Multi spectral image classification based on deep feature extraction using deep learning technique

Remote sensing image classification accuracy depends on the extraction of Deep Feature Extraction. Unsupervised deep feature extraction employs single-layer and deep convolutional networks. Application of supervised convolutional networks is highly challenging for multi- and hyper-spectral imagery when input data dimensionality is high and labelled set is limited. To accomplish the mentioned, greedy layer-wise unsupervised pre-training combined with an appropriate algorithm for unsupervised learning of sparse features is proposed. This algorithm concentrates on sparse representations and sparsity of the extracted features at a time. The proposed method is applied for land use/cover classification of different spatial/spectral remote imagery. Comparing the current algorithms for classification, the proposed method performs well. Extraction of powerful discriminative features is possible with single-layer convolutional networks to obtain detailed results in classification. Different spatial/spectral parameters are calculated to quantify the results.