ECA_FNet: a ConvNeXt-based model for soil type identification

Abstract

Soil classification and mapping are essential for agricultural resource management and sustainable production. Although deep learning—particularly convolutional neural networks (CNNs)—has advanced digital soil mapping, models frequently suffer from poor generalization across diverse environments due to background complexity and fine-grained variations in soil texture and porosity. To address these limitations, we propose ECA_FNet, a novel, highly efficient soil identification framework based on ConvNeXt. Specifically, we integrate a Faster Block into the backbone to capture discriminative soil features while drastically reducing computational overhead. Simultaneously, we employ an Efficient Channel Attention (ECA) mechanism to adaptive enhance feature extraction. Compared to the baseline ConvNeXt, our network is substantially more lightweight yet more precise. Extensively evaluated on an upland soil dataset from Bishan, Chongqing, ECA_FNet achieves over 90% across accuracy, precision, F1-score, and recall, specifically reaching a peak accuracy of 93.31% and consistently outperforming state-of-the-art methods.