Deep Carbon: A Multiscale Feature-Time Fusion Approach for Field Level Digital Soil Organic Carbon Mapping

Abstract

Soil organic carbon (SOC) plays a key role in soil health and ecosystem services. This study introduces Deep Carbon, a modelling framework that integrates static and time-series environmental covariates for high-resolution SOC prediction at the field scale. Time-series data were encoded using a stacked long short-term memory (LSTM) neural network to extract temporal patterns of dynamic features. These encoded time-series representations were combined with static covariates and used as inputs to train machine learning models at multiple spatial resolutions (5 km to 10 m). Individual predictions at each scale were then fused using a partial least squares regression (PLSR) model to generate SOC maps at 10 m resolution. The best accuracy was observed at 5 km scale (R² = 0.75; RMSE = 0.30% in log scale), while the fused 10 m prediction yielded a testing R² of 0.58 and RMSE of 0.44%. Fusion modelling identified 30 and 250 m resolutions as the most influential predictors. The approach successfully captured both high- and low-frequency SOC variations and demonstrated good transferability when tested on new observations from 2022. This multi-scale feature-time fusion approach uses legacy ground samples and satellite data to enable scalable and accurate digital SOC mapping.