Deep Learning Techniques for Water Level Forecasting in Satellite Images

As the climate continues to change, the ability to automatically map bodies of water and predict their future variations is certainly very important, both to ensure that people are able to prevent and detect certain extreme weather events, such as floods and droughts, and to aid urban, agricultural and industrial planning around large lakes or waterways. Since the launch of the first satellites, images from orbit have been used to detect and monitor our planet's water masses, such as oceans and lakes. Over the years, increasingly accurate methods have replaced human work, and in recent years, deep learning algorithms such as Convolutional Neural Networks (CNNs) and Visual Transformers (ViTs) have achieved near-perfect performance in this task. However, although the detection of even the smallest lake from Earth orbit is now near perfect, we are still unable to predict changes in the size and shape of bodies of water with high accuracy. This thesis aims to address this issue by comparing the performance of different state-of-the-art deep learning models in both water detection and water level prediction using freely available satellite data. We not only compared different models and architectures, such as U-Nets and DeepLab for segmentation or ResNet and LeViT for forecasting, but also how varying the properties of the input data, such as the number and type of electromagnetic bands or the number of temporal observations, can change the behaviour of the models and the results obtained. The experimental results show that CNNs perform better in these tasks than previous state-of-the-art machine learning approaches, such as SVM and Random Forest, and classical index-based methods.