This framework demonstrates improved accuracy in land cover classification using EfficientNet-B4, suggesting better architecture search methods for image classification tasks.
Key Points
The model achieves an impressive overall classification accuracy of 0.9881 for RGB images, indicating significant improvement over previous networks.
Using pretrained weights from EfficientNet-B4, feature extraction was performed effectively, enhancing the model's performance in land cover classification.
Dense layer structures were optimised using meta-heuristic algorithms, allowing for dynamic tuning of layers and nodes—improving classification outcomes.
Training and validation were conducted on the Sentinel-2 EuroSAT benchmark, which provides a rich dataset of 27,000 RGB image tiles across 10 land-cover classes.