An improving spectral PTF for mining area soil water content prediction: combining 2D correlation spectroscopy and...

Conventional methods for soil sampling and soil water content (SWC) measurement are often labor-intensive and time-consuming. The Pedo-transfer function (PTF) integrating soil spectroscopy with soil physicochemical properties provides a more efficient approach for SWC estimation. However, existing studies highlight regional limitations in the accuracy of PTFs across diverse geographical regions. To improve the accuracy of SWC estimation, a study was conducted in mining-induced subsidence areas with diverse mining types in the central-northern Huaibei mining region (Anhui Province, China). A total of 383 soil samples (0-20 cm depth) were collected across these areas. Based on these collected samples, a multimodal dataset was constructed by combining laboratory-measured soil physicochemical properties, crop growth indicators, raw soil spectral data, and two-dimensional correlation spectroscopy (2Dcos) images. A hybrid ResGRU model-integrating residual neural networks (ResNet) for spatial feature extraction and gated recurrent units (GRU) for one-dimensional (1D) sequence modeling-was developed for multimodal SWC prediction. The results demonstrated that the ResGRU model achieved superior performance when fusing 2Dcos imagery, 1D spectral data, and soil-crop indicators, achieving a coefficient of determination (R²) of 0.94 and a root mean square error (RMSE) of 0.01. Compared to traditional PTFs and machine learning models, the proposed method improved prediction accuracy by 9-25 %, underscoring its effectiveness and superiority for SWC estimation in mining-impacted soils.