Selecting the best backbone model: A comprehensive evaluation of deep learning models for satellite image-based land cover classification

Satellite image-based scene classification or land cover classification is an active area of research that has significantly benefited from the state-of-the art deep learning models. To address various problems in this domain, such as high resolution satellite image analysis, real-time monitoring via satellite, invariantly detect/recognize the landmark objects etc, most of the times, the newly proposed deep learning architectures extend the existing ones, use them as backbone architectures for image features extraction or employ their representative blocks. Hence, there is a dire need to extensively evaluate those existing architectures or their pre-trained models for the task of land cover classification in satellite images. Such evaluation will make the selection process of the backbone model convenient thus allowing for a more robust, efficient and accurate novel architecture. To this end, I extensively evaluate 50 pre-trained models on 10 publicly available land cover classification datasets. The pre-trained models encode/represent images which are then used to train and evaluate a linear Support Vector Machine (SVM). The motivation to use this methodology is to keep the comparison process simple and quick while at the same time demonstrate the performance of pre-trained models on all the datasets. I report the Top-1%, Top-3% and Top-5% accuracies achieved by all the pre-trained models on all the datasets. Additionally, I report the time taken for 5-fold cross-validation by each pre-trained model, enabling the identification of models that are both highly accurate and time-efficient.