Hiba Najjar

• Remote sensing (RS) provides abundant and diverse data for Earth Observation (EO) applications. Machine learning leverages the available data through deep neural networks and specialized architectures. However, increasing model complexity often compromises its interpretability, which is crucial for many EO applications that monitor sensitive human activities or support natural disaster response efforts. This thesis contributes to advancing the interpretability and explainability of complex AI models for various RS applications, with a specific focus on agricultural activities. Our work employs eXplainable AI (XAI) methods to address two main objective for understanding and improving the model predictions within EO applications. First, we focus on XAI for Justification where the model behavior is justified by analyzing how different input features contribute to the outputs. The explanation of individual predictions are leveraged and aggregated to provide a broader understanding of the model’s behavior. We apply and evaluate existing model-agnostic and model-specific methods, while also developing new techniques when necessary. We further explore how multi-task learning can further enhance the explainability of predictions. Second, we apply XAI for Improvement based on insights from our prior model justification results. On one hand, we identify the features that are necessary and sufficient for accurate modeling across different contexts. On the other hand, we focus on optimizing the selection and design of vegetation indices, a key component in EO analysis and modeling. We benchmark our explainability objectives across multiple datasets, covering a range of tasks in EO. We particularly focus on multi-modal datasets, commonly used in EO, to mitigate the research gap regarding the explanation of complex multi-modal networks. The results demonstrate that our approach effectively explains the models by verifying that the model reasoning aligns with expert knowledge. Additionally, our experiments on vegetation indices and the optimization of models through feature reduction yielded promising results, and contributed to enhanced overall model performance and interpretability. Overall, this thesis provides a thorough examination of the interpretability of ML models under complex modeling scenarios. It leverages various explainability tools and objectives to justify model predictions and improve the modeling strategy and performance. This work contributes an important building block towards more transparent and better performing ML models designed for EO applications.