A Soil Organic Carbon Mapping method based on transfer learning without the use of exogenous data

Jingfeng HanMujie WuYanlong Qi^*Xiaoning LiXiao ChenJing WangJinlong ZhuQingliang Li

• Department of Computer Science and Technology, Changchun Normal University, Changchun, China

Accurate, robust, and cost-effective mapping of soil properties such as soil organic carbon (SOC) is essential for supporting policy decisions. Deep learning-based methods have been highly successful in the field of soil mapping. Deep learning methods rely on large amounts of data; however, the availability of site data for soil mapping is often limited. To make more efficient use of information from soil sites, we propose a novel transfer learning approach to improve predictions of SOC at different depths, based on a Convolutional Neural Network (CNN) model that does not rely on exogenous data. When predicting data for one layer, we initialize the model with data from all layers and then fine-tune it using data from that specific layer. The results show that transfer model generally outperforms other machine learning models, including the Random Forest (RF) model,show that the coefficient of determination (R²) and root mean square error (RMSE) of the transfer model are 0.374 and 2.937%, respectively, demonstrating superior performance compared to the RF, CNN, and MTCNN models. The results indicate that, under conditions of limited data availability, the proposed method offers clear advantages for digital soil mapping and provides an efficient alternative for accurately assessing soil carbon content.