High-Resolution Satellite Image Classification Using Lightweight CNN with Depth wise Separable Convolutions.
📌 Introduction
This repository contains the implementation of multiple state-of-the-art deep learning architectures for satellite image classification using high-resolution satellite imagery. The models are designed to optimize computational efficiency while maintaining high classification accuracy, achieving an average accuracy of 97.3% across all models.
🚀 Models Used
The following architectures have been fine-tuned and optimized to enhance disease classification performance:
-
DenseNet121 (DenseNet121_model.ipynb): Efficient feature propagation and reduced computational overhead.
-
NASNetALarge (NASNetALarge_model.ipynb): Reinforced learning-based architecture with adaptive feature recalibration.
-
SENet (SENet_model.ipynb): Squeeze-and-Excitation blocks for dynamic channel-wise feature enhancement.
-
SE-ResNeXt50 (Se_ResNeXt50_model.ipynb): Multi-branch feature extraction with residual learning and attention mechanisms.
-
TinyNet (TinyNet_model.ipynb): Lightweight yet powerful convolutional model for real-time inference.
📊 Key Features & Enhancements
-
Adaptive Feature Recalibration: Models incorporate channel and spatial attention to refine feature extraction.
-
Progressive Resizing & Test-Time Augmentation (TTA): Enhances model robustness and generalization across different image scales.
-
Self-Distillation & Knowledge Transfer: Improves feature learning and optimizes the performance of smaller architectures.
-
Class-wise Dataset Partitioning: Balanced dataset representation to mitigate bias and improve performance consistency.
📈 Performance Metrics
Each model has been evaluated using rigorous performance metrics:
-
Accuracy: Average 95.7% across all models.
-
ROC-AUC Curve: Comprehensive evaluation of classification thresholds.
-
Precision-Recall Analysis: Ensures robustness against class imbalance.
-
Confusion Matrix Visualization: Provides insights into misclassifications.
| Parameter | Value |
|---|---|
Optimizer |
Adam |
Learning Rate |
0.0001 |
Batch Size |
32 |
Epochs |
50 |
Loss Function |
Categorical Crossentropy |
Weight Initialization |
Xavier Initialization |
Activation Function |
ReLU (Hidden Layers), Softmax (Output Layer) |
Dataset Split |
70% Training, 15% Validation, 15% Testing |
Image Size |
256 x 256 pixels |
Number of Classes |
5-Fold |
📬 Contact
For queries, feel free to reach out:
GitHub: ArnavGhosh999
Email: arnav032919@gmail.com