Bare soil erosion dust plays a significant role in the formation of haze weather, occupying an important position in the source of air particulates. Accurate monitoring of bare soil is an urgent need for environmental governance. High-resolution remote sensing images, rich in semantic information, are beneficial for the extraction of bare soil, but they are easily influenced by similar bare soil and background conditions. To address this, this paper proposes a Multi-branch Semantic Information Fusion Network (MSIF-Net) framework for the automatic detection of bare soil in high-resolution remote sensing images. MSIF-Net consists of encoding and decoding parts. The encoding part uses ResNet101 for preliminary feature extraction and outputs four levels of features. The highest-level feature is input into the constructed Dense attention-based Atrous Spatial Pyramid (DASP) module for in-depth extraction of multi-scale features. The medium and low-level features are input into the Multi-branch Detail Information Extraction (MDIE) module, fully utilizing the semantic and detail information of different branches to enhance feature extraction and refinement of local features. In the decoding part, the processed features of different resolutions are input into the constructed Multi-element Flow Alignment (MFA) module, which uses multi-branch stepwise decoding. Shallow spatial information guides the deep semantic information to achieve more flexible feature alignment and improve edge segmentation effects. Experiments on 2 m resolution remote sensing images obtained by the China-Brazil Earth Resources Satellite demonstrate that MSIF-Net has significant advantages over networks like U-Net, DeepLabV3+, MF2AM, and BuildFormer. Its Pixel Accuracy and Intersection over Union reached 93.84% and 85.32%, respectively, enabling high-precision automatic extraction of bare soil.