A Comprehensive Review on Segmentation Techniques for Satellite Images

Segmentation of satellite images is the noteworthy and essential step for better understanding and analysis in various applications such as disaster and crisis management support, agriculture land detection, water body detection, identification of roads, buildings, transformation analysis of forested ecosystems, and translating satellite imagery to maps, where the satellite image can be utilized for remotely monitoring any specified region. This manuscript contemplates the comprehensive and comparative analysis of existing satellite image segmentation techniques with their advantages, disadvantages, experimental results, and futuristic discussion. The comprehensive and comparative analysis provides the basic platform and a new direction of research to perspective readers working in this area. In this review, existing segmentation techniques are extensively analyzed and categorized on the basis of their methodology similarities. In the reviewing process of state-of-the-art satellite image segmentation techniques, it has been noticed that the problems of semantic and instance segmentation are solved effectively using deep learning approaches. The entire review process exhibits the problem of the limited dataset, limited time to train a network, objects appearing differently from different imaging sensors, and class imbalance in semantic and instance segmentation. A fully convolutional network, U-Net, and its variants are utilized to solve these problems by applying transfer learning, synthetic data generation, artificially generated noisy data, and residual networks. This manuscript focuses on the existing work and helps to provide comparative results, challenges, and further improvement areas.