6533b862fe1ef96bd12c5e6e

RESEARCH PRODUCT

Exploiting deep learning algorithms and satellite image time series for deforestation prediction

subject

description

In recent years, we have witnessed the emergence of Deep Learning (DL) methods, which have led to enormous progress in various fields such as automotive driving, computer vision, medicine, finances, and remote sensing data analysis. The success of these machine learning methods is due to the ever-increasing availability of large amounts of information and the computational power of computers. In the field of remote sensing, we now have considerable volumes of satellite images thanks to the large number of Earth Observation (EO) satellites orbiting the planet. With the revisit time of satellites over an area becoming shorter and shorter, it will probably soon be possible to obtain daily images of any area in the world. This availability of images allows to create time series of data called Satellite Image Time Series (SITS). SITS can be used for multiple real-world applications such as the prediction of land cover changes in general, and in particular the deforestation. The aim of this thesis is to exploit the potential of deep learning methods and the availability of SITS to create predictive models based on deep learningarchitectures, that will analyze historical data of a given area and will predict the deforestation in that area. Four main contributions are noted at the end of this thesis work: 1) Proposal of a workflow for batch collection and preprocessing of Sentinel-1 satellite images; 2) Comparison of three DL architectures for the task of predicting the next occurrence in a STIS; 3) Validation of DL methods for predicting land cover changes by comparison with the most method used in the literature (CA-Markov method); 4) Proposal of a model called (Deforest_Pred) for the prediction of deforestation around the Dja Biosphere Reserve (Cameroon). The Deforest_Pred model is based on a hybrid CNN-LSTM architecture and trained on Sentinel-1A images, augmented by an image fusion technique proposed in this study.