Change Detection on Sentinel-2 multi-spectral images: A Semi-(Non-)Supervised Learning approach

Abstract

Land cover change detection is a crucial task to automate for many applications ranging from efficient natural disaster monitoring or military surveillance to property insurance optimization. Space missions, notably Sentinel-2 from the European Space Agency have opened the path to a much wider use of multispectral geospatial imaging with high temporal resolution which proves ideal for the task of bitemporal change detection. In that context, the company Aerospacelab has implemented a neural network trained on a synthetic dataset in order to cope for the lack of labeled data. This thesis has two aims: evaluating the potential of multispectral channels in contrast to the three typical visual spectral channels Red-Green-Blue for the task of general change detection and establishing whether or not unsupervised or semi-supervised methods can better cope with the common lack of available labeled data.

To provide these insights, an unsupervised approach initially proposed by Gong et al. called Generative Discriminatory Classified Network (GDCN) is exploited, derived, implemented and tested. It consists in a generator to produce bitemporal multispectral satellite images and a discriminator which simultaneously discriminates between real and fake pairs of images as well as classify each of its pixels as change or non-change. To determine the impact of each of the approach's components on the performances for the task of change detection, three models are implemented: GDCN itself, ConvGDCN which is a scalable derivation of the model inspired from the Deep Convolutional GAN framework proposed by Radford et al. and finally, ConvCN which is a copy of ConvGDCN from which were removed the parts dedicated to the generation task in order to evaluate just how useful it is. Each model is trained in the unsupervised manner proposed by Gong et al. relying on pseudo labeled samples provided by another notable unsupervised method: Compressed Change Vector Analysis by Bovolo et al. The unsupervised models are tuned on a validation set and the best of each model trained on both RGB and multispectral images are fine-tuned with a small amount of hand-labeled samples.

All unsupervised and tuned models with both RGB and multispectral images are tested on a test dataset and the results provide insights to the company's questions. Specifically, this work shows that RGB images carry most of the information relevant to the task of general change detection and that most significant improvements can be obtained by better exploiting these RGB bands. Additional multispectral bands can still be used in specific contexts, for instance in the form of known spectral indices but otherwise complicate the learning task by adding more noise than useful information which creates a need for more complex models and larger datasets. As to the unsupervised and semi-supervised methods, this work shows that they can compare to the synthetic dataset-trained approach while not outperforming it. Nevertheless, our appproach can provide good pretrained weights to be tuned for a more specific task of change detection.