Deep learning-based national scale soil organic carbon mapping with Sentinel-3 data

Abstract

Mapping of soil organic carbon (SOC) at the regional level is critical for climate change policy and the mitigation of its adverse effects. However, reliable SOC estimates particularly over a large extent remains a major challenge due to among others limited sample points, quality of simulation data and the algorithm adopted. Remote sensing (RS) strategies have emerged as a suitable alternative to field and laboratory SOC determination, especially at large spatial extent. The use of Sentinel-3 sensor, the latest of the Sentinel series is minimal and has not been fully developed, despite its impressive attributes that include high spectral-temporal resolution and large coverage. Compared to linear and classical ML models, deep learning (DL) models offer a considerable improvement in data analysis due to their ability to extract more representative features and identify complex spatial patterns associated with big data. Yet, there is paucity in literature on the application of dl-based remote sensing strategies for SOC prediction. Consequently, this study adopted a deep neural network (DNN) to predict SOC at a national scale, using Sentinel-3 image, and compared the results with random forest (RF), support vector machine (SVM) and artificial neural network (ANN) models. The models were trained based on 10-fold cross-validation with 1936 soil samples and 31 predictors. The DNN model generated the best result with a root mean square error (RMSE) of 10.35 t/ha (26 % of the mean), followed by RF (RMSE = 11.2 t/ha), ANN (RMSE = 11.6 t/ha) and SVM (RMSE = 13.6 t/ha). The analytical prowess of the DNN, together with its ability to handle big data by learning patterns through a series of hidden layers (10) to draw conclusions, gives it an edge over other classical ML models. The study concluded that the DNN model with Sentinel-3 data is promising and provides an effective framework for continuous national level SOC modelling.