Enhancing Hyperspectral Image Classification for Land Use Land Cover With Dilated Neighborhood Attention Transformer and Crow Search Optimization

Abstract

The classification of Land Use Land Cover (LULC) can be accomplished with the help of hyperspectral imaging, which is a cutting-edge technology. Nevertheless, despite its efficacy, the utilization of hyperspectral images for LULC classification continues to present difficulties and demands a significant amount of time. The limited availability of training samples for hyperspectral images poses a challenge in achieving accurate classification of LULC. Nevertheless, through meticulous deliberation and examination, this impediment can be surmounted. To tackle the task of LULC classification, we have developed a Dilated Neighbourhood Attention Transformer (DNAT). Firstly, we employ LeNet-5 to extract features from the provided data. Subsequently, we perform band selection using Crow Search Optimization (CSO). Following the extraction of features and selection of bands, we proceed to classify LULC. In our study, we used the Salinas, Indian Pines (IP), and Washington DC Mall datasets for LULC classification. The performance of our proposed classification approach is evaluated using the commonly used metrics, namely, Average Accuracy (AA), Overall Accuracy (OA), and Kappa Coefficient (KC). We have achieved 99.85% as OA, 99.83% as AA, and 99.73% as KC for the Salinas Dataset. This is the highest accuracy we have achieved using the DNAT classifier. The experimental results proved beyond a reasonable doubt that the proposed method achieved the highest possible performance, surpassing all prior methods.