Knowledge-Based Progressive Granular Neural Networks for Remote Sensing Image Classification

Reforming the architecture of a granular neural network (GNN) with varying input information further improves its operational transparency in the classification of land use/land cover of remote sensing data. Recently, an evolving GNN (eGNN) has been originated with an advancement to the GNN that has the ability to explore and accommodate input information more effectively and results in the performance improvement of classification systems. For each input, the shape and the size of granules are adjusted with randomly selected weight parameters, and accordingly, the network architecture of the eGNN is determined. In this direction, the present paper aims to improve the performance of eGNN by deterministically initializing its weight parameters and reforming the network architecture through encoding the domain knowledge using a rough-set-theoretic method. We also propose to use a back error propagation learning algorithm for the eGNN. The resultant network called a progressive GNN(PGNN) becomes more dynamic in the learning process and thus minimizes the learning time with improved accuracy. To justify the claim, we developed a PGNN-based classification model for land cover classification of both multispectral and hyperspectral remote sensing images. Superiority of the model to similar other methods has been validated with performance measurement metrics such as overall accuracy, precision, recall, kappa coefficient, dispersion score, and computational time.