Transfer Learning with Res2Net for Remote Sensing Scene Classification

Remote sensing is the method of obtaining data of a particular geographic region of interest with the help of sensors. Often, this data is in form of images, and analyzing such images can give useful information that can help mankind in several ways. In recent years, several methods have been proposed for the same purpose. Most of the successful ones are based on deep learning. Deep learning architectures have a large number of randomly-initialized weights (or parameters) that are trained to extract features from images. Often these weights fail to fully train due to a lack of proper and sufficient training data. To address this issue, transfer learning was introduced. Transfer learning is a technique inspired by humans, which involves initializing deep learning models with weights pretrained on a separate task and fine-tuning them further by training on the desired data. In this work, the same technique is used for classifying remote sensing scenes, with a newly found novel architecture called Res2Net. A single Res2Net block extracts multi-scale features from the input using several receptive layers hierarchically connected to each other with residual-like connections. Although this technique slightly increases the total number of weights, it is capable of capturing relevant and complex features at a granular level. The performance of the proposed approach is evaluated on three remote sensing datasets - UC Merced, Brazilian Coffee Scenes, and EuroSAT dataset and classification accuracies of 98.76%, 93.25%, and 97.50% are achieved, respectively. The method is tested with two fine-tuning techniques on different variations of Res2Net-50. Comparisons with other recently proposed methodologies are shown and confusion matrices are further plotted to better understand the classification potential. The PyTorch code for this work is available at https://github.com/iamarijit/res2net-remote-sense.