Evaluation of deep learning and transform domain feature extraction techniques for land cover classification:...

Area of Science:

• Remote Sensing
• Machine Learning
• Image Processing

Background:

• Accurate land cover classification is crucial for environmental monitoring.
• Traditional methods struggle with feature extraction and imbalanced datasets in SAR data.
• Deep learning and transform domain techniques offer potential improvements.

Purpose of the Study:

• To compare deep learning and transform domain feature extraction for SAR land cover classification.
• To evaluate the effectiveness of fusing Convolutional Autoencoders (CAE) and Haar Wavelet Transforms (HWT) features.
• To assess the impact of Synthetic Minority Oversampling Technique (SMOTE) on imbalanced datasets.

Main Methods:

• Utilized Convolutional Autoencoders (CAE), Variational Autoencoders (VAE), and Haar Wavelet Transforms (HWT) for feature extraction.
• Implemented feature fusion of CAE and HWT for improved information capture.
• Applied Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance.
• Employed hierarchical classification to differentiate similar land cover types.

Main Results:

• Feature fusion of CAE and HWT effectively combined high- and low-frequency information, boosting classification accuracy.
• SMOTE significantly increased training samples, mitigating issues with imbalanced data.
• Hierarchical classification improved discrimination between visually similar classes like forest and agricultural land.
• The study demonstrated improved classification performance on RISAT-1 and AIRSAR datasets.

Conclusions:

• Fusing CAE and HWT features offers a robust approach for SAR land cover classification.
• SMOTE is effective in handling imbalanced datasets for improved classification outcomes.
• Hierarchical classification combined with advanced feature extraction provides a valuable guidance for accurate land cover mapping.