No abstract

No abstract

Hyperspectral sensor generates huge datasets which conveys abundance of information. However, it poses many challenges in the analysis and interpretation of these data. Deep networks like VGG16, VGG19 are difficult to directly apply for hyperspectral image (HSI) classification because of its higher number of layers which in turn requires high level of system resources. This article suggests a novel framework with lesser number of layers for hyperspectral image classification (HSIC) that takes into account spectral-spatial context sensitivity of HSI, which focuses on enhancing the discriminating capability of HSIC. The model uses available spectral feature as well as spatial contexts of HSI and consecutively learn the distinctive features. A small training set has been used to optimize the network parameters while the overfitting problem is alleviated using the validation set. Regularization has been performed using batch normalization (BN) layer after each convolution layer. The cost of the model is measured in terms of training and testing time duration under the same platform, which has further been compared with some ensemble learning methods, SVM and other three recent state-of-the-art methods. Experimental results establish that the proposed model performs very well with the three benchmark datasets: Indian Pines, Salinas and University of Pavia, which mostly contain land cover of agriculture, forest, soil, rural, and urban area.

No abstract

No abstract