Soil Texture Classification Based on Deep Learning Networks and Multi-Source Remote Sensing Images

To address the inefficiency and limited applicability of traditional soil texture analysis methods, this study proposes a soil texture classification approach integrating multi-source remote sensing images with deep learning networks. Using Landsat- 8 and Sentinel-2 remote sensing data, we constructed 256 $\times 256$ sample sets and conducted comparative experiments on five classical neural networks: SegFormer, U-Net, PSPNet, DeepLabV3+, and SegNet. The results demonstrate that SegFormer achieves optimal overall performance on Sentinel-2 data ($\mathrm{mIoU}=0.6942$), with its Transformer architecture effectively capturing global contextual features of soil textures. This validates the significant impact of multi-scale feature processing capabilities on model performance. For Sentinel-2 data, SegNet exhibits the highest classification accuracy for gravelly soil and hard soil (IoU are 0.7796 and 0.7899, respectively), but underperforms for ordinary soil (IoU is less than $\mathbf{0. 4 3}$). The study confirms that fusing deep learning models with multi-source remote sensing data can overcome the geographical limitations of traditional methods, with SegFormer showing cross-data-source stability (mIoU >0.69 for both datasets). SegFormer emerges as a novel solution for large-scale soil texture monitoring, though boundary ambiguity in hard and ordinary soil classification requires further optimization. This research provides an efficient technical pathway for digital soil mapping and precision agricultural management.