Improving and exploiting raw inferences from CoSIA maps

This thesis investigates the performance and enhancement of CoSIA (Couverture du Sol par Intelligence Arti cielle ; Landcover by Arti cial Intelligence in English) land-cover maps produced by two deep-learning models, U-net Resnet34 and Swin UPerNet. The research is divided into two main parts. The rst part assesses the raw inferences from these AI (Arti cial Intelligence) models, developing protocols to improve map quality and identifying areas with unreliable predictions. A novel staggered sieve method was introduced, improving the metrics (Intersection over Union) for individual classes, justifying its integration into the production process of vectorized CoSIA maps. Additionally, the research highlights the importance of model calibration, noting that while most classes are well-calibrated, certain challenging classes, such as vineyards and swimming pools, showed less accurate predictions. The development of a hesitation index further enhanced the evaluation process by visualizing the spatial distribution of model uncertainty, providing valuable insight for future re nements of CoSIA maps. The second part of the thesis demonstrates the practical applications of CoSIA maps through two use cases: a bi-temporal change detection protocol and a risk-assessment tool for windmill placement. These applications showcase the maps' utility in targeted change detection and preventive risk analysis, respectively. Overall, this work lays a foundation for future improvements and broader applications of CoSIA maps, providing powerful tools for land-cover mapping and analysis

This thesis investigates the performance and enhancement of CoSIA (Couverture du Sol par Intelligence Arti cielle ; Landcover by Arti cial Intelligence in English) land-cover maps produced by two deep-learning models, U-net Resnet34 and Swin UPerNet. The research is divided into two main parts. The first part assesses the raw inferences from these AI (Arti cial Intelligence) models, developing protocols to improve map quality and identifying areas with unreliable predictions. A novel staggered sieve method was introduced, improving the metrics (Intersection over Union) for individual classes, justifying its integration into the production process of vectorized CoSIA maps. Additionally, the research highlights the importance of model calibration, noting that while most classes are well-calibrated, certain challenging classes, such as vineyards and swimming pools, showed less accurate predictions. The development of a hesitation index further enhanced the evaluation process by visualizing the spatial distribution of model uncertainty, providing valuable insight for future re nements of CoSIA maps. The second part of the thesis demonstrates the practical applications of CoSIA maps through two use cases: a bi-temporal change detection protocol and a risk-assessment tool for windmill placement. These applications showcase the maps' utility in targeted change detection and preventive risk analysis, respectively. Overall, this work lays a foundation for future improvements and broader applications of CoSIA maps, providing powerful tools for land-cover mapping and analysis