Classification of Soil Images Based on Frequency-Space Feature Fusion and Dynamic Convolution

This paper proposes an improved feature extraction method to enhance the accuracy of soil image classification, with the aim of more effectively capturing the diverse structures, textures, and components within soil images. Traditional convolutions often fail to fully exploit deep features when processing complex images. To address this limitation, we replace the original standard convolution with ODConv, which optimizes the convolutional structure and significantly enhances feature extraction capabilities. Additionally, a multi-scale fusion strategy, combining frequency and spatial features, is introduced, enabling the model to more precisely capture feature information at various scales, thereby improving classification robustness. Experimental results demonstrate that the improved model achieves a classification accuracy of 91.07% on the soil image dataset, a 1.09 percentage point increase over the unmodified MobileNetV3. Further experiments indicate that the enhanced model exhibits greater stability and stronger generalization when applied to different types of soil images, thus validating the effectiveness of the proposed method.