Spectral-spatial classification of remotely sensed hyperspectral images hasbeen the subject of many studies in recent years. Current methods achieveexcellent performance on benchmark hyperspectral image labeling tasks when asufficient number of labeled pixels is available. However, in the presence ofonly very few labeled pixels, such classification becomes a challengingproblem. In this paper we propose to tackle this problem using convolutional neuralnetworks (CNNs) and data augmentation. Our newly developed method relies on theassumption of spectral-spatial locality: nearby pixels in a hyperspectral imageare related, in the sense that their spectra and their labels are likely to besimilar. We exploit this assumption to develop 1) a new data augmentationprocedure which adds new samples to the train set and 2) a tailored lossfunction which penalize differences among weights of the network correspondingto nearby wavelengths of the spectra. We train a simple single layerconvolutional neural network with this loss function and augmented train setand use it to classify all unlabeled pixels of the given image. To assess the efficacy of our method, we used five publicly availablehyperspectral images: Pavia Center, Pavia University, KSC, Indian Pines andSalina. On these images our method significantly outperforms other baselines.Notably, with just 1% of labeled pixels per class, on these dataset our methodachieves an accuracy of 99.5%, etc. Furthermore we show that our methodimproves over other baselines also in a supervised setting, when no overlapbetween train and test pixels is allowed. Overall our investigation demonstrates that spectral-spatial locality can beeasily embedded in a simple convolutional neural network through dataaugmentation and a tailored loss function.
展开▼
