Deep Transfer Learning for Land Use and Land Cover Classification: A Comparative Study

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• High-resolution remote sensing imagery is crucial for accurate land use and land cover (LULC) classification.
• Deep learning, particularly convolutional neural networks (CNNs), has significantly advanced image classification.
• Transfer learning offers a powerful approach to fine-tune pre-trained models for specific LULC tasks, especially with limited data.

Purpose of the Study:

• To implement and evaluate transfer learning for LULC classification using pre-trained Visual Geometry Group (VGG16) and Wide Residual Networks (WRNs).
• To compare the performance and computational efficiency of VGG16 and WRNs on the EuroSAT dataset.
• To optimize the classification process using techniques like early stopping, gradient clipping, adaptive learning rates, and data augmentation.

Main Methods:

• Fine-tuning pre-trained VGG16 and WRN models by replacing their final layers for LULC classification.
• Utilizing the red-green-blue version of the EuroSAT dataset for training and validation.
• Applying optimization techniques including early stopping, gradient clipping, adaptive learning rates, and data augmentation to improve performance and efficiency.

Main Results:

• The transfer learning approach successfully addressed the limited-data problem in LULC classification.
• Wide Residual Networks (WRNs) achieved a superior accuracy of 99.17%, outperforming previous benchmarks.
• The WRN-based method demonstrated enhanced computational efficiency compared to other approaches.

Conclusions:

• Transfer learning with fine-tuned WRNs is a highly effective method for high-resolution LULC classification.
• The proposed approach offers a computationally efficient and accurate solution for remote sensing image analysis.
• This study sets a new standard for accuracy and efficiency in LULC classification using deep learning.