Forecasting landslide behavior with Deep Learning algorithms for Early Warning purposes

Development of a Deep Learning algorithm for landslide forecasting. It uses past displacements and other optional information to predict the next displacements. Many versions of this algorithm are tested, changing input data, underlying Neural Network, loss function, and other characteristics. Development of a Deep Learning algorithm for tree species recognition from drone images. This is achieved with models pre-trained on ImageNet, and subsequently fine-tuned on task-specific images. The output produced can be used as additional input for the landslide forecasting algorithm. Development of various auxiliary functions like custom loss, custom metrics, automatic hyperparameter space search, data cleaning and processing, data augmentation… Test of both algorithms on real world data with presentation and discussion of the results.