Convolutional LSTMs for automatic mapping of spatio-temporal evolution of natural and anthropogenic features in West African Littoral

Abstract:

The West African coast is one of the worldwide littoral areas most threatened by the ongoing climatic and anthropogenic global changes: with a low topography, a weak geological substratum, a poor fresh water supply, and a dense and rapidly expanding population, the West African littoral is highly vulnerable to current sea level rise, extreme climatic phenomena, erosion, and modifications of the ecosystems and resources. With the development of publicly available medium-high resolution remote sensing data such as the multi-spectral image time series provided by the Sentinel-2 mission, the West African littoral can now be monitored on a large scale and this calls for increasingly high-performing automated machine-learning based mapping tools. In this work, we exploit both the temporal and spatial information provided by the Sentinel-2 Earth observation mission, to generate accurate land cover maps, that can serve to monitor the West African coast and track changes. Our approach consists in extracting multi-temporal features, which are then filtered to reduce temporal noise induced by atmospheric perturbation or surface reflectance variation. For this purpose, we propose a hybrid deep learning architecture combining a Fully Convolutional Network (FCN) with a Recurrent Neural Network (RNN). Our FCN is based on an adaptation of U-NET that captures the spatial representation of features in the image. It is made up of a contracting path that captures the context and a symmetric expanding path that enables precise localization. Our RNN is based on a Convolutional Long Short Term Memory (ConvLSTM), a type of RNN able to learn the temporal information, while preserving the spatial information. Since we require public and free data, we use online map data from Open Street Map which provides semantic labeling for many places in the world. Preliminary results show that the temporal information provided by time series images allows increasing the accuracy of land cover classification, thus producing up-to-date maps that can help in identifying spatio-temporal changes in the West African Littoral.