The Role of AI in Improving Satellite Image Classification: A Literature Review

Abstract

As the need for precise land cover mapping and environmental monitoring continues to expand, better satellite image classification has grown in significance. New developments in artificial intelligence, from traditional forms to deep learning, have created even more promise for accurate and efficient classification of high-resolution satellite imagery. This literature review assesses AI, specifically Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), on their portfolios toward improving satellite image classification. We draw from fundamental experimental results found in three extensive studies, while presenting a comparative discussion of each of the AI-enabled models examined: ResNet, DenseNet, EfficientNet, and state-of-the-art Transformer models. In our review, we found that deep learning models, particularly those aided by attention mechanisms or multiscale feature fusion input layers, are adept at disambiguating the spectral, spatial, and temporal complexities of satellite data, which has traditionally caused challenges for remote imagery analysts. Current obstacles of data scarcity, an upper limit on what could be annotated with high financial cost, and computing infrastructure further perpetuate flawed and ineffective classification accuracy. This paper identifies key advancements made, the frequent issues associated with satellite imagery, and future directions, like self-supervised or any other potential learning model, a more efficient class of architectures that lessen compute time, and, more importantly, how to engage with satellite data in 3D. We hope that addressing these points forms the knowledge-sharing basis for other researchers interested in utilizing AI to enhance the value of satellite image classification.

Keywords: Satellite Image Classification, Deep Learning, Convolutional Neural Networks, Vision Transformer, Remote Sensing, Multispectral Imagery, Semantic Segmentation, AI in Geospatial Analysis