A Hybrid Framework Combining GAN-Based Augmentation and Ensemble Learning for Improved Satellite Image Classification

Satellite image classification is a complex task lies at the intersection of remote sensing, computer vision, and machine learning. The significant variability present in satellite data makes many existing land use land cover classification methods inadequate for effectively handling such datasets. Machine learning and transfer learning models have been widely used in satellite image classification. In contrast, ensemble learning approaches remain relatively new in this domain and have not been explored extensively, despite showing significant potential in many other domains. Another less explored area in satellite imagery is the application of Generative Adversarial Network (GAN), which can be used to augment the satellite image dataset and further increase the model performance. In this paper, we propose a framework for satellite image classification by incorporating GAN with ensemble learning. We aim to address the gap of using GAN and ensemble learning in satellite image classification and introduce a more efficient approach for classifying diverse scenarios across a broad spectrum of satellite images. Two different variants of GAN have been used in this research, i.e., DCGAN and WGAN-GP to generate synthetic data. In addition, we applied several transfer learning models for extracting features from the images. Afterwards, we used a feature selection algorithm, Particle Swarm Optimization (PSO) on the extracted features to select the most important and relevant features, reducing the computational complexity. Finally, we employed an ensemble learning classifier consisting of three machine learning models as base learners and Logistic Regression as the meta-learner to perform the final prediction. The proposed framework is empirically evaluated with several state-of-the-art approaches where it outperformed the existing approaches by achieving 96.03% accuracy.