A Segment Anything Model-Based Geological Remote Sensing Interpretation Method with a Distributed Data-Parallel Deep Learning Framework

The interpretation of remote sensing images is pivotal in extracting geological elements of interest. Recent studies using deep learning models often fail to provide accurate boundaries between geological elements due to high interclass similarity and imbalanced data distribution. Furthermore, these models are computationally intensive. Inspired by expert interpretation practices, which involve first delineating boundaries and then identifying semantics, we leverage the vision foundation model and propose a distributed interpretation framework including distributed training and inference phases based on data parallelism in distributed architectures. First, a conventional semantic segmentation model, DeepLabV3, is trained. Subsequently, we integrate the Segment Anything Model (SAM) with the trained model, completing the final mapping during the inference phase. Specifically, the inference model features two branches: one for extracting masks by SAM and another for categorizing geological elements. Then, a semantic voting module combines the information from both branches to determine the category for each mask. Experiments demonstrate that the method enhanced the performance metric by more than 2%. Moreover, the method achieves a 3.85x speedup in training and a 3.95x speedup in inference on a 4-GPU machine. We further exploit simulation experiments of 16 GPUs and achieve up to 15.1x training speedup.