Abstract
Abstract. Many Earth observation programs such as Landsat, Sentinel, SPOT, and Pleiades produce huge volume of medium to high resolution multi-spectral images every day that can be organized in time series. These time series are a great opportunity to detect and measure the space and time changes of anthropogenic and natural features. In this work, we thus exploit both temporal and spatial information provided by these images to generate land cover maps. For this purpose, we combine a fully convolutional neural network with a convolutional long short-term memory. Implementation details of the proposed spatio-temporal neural network architecture are provided. Experimental 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 changes on earth in both time and space.
Highlights
Several places on Earth are threatened by the ongoing climatic and anthropogenic global changes: in particular, with a low topography, a weak geological substratum, a poor fresh water supply, and a dense and rapidly expanding population, some littoral areas are highly vulnerable to current sea level rise, extreme climatic phenomena, erosion, and modifications of the ecosystems and resources
Many Earth observation programs such as Landsat, Sentinel, SPOT and Pleiades produce huge volume of medium to high resolution multispectral images every day that can be organized in time series and used to produce accurate and up-todate land cover maps that can monitor environmental changes at different places and time ranges
We focus on Long Short Term Memory (LSTM), an evolution of Recurrent Neural Networks (RNN) which solves the problem of gradient explosion and gradient disappearance in RNNs
Summary
Several places on Earth are threatened by the ongoing climatic and anthropogenic global changes: in particular, with a low topography, a weak geological substratum, a poor fresh water supply, and a dense and rapidly expanding population, some littoral areas are highly vulnerable to current sea level rise, extreme climatic phenomena, erosion, and modifications of the ecosystems and resources. Land cover mapping is a semantic segmentation problem: each pixel in a satellite image must be classified into one of the land cover classes of interest These classes describe the surface of the Earth and are typically broad categories such as ”water”, ”roads”, ”low vegetation”, ”forest”, ”building”, etc. Throughout the years of research, a wide family of methods have been proposed, ranging from the classification of individual pixels with machine learning techniques, to the incorporation of higher-level information such as shape features. For this task, supervised machine learning algorithms have shown their potential, especially traditional algorithms such as Random Forests (RFs) and Support Vector Machines (SVMs). In (Thanh Noi, Kappas, 2018), the authors provide a comparison of Random Forest, k-Nearest Neighbor, and SVM classifiers for land cover classification using Sentinel-2 imagery
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
