A learning architecture based on Convolutional Neural Network (CNN) and Graph Theory, applied on ground truth WND cases and satellite images is applied and tested. This process produces AI based risk maps to compare with classical statistical methods for evaluating the degree of improvement in forecasting the disease spread.
Train, fine-tuning and validation of the AI model
AI models/algorithms for the analysis and prediction of WND “behaviour” are developed and parameters estimated. Graph-based DNN models are explored for merging geo-referenced local sites information with satellite images, the latter being processed through Convolutional Neural Networks (pre-trained or trained from scratch). Temporal deep models (e.g. RNN - Recurrent Neural Networks, LSTM - Long-short term memory) are then employed for an effective forecasting of the behaviour based on EO data.
The accuracy of the chosen model are consequently evaluated, together with the need to include additional data or to change the train model hyper-parameters. In conclusion, the final model is compared with the classical statistical models developed in phase 2.