Unsupervised Deep Feature Extraction for Remote Sensing Image Classification

This paper introduces the use of single layer and deep convolutional networks\nfor remote sensing data analysis. Direct application to multi- and\nhyper-spectral imagery of supervised (shallow or deep) convolutional networks\nis very challenging given the high input data dimensionality and the relatively\nsmall amount of available labeled data. Therefore, we propose the use of greedy\nlayer-wise unsupervised pre-training coupled with a highly efficient algorithm\nfor unsupervised learning of sparse features. The algorithm is rooted on sparse\nrepresentations and enforces both population and lifetime sparsity of the\nextracted features, simultaneously. We successfully illustrate the expressive\npower of the extracted representations in several scenarios: classification of\naerial scenes, as well as land-use classification in very high resolution\n(VHR), or land-cover classification from multi- and hyper-spectral images. The\nproposed algorithm clearly outperforms standard Principal Component Analysis\n(PCA) and its kernel counterpart (kPCA), as well as current state-of-the-art\nalgorithms of aerial classification, while being extremely computationally\nefficient at learning representations of data. Results show that single layer\nconvolutional networks can extract powerful discriminative features only when\nthe receptive field accounts for neighboring pixels, and are preferred when the\nclassification requires high resolution and detailed results. However, deep\narchitectures significantly outperform single layers variants, capturing\nincreasing levels of abstraction and complexity throughout the feature\nhierarchy.\n