Predicting the occurrence of gullies at the continental scale with deep learning

Abstract

Gully erosion is the removal of soil by surface runoff, typically on hillsides. This process creates what we called a gully, which is represented by a ravine with a head-cut and high slope variations. Gully erosion has negative impacts on people's activities but also on the infrastructures. Some regions, such as Africa, suffer more from these impacts. Several researches were made to understand Gully erosion and build models to detect gullies in order to potentially reduce its impacts. However, current models do not work at a country-wide or continent-wide scale. The purpose of this master thesis is to determine if neural networks can be used as models to predict gullies with the help of freely worldwide available satellite data. The first step was to create a database composed of satellite images representing gully and no-gully points. To perform the predictions, three different models were used: a baseline, a CNN, and a ResNet. These models were tested on different data sets i.e., Ethiopia and Africa data sets. The data sets are based on the satellite images database from which different features are extracted. Results obtained with baseline model are closed to a random classifier for any data set configuration. The CNN and ResNet perform better on the configuration of some data sets, in particular on the data sets that do not separate sites between train and test sets. This non-distinct separation between sites implies models are good at predicting similar images than those seen during the training. When data set configuration forced a distinction in sites used to train and test the models, which is the configuration required to have models able to predict gullies on a totally new site, results obtained are also close to a random classifier. These results point out some limitations of the current database i.e. data points similarity, the quantity of labeled data, and the resolution of satellite images. However, these limitations could be reduced and future work ideas, with preliminary results, are proposed in this work. These perspectives show some improvements with an accuracy close to 0.6. The work achieved in this master thesis is a first step for the development of deep learning architectures to predict gully headcuts at a large scale. This brings some guidance to improve results and possibly to have powerful model.