A framework for fusing transfer learning and meta-learning for enhanced soil thickness prediction in unsampled areas

Abstract

Soil thickness is a crucial parameter that affects surface processes and environmental evolution. Its spatial distribution characteristics hold significant guiding importance in various fields. However, existing soil thickness prediction methods have significant limitations in cross-regional applicability, restricting their widespread use. To address this challenge, this study introduces a model framework based on the combination of feature transfer and meta-learning (MTL) to predict soil thickness in unsampled areas. We select two geologically similar regions in the southeastern hills of China as the study areas. After verifying the linear independence of the environmental covariates between the source and target domains, transfer component analysis (TCA) is applied to align their covariate distributions. A prediction model is then constructed by integrating the model-agnostic meta-learning (MAML) algorithm to predict the spatial distribution of soil thickness in the target domain. Comparative analysis reveals that the MTL model outperforms traditional multiple layer perceptron (MLP), transfer learning (TL), and meta-learning (ML) models, achieving a root mean square error (RMSE) of 14.06 cm and a coefficient of determination (R²) of 0.725. Additionally, the predicted spatial distribution of soil thickness aligns more closely with actual conditions. The transfer learning and meta-learning fusion framework proposed in this study provides a valuable reference for soil thickness prediction in unsampled areas.