The Role of Convolutional Neural Networks in Efficient Classification of Remotely Sensed Hyperspectral Images

In the remote sensing domain, significant advances have been made due to the development of several techniques for Hyperspectral imagery (HSI) classification. Achieving accurate classification of the remote sensing hyperspectral data is a huge challenge since they have complex characteristics. Several researchers have shown that the usage of Convolutional Neural Networks (CNN) along with several Deep learning models can be very efficient in addressing many hurdles of HSI classification like the inadequate extraction of spectral-spatial features, unannotated samples, excessive noise, the possibility of overfitting and non-linear problems of HSI. This paper provides a systematic overview of few CNN strategies that can efficiently perform accurate classification of different types of Hyperspectral images. Initially, we identify the significance and various challenges of HSI classification and how each variation of CNN combined with deep learning concepts offers good results and high-quality classification. Moreover, given the condition that in the remote sensing domain, existing samples needed for training are typically not sufficient, and deep networks need many samples for training, we discuss some approaches to alleviate classification performance. Further, we compare the performance, limitations and advantages of each CNN strategy so that we can assist future researchers to enhance the performance accuracy in case of hyperspectral image classification.