Comparative Analysis of Regression and Classification‐Based Deep Learning and Machine Learning Models for Soil Erodibility Prediction

ABSTRACT Soil erodibility (USLE‐K) is a critical indicator of soil susceptibility to erosion; however, its accurate estimation remains challenging due to complex and nonlinear interactions among soil physicochemical properties. Most existing studies focus either on traditional machine learning techniques or on individual deep learning models, often addressing classification or regression tasks separately and rarely integrating spatial validation of predictive results. The present study addresses these limitations by proposing an integrated comparative framework that simultaneously evaluates machine learning and advanced deep learning algorithms for both classification and continuous prediction of soil erodibility. Random Forest (RF) was compared with deep learning architectures including 1D CNN, CNN + LSTM, TabNet, RNN, LSTM, and GRU using a large dataset derived from field measurements and laboratory soil analyses in the Sakarya Basin. Model performance was assessed using standard classification and regression metrics, while predictive outputs were further evaluated through geostatistical interpolation to examine spatial consistency. The TabNet classifier achieved the highest classification accuracy (94%), demonstrating its effectiveness in capturing complex feature interactions in tabular soil data. For regression‐based prediction, the GRU model exhibited superior performance with an R 2 value of 94%, outperforming both other deep learning models and the classical RF approach. Statistical analyses indicated that regression‐based deep learning models produced largely similar predictions, a finding supported by strong agreement in spatial distribution maps. The added value of this study lies in its combined evaluation of predictive accuracy and spatial coherence, demonstrating that deep learning models—particularly TabNet and GRU—provide reliable and spatially consistent soil erodibility assessments. The proposed framework offers a practical decision‐support tool for soil erosion risk mapping, land‐use planning, and the development of effective soil conservation and erosion control strategies in heterogeneous environmental systems.