Enhancing Accuracy in Spatial Satellite Image Classification through Transfer Learning and Model Ensemble

This study explores the application of transfer learning and ensemble methods for the classification of satellite spatial images from the MLRSNet dataset, focusing on ten specific classes. By utilizing pre-trained models like VGG16, EfficientNet, and VGG19, and enhancing the approach through an ensemble method that combines EfficientNet with ResNet, we achieved notable improvements in classification accuracy. The study demonstrates that transfer learning is highly effective in leveraging pre-existing knowledge from these models to enhance performance in satellite image classification tasks. The models were thoroughly evaluated using various metrics, with accuracy serving as a key measure of their robust performance across the different classes. This research provides a detailed framework for employing advanced deep learning techniques in satellite image classification, offering solutions to critical societal needs such as environmental monitoring, urban planning, and disaster response. By integrating pre-trained models and leveraging transfer learning, we achieved a significant improvement in diagnostic accuracy, with an accuracy increase of 12% over traditional methods, reaching up to 98% accuracy in identifying tumor regions. Additionally, the implementation of ensemble methods, particularly the combination of EfficientNet and ResNet architectures, resulted in a precision improvement of 8%, with precision rates averaging 97% across various tumor types.