Efficient Classification of Hyperspectral Image s Using Multiscale Relation Learning

The classification of hyperspectral images (HSIs) has emerged as a key area of interest in remote sensing and has prompted the investigation of several approaches in recent years. Because of its strong feature extraction capabilities, deep learning has become popular; nevertheless, because of its low discriminative capability, traditional methods frequently produce subpar results. Furthermore, the lack of labelled data makes it difficult to achieve high classification accuracy with small sample sizes in HSI classification. This study suggests a novel method that combines convolutional neural networks (CNNs) with different feature learning techniques to address these problems. The model creates pertinent feature maps for each input feature by feeding a customized CNN architecture with a wide variety of characteristics that are directly retrieved from raw images. These joint feature maps are then used by further layers to forecast final labels for each pixel in the hyperspectral image and this method takes advantage of CNNs' improved feature extraction capabilities while efficiently combining spectral and spatial information to maximize the generated features' discriminative potential. The development of hyperspectral imaging technology, which enables sensors to take pictures in hundreds of bands, highlights the importance of hyperspectral image classification for remote sensing applications. This technology is useful in many different fields, such as monitoring geological disasters, military reconnaissance, monitoring vegetation and ecology, and atmospheric assessment. With multiscale relation learning we have achieved the accuracy rate of 96.13% and 82.6% recall as well.