Soil-SAM: Segment anything model for soil pore identification

High-precision pore segmentation results are a critical step in exploring soil internal structures. Due to the complex shapes and blurred boundaries of soil pores, existing methods struggle to automatically and accurately segment pores, leading to inaccurate characterization of soil structure. Recently, the large model Segment Anything Model (SAM) has gained wide application in the field of image segmentation due to its exceptional generalization ability. However, no research has yet investigated its performance in soil pore identification. Therefore, this study proposed a soil pore segmentation model (Soil-SAM) to improve soil pore structure identification precision and explore a new research paradigm for customized large models in soil image segmentation. First, Soil-SAM enhances the original SAM image encoder by integrating additional trainable Low-Rank Adaptation (LoRA) layers and multi-scale adapter layers to extract semantic information at different scales, thereby recognizing pores of varying sizes and complex shapes. Then, the model adopts a multi-feature fusion up-sampling module to combine the features of the image encoder with those of the mask decoder, leveraging multi-level semantic features to enhance the segmentation of complex pores. Compared to conventional image processing software and other deep learning methods, the proposed Soil-SAM method achieved superior pore segmentation performance, with the highest accuracy (99.27 %) and harmonic mean (80.25 %), surpassing the second-best method by 6.29 % and 4.84 %, respectively. This study demonstrated that the proposed Soil-SAM model can automatically and accurately identify complex soil pores, laying a foundation for further exploration of soil internal structures.