydrological states and fluxes. This study presents an automated, end-to-end modelling pipeline that connects real-time sensor observation, land surface modelling, data assimilation, and ensemble forecasts to deliver operational, site-specific soil moisture forecasts for agricultural stakeholders. The framework integrates in situ measurements from a Cosmic-Ray Neutron Sensor (CRNS) with the Community Land Model version 5 (CLM5) and the Parallel Data Assimilation Framework (PDAF). Observations are assimilated using an Ensemble Kalman Filter (EnKF) to constrain model states and reduce uncertainties. For forecasting, the system includes medium-range ensemble weather predictions from the European Centre for Medium-Range Weather Forecasts (ECMWF), which provide probabilistic information on future atmospheric conditions by running multiple simulations with perturbed initial states. Applied at the Selhausen agricultural research site (Germany), the framework shows how data assimilation improves the agreement between simulated and observed soil moisture by approximately 20 % (e.g., reduced RMSE), and yields short-term forecasts with skill comparable to or exceeding the open-loop at short-to-intermediate lead times, with added value becoming particularly apparent when unresolved wetting or drying events occur later in the forecast horizon. By translating model outputs into threshold-based indicators, the system enables actionable insights for irrigation and field management, thereby bridging the gap between scientific modelling and practical agricultural decision-making. Overall, the proposed framework represents an important step towards an operational, site-specific decision-support tool for short-term irrigation and agricultural water management under uncertain weather conditions.
This manuscript presents a well-integrated pipeline connecting CRNS observations, CLM5–PDAF data assimilation, ECMWF ensemble forecasts, and an accessible display of soil-moisture conditions. The application at Selhausen is clearly motivated, and the comparison with independent SoilNet measurements is a strength. The study provides a useful demonstration of the technical feasibility of this workflow. My recommendation is minor revision, principally to align the interpretation with the evidence presented. The results establish improved agreement of the assimilated model state with CRNS observations and illustrate one 10-day forecast. They do not yet establish general forecast skill or the effectiveness of irrigation decisions. Please find below some comments:
1. Figure 9 shows a forecast initialized on 1 July 2022, but the abstract and conclusions make general claims about lead-time skill (abstract, lines 11–17; lines 410–426, 518–532). The single example cannot establish when DA improves forecasts or how long that benefit persists. In case you wish to make a general claim, you shall run forecasts from multiple dates spanning wet and dry periods. Report error against subsequent observations by lead day for both DA- and open-loop-initialized forecasts, alongside ensemble coverage or a proper probabilistic score.
2. Figure 7 shows modest RMSE improvements at 5–50 cm and a deterioration at 100 cm. The substantial shallow-depth discrepancy after SoilNet sensors were removed and reinstalled also limits interpretation. A brief, balanced summary of these findings should accompany the stronger CRNS-based result. Please clarify that point measurements and the spatially integrated CRNS observation have different spatial support.
3. The site was irrigated in 2022, but its timing and amount are unknown, while the model is configured as rainfed potato (lines 206–213). This is an unmeasured water input and a serious limitation for both water-budget interpretation and forecast verification. The possible irrigation-related wetting around 5–6 July cannot have been predicted by a forecast initialized on 1 July unless the future input was specified. Later daily assimilation may correct the subsequent analysis, but that does not demonstrate forecast recovery (lines 359–365). Separate those two statements and, if records cannot be recovered, evaluate periods with and without suspected management inputs.
4. A single 100 m × 100 m column represents the model, whereas the stated CRNS footprint extends roughly 150–240 m in radius and has moisture-dependent depth sensitivity (lines 92–98, 203–205, 279–288). Explain what land covers and soil conditions lie within that footprint, how horizontal weighting is handled, and why one column is a defensible observation counterpart. Likewise, SoilNet nodes within about 30 m of the instrument provide independent measurements, but they do not sample the full CRNS footprint (lines 227–235). These scale differences affect the observation error and interpretation of Figure 7.
5. The observation weights depend on soil moisture and sensing depth, yet Equations (2)–(3) present a fixed linear H_t? (lines 279–288). State whether the weights are recomputed for each member and cycle, frozen during an analysis, or otherwise approximated. Describe exactly which CLM layers are updated, how increments are distributed, and how bounds on water content are enforced. Report cumulative DA water increments in mm alongside precipitation, evapotranspiration, and drainage. Without this accounting, an improved moisture fit could conceal substantial artificial water additions or removals.
6. The CRNS estimates are calculated using a 24-hour rolling mean and assimilated daily at 12:00 UTC with an assumed uncertainty of 0.02 cm³ cm⁻³ (lines 220–224, 279–288). Please specify whether each averaging window ends at the assimilation time; a centered window would include observations unavailable at that time. Clarify whether the model observation operator uses a corresponding 24-hour average or an instantaneous state. Finally, explain what the prescribed uncertainty includes, particularly whether it accounts for calibration and differences between the CRNS footprint and the modeled column. A brief discussion of how this assumption may affect the results would suffice.
7. Figure 10 classifies 10 cm soil moisture below 22 vol% as “irrigation needed” (lines 380–407). This surface-layer threshold is useful for communicating drying risk, but irrigation timing for potato depends on water availability throughout the root zone, crop stage, and expected rainfall. FAO-56 likewise frames irrigation scheduling in terms of root-zone depletion and readily available water. Please explain why the 10 cm threshold is appropriate for this display and describe the categories as illustrative indicators unless their link to irrigation decisions has been evaluated. Also clarify whether the reported 84% favourable and 16% drying shares apply to one forecast day or are aggregated across the forecast period.
Specific comments: