Multi-crop classi cation using hyperspectral data: comparison of classical machine learning and convolutional neural network techniques

Abstract

Hyperspectral remote sensing is one of the most exciting elds of remote sensing with the enormous size of data due to high spectral resolution. Hyperspectral data has both spatial and spectral components making a datacube which is very large to process due to a high number of features, a large volume of pixels and a low number of labeled pixels. Classi cation of crops classes using hyperspectral data is challenging due to small-sized classes and spectral similarity between various crops, so di erent pattern recognition techniques are developed to classify the data accurately. In the past, Spectral Angle Mapper and GIS techniques were used land cover classi cation. Classi cation using the supervised machine learning algorithms Support Vector Machine, Random Forest, k-Nearest Neighbors, and Multinomial Logistic Regression are used for classi cation. The performance at various training samples percentages (10%, 30%, 60%, and 80%) is analyzed for all classi ers. The classi ers' parameters are ne-tuned to extract the best decision boundary and better confusion matrix. SVM-RBF was the best classi er with an accuracy of 89.1%, 95.9%,94.8% and 99.8% across Indian Pines, Pavia, Salinas, and AVIRIS-NG datasets respectively In the second part of the thesis, Neural Networks and their various architectures have been analyzed. Convolutional Neural Network(CNN) and Recurrent Network are the main focus for Hyperspectral Remote Sensing Classi cation. The convolution operation helps in identifying neighborhood information and perform spatial feature extraction to get important features or patterns. The recurrent network helps train on a limited number of samples by getting better feature representation from the input data. 1D, 2D, and 3D CNN are the architectures on which the hyperspectral datasets are trained on with overall accuracies of 97.7%, 96.32%, and 98.80% respectively. The classical and deep learning techniques are compared and their results on open datasets and AVIRIS-NG dataset are discussed. 3D CNN deep learning model performance is found to be comparable with the SVM-RBF for classi cation purpose.