Land Cover Classification Model Using Multispectral Satellite Images Based on a Deep Learning Synergistic Semantic...

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Computer Vision

Background:

• Land cover classification (LCC) is crucial for mapping and monitoring Earth's surface using satellite imagery.
• Traditional LCC methods face challenges with increasing data complexity and demand for higher resolution.
• Deep learning, particularly Convolutional Neural Networks (CNNs), has advanced satellite image segmentation, but training is resource-intensive.

Purpose of the Study:

• To develop an improved semantic segmentation method for multispectral satellite images.
• To enhance the accuracy and robustness of land cover classification.
• To address the computational demands of deep learning by utilizing pre-trained networks and a novel post-processing technique.

Main Methods:

• A hybrid synergistic semantic segmentation approach integrating the Deeplab v3+ network with a clustering-based post-processing scheme.
• The post-processing involves a spectral bag-of-words model and K-medoids clustering to refine segmentation outputs.
• Utilizing pre-trained networks to mitigate the need for extensive hardware and training time.

Main Results:

• The proposed method accurately classifies diverse land cover types, including forests, urban areas, and water bodies.
• The hybrid approach improved the Matthews correlation coefficient (MCC) by approximately 5.7% compared to baseline deep learning methods.
• The method demonstrated robustness against data imbalance and seasonal variations, dynamically updating its classification codewords.

Conclusions:

• The synergistic semantic segmentation method offers a significant advancement in satellite image classification accuracy.
• The integration of deep learning with clustering-based post-processing effectively refines segmentation results.
• The proposed approach outperforms existing state-of-the-art methods, achieving at least a 6% higher MCC in benchmark tests.