Efficient Patch-Wise Semantic Segmentation for Large-Scale Remote Sensing Images.

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Semantic segmentation is crucial for remote sensing image analysis.
• Traditional methods struggle with high-resolution and large-scale remote sensing data.
• Challenges include color imbalance, object rotation, and scale variations.

Purpose of the Study:

• To develop an efficient and accurate semantic segmentation method for common land resources in remote sensing images.
• To address the limitations of existing methods in handling high-resolution and large-scale data.
• To improve the precision of land resource classification.

Main Methods:

• A novel patch-wise semantic segmentation method using fully convolutional networks.
• Image splitting into local patches for high-resolution data processing.
• Data preprocessing to handle color imbalance, rotation variations, and lens distortion.
• A multi-scale training strategy to address scale variation.
• Investigation of Conditional Random Field (CRF) for precision enhancement.

Main Results:

• The proposed method achieved a mean intersection over union (MIoU) of 54.96% on a Gaofen-2 satellite dataset.
• Demonstrated superior performance compared to other state-of-the-art methods in remote sensing image segmentation.
• Successfully segmented ten common land resources, including Grassland and Road.

Conclusions:

• The developed patch-wise semantic segmentation method with a new training strategy is effective for remote sensing image analysis.
• The approach successfully handles high-resolution, large-scale images and various data characteristics.
• This method offers a significant advancement in classifying common land resources from satellite imagery.