GrI-CNN: Granulated Deep Learning Model with Interpretable Architecture for Remote Sensing Image Classification

Convolutional neural networks (CNNs) are highly effective deep learning architectures for remote sensing (RS) image classification. However, the interpretability of CNN architecture remains challenging for further performance improvement. To address this issue, we propose an end-to-end interpretable CNN architecture called granulated interpretable CNN (GrI-CNN) within the granular computing (GrC) framework. The GrI-CNN uses fuzzy and rough sets to make each architecture component functionally interpretable. Fuzzy sets perform class-dependent granulation of the input feature space, while rough sets granulate the information with operations such as reduct for dimension reduction, functional dependency (FD) of samples for the optimal selection of filters, and weighted membership of granules. The decision layer of GrI-CNN measures the roughness of overlapping granules, encodes the domain knowledge, and initializes the weights using weighted membership and roughness measures. We combined two classification networks at the decision layer to achieve the best possible performance: FD-based interpretable-extreme learning machine (I-ELM) and knowledge-encoded evolving granular neural network (e-GNN). E-GNN is a kind of GNN in which the shape and size of granules evolve based on the user's needs. Thus, GrI-CNN uses only the required weight parameters and reduces computational time. We have demonstrated the superiority of GrI-CNN over similar state-of-the-art models for classifying multispectral and hyperspectral RS images.