Research Projects

Remote sensing images are crucial for observing Earth's surface and providing valuable information for intelligent Earth observation. However, the large size of these images demands efficient classification of pixels and scenes. Challenges in classifying RS images include significant heterogeneity within classes, high similarity between classes, considerable variation in scene/object scales, and the presence of multiple ground objects. This study aims to propose efficient and robust classification models to address these issues and challenges. The study aims to formulate a strategic hybridization of two frameworks: granular computing (GrC) and deep learning (DL)-based representative feature extraction and classification. GrC works with human principles through granulation of information and abstract reasoning, reducing computational burden and facilitating the use of information granules. The benefits of GrC can be realized by incorporating the framework with neural networks (NNs). The second framework of the proposed model, DL NN, has been successfully used to address complex decision-making systems issues. High-quality features are needed to effectively classify real-time and complex data sets, like RS images. Engineered and hand-crafted features from standard methods do not meet these requirements and involve significant human intervention, leading to data-independent decision-making and inferior performance. The proposed model focuses on auto-encoder and convolution neural networks. The objectives include a comprehensive study on the basic architectures of DL NN and GrC frameworks, a strategic hybridization framework for RS image classification, and testing the applicability of the proposed models to multispectral RS images and hyperspectral RS images.

0