Hyperspectral image classification a comparative analysis of machine learning and deep learning approaches

Abstract

In recent years, remote sensing has become a highly interesting field due to its diverse applications. Among these, hyperspectral imaging stands out because of the vast amount of data contained within its hundreds of spectral bands and spatial features, making it highly effective for remote object identification, which is the focus of this work. Given its utility and necessity across various applications, it has become inevitable to delve deeper into this technology. However, handling such a vast and complex amount of data presents significant challenges. That’s why we proposed several methods to process this data by extracting the spectral and spatial features contained in hyperspectral images for pixel classification and, consequently, object identification with the highest accuracy, which is our ultimate goal. Firstly, we explored traditional machine learning methods, specifically K-Nearest Neighbors and Support Vector Machines, and demonstrated their limitations in Hyperspectral image classification. We then turned to more advanced deep learning methods, particularly convolutional neural networks. These techniques act similarly to the human brain, which is a very powerful tool for managing large and complex data. Our first approach focused on models that handle only spectral information of the HSI, specifically 1D-CNN, along with two additional techniques which are FFT and the Atrous algorithm to improve the processing and, hence, the accuracy of the models using only spectral information. The second approach dealt with the spatial features of the HSI using a 2D-CNN. The final method combined the processing of both spectral and spatial features while using a 3D-CNN. All the models were tested on two datasets, namely the Indian Pines and Pavia University datasets. The classification and identification results of our models on these datasets were promising, where the best performer was 3D-CNN with accuracy of 99.88 .