Advanced Prediction of Soil Organic Carbon: A Hybrid Transformer Network with Cost-Sensitive Learning Using Remote Sensing and Climate Data

The analysis of soil organic carbon (SOC) contents and stocks provides crucial insights into soil health, biodiversity support, erosion control, and climate change mitigation. To enhance SOC level estimation and reveal patterns across diverse ecosystems, advanced deep learning techniques are integrated with remote sensing data. Despite this progress, designing a powerful network capable of handling unevenly distributed ground truths remains a challenge, reflecting the inherent complexity of the data in such scenarios. In this study, we propose a hybrid deep learning network that incorporates a vision transformer (ViT) for processing remote sensing images and a transformer for handling climate features. To address the challenges posed by particularly difficult-to-predict samples, we apply a cost-sensitive loss function within this network. Experimental results on the LUCAS dataset demonstrate that this new loss function significantly outperforms conventional ones.¹