Synergistic Use of Low-Cost Nir Scanner and Geospatial Covariates to Enhance Soil Organic Carbon Predictions Using Dual Input Deep Learning Techniques

This study presents a methodological framework that predicts topsoil organic carbon (SOC) using spectral recordings from a low-cost near-infrared (NIR, 1350 - 2550 nm) device and geospatial covariates related to landform, climate, and vegetation. Initial attempts using the extreme gradient boosting algorithm yielded moderate accuracy metrics (RMSE 0.36 log-SOC), emphasizing the necessity for synergistic integration with geo-covariates to address existing limitations. To address this, a dual-input deep learning framework was utilized, where diverse data inputs are fed into two distinct streams. One stream utilizes a convolutional neural network with NeoSpectra sensor spectral signatures, while the other employs 214 raster spatial layers in a fully connected neural network. These streams are combined into a unified feature vector for a dense layer, generating the final prediction. Preliminary results indicate that this dual-input approach holds significant potential in enhancing predictive accuracy for handheld NIR sensors, resulting in a reduced RMSE of 30.56%.