Research on land cover type classification method based on improved MaskFormer for remote sensing images

Area of Science:

• Remote Sensing
• Ecology
• Computer Vision

Background:

• High-resolution remote sensing images offer extensive coverage and spectral data for land cover analysis.
• Existing datasets often lack representation of highland mountainous areas, differing significantly from plain regions.
• There is a need for specialized datasets to support ecological protection in unique terrains.

Purpose of the Study:

• To construct a grassland ecological element dataset for Qilian County to aid highland ecological protection.
• To develop and validate a novel segmentation network for accurate land cover classification in challenging environments.

Main Methods:

• A Qilian County grassland ecological element dataset was created, focusing on RGB and near-infrared spectra.
• A Shunted-MaskFormer network was proposed, incorporating mask-based classification, multi-scale feature extraction, and data-dependent upsampling.
• The network's performance was evaluated on the Qilian grassland dataset and the Gaofen Image Dataset (GID).

Main Results:

• The optimized Shunted-MaskFormer network achieved a Mean Intersection over Union (MIoU) of 80.75% on the Qilian grassland dataset.
• The model demonstrated superior segmentation for small sample classes in imbalanced datasets.
• A highest MIoU of 72.3% was obtained on the GID dataset.
• The optimized model's size was reduced to one-third of the suboptimal model.

Conclusions:

• The developed Shunted-MaskFormer network provides effective grassland segmentation for highland areas.
• The Qilian County grassland ecological element dataset supports crucial ecological protection efforts.
• The model shows strong generalization capabilities and improved efficiency.