12-05, 19:30–20:00 (UTC), AI/ML Track
This talk will uncover the power of AI in combating Amazon deforestation through an innovative cattle detection system. We present a cutting-edge approach to monitoring illegal ranching, a primary driver of deforestation, using very high-resolution satellite imagery and deep learning. We'll dive into the unique challenges of detecting cattle from space – from congested scenes with small, clustered targets to diverse and cluttered backgrounds – and how we overcame them with a two-step neural network approach. By combining classification and density estimation techniques, our model efficiently identifies potential cattle locations and estimates herd sizes across varied landscapes. Discover how this interdisciplinary project, developed in collaboration with Brazilian prosecutors, leverages data science to drive real-world impact in environmental conservation and sustainable land management. Join us to explore the intersection of computer vision, geospatial analysis, and environmental advocacy, and learn how AI can be a powerful tool in the fight against deforestation in the Amazon and beyond.
This presentation is aimed at individuals interested in bridging the gap between computer vision challenges and practical applications that yield tangible impacts. Our interdisciplinary research focuses on addressing the urgent need to mitigate deforestation in the Amazon. The talk will be structured as follows:
Motivation (0'-3'):
Discuss the critical importance of halting deforestation in the Amazon and its implications for biodiversity, climate change, and local communities.
Introduction (3'-7'):
Overview of animal counting using satellite imagery: a review of existing methodologies and models utilized in previous studies.
Model Selection (7'-10'):
Justification for employing a congested scene recognition approach over traditional object detection frameworks. Explore the advantages of this model in handling the unique challenges presented by cattle detection in satellite images.
Methods in Detail (10'-15'):
Framework for establishing a deep learning model tailored to environmental challenges:
Training Dataset: Strategies for assembling and ground-truthing the dataset, ensuring accuracy and reliability.
Model Architecture: Discussion of the classification model and density estimation techniques implemented.
Evaluation Metrics: Overview of metrics used to assess the performance of the test set.
Results (15'-25'):
Focus Regions: Examination of data science challenges related to cattle detection in Indigenous lands.
Scalability: Methodology for generating a comprehensive cattle density map for extensive regions in Brazil.
Caveats: Address limitations concerning image availability, temporal and spatial coverage, and the variability of image backgrounds.
Conclusion and Wrap-Up (25'-30'):
Summarize key findings and implications of the study, emphasizing the potential of AI and geospatial analysis to foster sustainable land management practices both in the Amazon and globally.
No previous knowledge expected
Leonie is a land system scientist and postdoctoral researcher at the Global Land Use and Environment Lab at the University of Wisconsin-Madison. With a background in bioinformatics, she recently earned her PhD from ETH Zurich in Switzerland. Her research is dedicated to understanding deforestation trends in tropical regions and evaluating the effects of both private and public conservation interventions. Leonie's work employs AI-driven large-scale geospatial analysis, complemented by qualitative methods, to explore and analyze complex land systems.