Temporal Machine Learning and Spatial Mapping of Surface-Soil Moisture: Integrating OPTRAM Geometry with Ground-Based and Sentinel-2 Data

Field-scale surface-soil-moisture information is needed for crop water management, but optical trapezoid methods such as OPTRAM provide only relative, date-specific indices. We tested a two-component framework on one winter wheat field near Röszke, Hungary (6.79 ha), using 22 Sentinel-2 L2A scenes, three continuously logged TEROS10 sensors and six TDR350 campaigns. OPTRAM edges were fitted per date over a 564.5 ha calibration area, and the relative index W was converted to Volumetric Water Content (VWC) maps by a per-date linear trend at the sensors plus inverse-distance-weighted residuals. A gradient-boosting model estimated VWC at the sensors from spectral, OPTRAM, soil and meteorological predictors. Under leave-one-date-out validation it reached R2 = 0.90 (MAE = 1.7 vol.%) against 0.75 for 7-day persistence; with forward chaining using only earlier dates, R2 was 0.66 (0.14 without sensor-derived antecedent moisture). Ablations showed that this accuracy derives from soil, meteorological and antecedent predictors; Sentinel-2 and Sentinel-1 predictors added no measurable gain. The maps agreed with independent TDR350 data (pooled r ≈ 0.70), but plain interpolation of the three sensors agreed equally well (r ≈ 0.80); unanchored W correlated with TDR within dates (mean r = 0.57). Results apply to one field and one season.