Towards automated satellite image segmentation and classification for assessing disaster damage using data-specific features with incremental learning

Abstract

In this paper, an online classification strategy is adopted where a classifier is built incrementally using the streaming damage labels from various sources as training samples, i.e. without retraining it from the scratch when new samples stream in. The Passive-Aggressive online classifier is used for the classification process. Apart from the classifier, the choice of image features plays a crucial role in the performance of the classification. The features extracted using recently reported deep learning approaches such as Convolutional Neural Networks (CNN), which learns features directly from images, have been reported to be more effective than conventional handcrafted features such as gray level co-occurrence matrix and Gabor wavelets. Thus in this study, the potential of CNN features is explored for online classification of satellite image to detect structural damage, and is compared against handcrafted features. The feature extraction and classification process is carried out at an object level, where the objects are obtained by over-segmentation of the satellite image. The proposed online framework for damage classification achieves a maximum overall accuracy of about 73%, which is comparable to that of batch classifier accuracy (74%) obtained for the same training and testing samples, however at a significantly lesser time and memory requirements. Moreover, the CNN features always significantly outperform handcrafted features.