Evaluation of deep learning and transform domain feature extraction techniques for land cover classification: balancing through augmentation

Abstract

The identification of features that can improve classification accuracy is a major concern in land cover classification research. This paper compares deep learning and transform domain feature extraction techniques for land cover classification of SAR data on balanced and imbalanced training sets. Convolutional autoencoders (CAE), variational autoencoders (VAE), and Haar wavelet transforms (HWT) are used and evaluated for feature generation capability. Variations in features of CAE and HWT help gather more information about the image patch. The fusion of CAE and HWT features provides a combination of high- and low-frequency coefficients, respectively, which improves classification accuracy for various land covers. To assess features generated through fusion of features, RISAT-1 C-band and AIRSAR L-band datasets are used. Furthermore, agricultural/grass land has similar features with open forest which leads to misclassification of forest in agricultural/grass land. Increasing the number of samples in each class using the Synthetic Minority Oversampling Technique (SMOTE) increases training samples. Hierarchical classification of the above-mentioned features, where agricultural/grass land and forest classes are discriminated, improves classification results. This paper evaluates all the three types of features and fused features and provides a guidance for land cover classification.

Keywords: Deep learning; Land cover; Machine learning.