A Contrastive Continual Learning for the Classification of Remote Sensing Imagery

Since the emergence of computer science, the researching community has been pushing the boundaries to produce an artificially intelligent machines. One of the most sought-after goals is to mimic many of human abilities like incremental learning. In our case, we intend to perform remote sensing (RS) classification task, which takes images of earth/land captured by unmanned aerial vehicles (UAVs) and classifies them to a specific land type/use. In this paper, we propose a contrastive learning-based deep neural network model to classify remote sensing images in an incremental learning environment. The deep learning model is continually confronted with new tasks and data and has to perform well without seriously suffering from catastrophic forgetting. Experiments are conducted on well-known RS dataset: UC-Merced. The experiments results show substantial improvement of overall accuracy using contrastive losses compared to traditional Cross-Entropy loss in an incremental learning scenario.