From Plant Detection to Satellite Mapping: A Multi-Scale AI Toolkit for Ragweed Surveillance under Climate Change

Climate change is altering weed population dynamics, presenting dynamic management challenges that static decision rules cannot address. This chapter introduces and assesses an open-source surveillance toolkit comprising artificial intelligence tools for identifying and mapping Ambrosia artemisiifolia, commonly known as common ragweed, at multiple spatial scales. The toolkit was tested in two phenological stages (seedling phase in September 2024 and adult phase in December 2024) within a lentil field in central Chile.

At the field level, the toolkit employs a YOLOv11 detection model in conjunction with Slicing Aided Hyper Inference, achieving an mAP50 score of 0.886. However, when deployed in new geographic locations, the single-site model exhibits a significant performance drop, with mAP50 plummeting to 0.108. This phenomenon is termed as catastrophic collapse. Nevertheless, multi-domain training effectively mitigates this issue, restoring performance to 0.874.

At the satellite scale, the PRESTO foundation model embeddings demonstrate a strong correlation with ground-truth weed density, with a correlation coefficient of 0.739. Furthermore, geographically weighted regression explains up to 91% of the local density variation. The methodology and tools developed throughout the study are made available as the ragweed-ai-toolkit, an open-source Python package designed for modularity and ease of adaptation to new species and regions. The toolkit is accessible at https://github.com/agroia-lab/ragweed-ai-toolkit.