Automated dense-layer architecture search on EfficientNet: A hybrid approach for scene-based land-cover classification

This article proposes a land cover classification framework based on EfficientNet-B4, a model from the EfficientNet family of convolutional neural networks developed by Google AI. EfficientNet models have been widely applied in various domains, including image classification, object detection, and medical imaging, due to their scalability and efficiency. EfficientNet-B4 and its pretrained weights were used for feature extractions, and the dense layer structure (number of layers, nodes, and dropout rates) was tuned using meta-heuristic optimisation algorithms. The model is trained and validated on the Sentinel-2 EuroSAT benchmark, which comprises 27,000 RGB (Red-Green-Blue) image tiles spanning 10 land-cover classes across Europe. The results show that this proposed model achieves an overall classification accuracy of 0.9881 for RGB images, which is higher than that of previous networks using similar datasets. This hybrid approach can be considered as an alternative solution to search for neural architectures for different applications.