Abstract

Soil property estimation is critical for precision agriculture, environmental modeling, and land management, but developing accurate models for diverse regions remains challenging due to data variability and scarcity. Transfer learning (TL), a machine learning approach that reuses pre-trained models, offers a solution by leveraging knowledge from data-rich regions to improve predictions in data-scarce ones. This study explores TL for estimating soil properties (e.g., organic carbon, clay content, pH) across three distinct agricultural regions using a convolutional neural network (CNN). We demonstrate that TL significantly improves prediction accuracy compared to region-specific models, particularly in data-limited areas. Results highlight the potential of TL to enhance soil mapping scalability and robustness, with implications for global soil management.