High-performance scene classification in remote sensing imagery using a custom deep CNN architecture.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Accurate multi-class image categorization is crucial for remote sensing data analysis.
• Existing Convolutional Neural Network (CNN) models face challenges in balancing accuracy, efficiency, and interpretability for these tasks.

Purpose of the Study:

• To introduce a novel, lightweight, and efficient CNN architecture for multi-class image categorization in remote sensing.
• To enhance model interpretability using Shapley Additive Explanations (SHAP) and Class Activation Mapping (CAM).
• To evaluate the model's performance and generalizability across diverse remote sensing datasets.

Main Methods:

• Development of a hybrid CNN framework tailored for remote sensing image classification.
• Evaluation on NWPU-RESISC45 and UC Merced Land Use datasets.
• Integration of SHAP and CAM for model interpretability.
• Comparison against five popular pre-trained CNN models.

Main Results:

• The proposed CNN architecture achieved high accuracy (0.9428 on NWPU-RESISC45, 0.93 on UC Merced) and outperformed existing models.
• Achieved competitive recall (0.94, 0.93), precision (0.95, 0.94), IoU (0.89, 0.86), and F1-scores (0.94, 0.93).
• Demonstrated efficient training times (3,692s for NWPU-RESISC45, 559s for UC Merced) with manageable GPU memory usage.

Conclusions:

• The novel CNN architecture offers a compelling solution for remote sensing image understanding, balancing performance, efficiency, and interpretability.
• The integrated explainability techniques (SHAP, CAM) enhance the model's reliability.
• The model's generalizability across diverse datasets confirms its robustness for various remote sensing applications.