Research on Soil Spotted Degradation Prediction in Qinling Tea-producing Area of China Based on Deep Learning

Lei Zhang^1,2Tingyu Zhang²Lin Guo^3*

• ¹Technology Innovation Center for Land Engineering and Human Settlements, Xi'an Jiaotong University, Xi'an, China
• ²Shaanxi Provincial Land Engineering Construction Group, Xi'an, China
• ³Shaanxi Institute of Forestry Survey and Planning, Xi'an, China

Soil Spotted Degradation (SSD) is a critical manifestation of land degradation that significantly impedes agricultural development. The mechanisms driving this phenomenon, along with methods for accurately predicting its occurrence, remain insufficiently understood. In recent years, SSD occurrences in the tea-producing regions of the Qinling Mountains, China, have become increasingly prevalent. Given the urgent need to address soil-related challenges in this area, accurate SSD prediction has become a priority. To address the limitations of sample compilation and the underutilization of feature information in SSD prediction, this study focuses on Shangnan County, Shaanxi Province, China. Utilizing the QGIS platform in conjunction with remote sensing technology and field investigations, authentic SSD samples were collected. A novel approach that integrates Stacked Autoencoders with Dense Residual Networks (SAE-DRN) is proposed and compared against SVM, CPCNN-RF, and U-net models. Experimental results reveal that the proposed method achieves the highest accuracy (Overall Accuracy = 0.87, F1 Score = 0.89) and an area under the receiver operating characteristic curve (AUC) of 0.92. This approach demonstrates superior adaptability for small-sample predictions, providing more precise and reliable results. These findings not only guide tea cultivation site selection and soil disease prevention but also highlight the potential of leveraging advanced technologies to tackle critical challenges in agriculture, environmental sustainability, and resource management. Furthermore, the method's applicability extends beyond tea cultivation regions, presenting promising prospects for broader adoption across diverse crops and ecosystems.