Towards a two-stage low-cost soil carbon estimation over large area using proximal sensing and digital soil mapping

Accurate and scalable estimation of soil organic carbon (SOC) is essential for sustainable land management and climate mitigation, yet conventional laboratory-based methods remain costly and impractical for large-area monitoring. This study presents a novel two-stage framework that integrates low-cost proximal red, green, and blue (RGB) soil imaging, deep learning, and machine learning-based digital soil mapping (DSM) to enable regional SOC prediction at 100 m resolution. In the first stage, SOC was estimated from RGB images of 405 surface soil samples collected across multiple agro-climatic zones of West Bengal, India, using a random forest model based on color features and a fine-tuned Visual Geometry Group 16-layer (VGG16) convolutional neural network. The deep learning model achieved superior performance, with a 37.5% higher validation R² and 13.1% lower mean absolute error (MAE) compared to the color feature-based model. In the second stage, RGB-derived SOC predictions were integrated with multi-source environmental covariates within a DSM framework, yielding regional SOC maps that slightly outperformed a benchmark wet chemistry-based DSM with 8.8% higher test R² and 32.1% lower MAE. The maps revealed strong agro-climatic control on SOC distribution, with higher SOC in the northern Terai Zone (1.1–2.0%) and lower SOC in the Red and Laterite Zone (