Summary
Highlights
Libby Bischoff introduces Dr. Tim Wallace and the annual Matson New York Times lecture, made possible by an endowed fund honoring Walter Matson. She shares Matson's connection to USM, his career at The New York Times where he championed modernization, and how Wallace's lecture aligns with Matson's forward-thinking spirit. Matson, a USM alumnus, rose from printer to president of The New York Times, computerizing the newsroom and launching the national edition, making him a fitting figure to honor with a lecture on advanced mapping technologies.
Dr. Matthew Edney introduces Tim Wallace, highlighting his unique career path from marine archaeologist to his current role at Descartes Labs. Wallace, now a creative director at Descartes Labs, focuses on interpreting and storytelling with data. He will discuss mapmaking in the age of artificial intelligence, aiming to demystify AI and machine learning for the audience, especially regarding its applications in geographic information and visualization.
Tim Wallace defines Artificial Intelligence (AI) as a computer system achieving goals in a complex environment and Machine Learning (ML) as a technique allowing computer systems to improve with experience and data. He distinguishes common robotic AI misconceptions from the code-based AI used in mapping. He illustrates ML with an example of classifying vegetation in New York City using satellite imagery, explaining how ML can refine traditional methods like NDVI (Normalized Difference Vegetation Index) by training algorithms with ancillary data like lidar to identify specific features, such as trees.
Wallace provides a historical overview of data collection for mapping, from 18th-century balloon aerial photography to early rocket-borne cameras in the 1940s and the explosion of satellite technology in the 1960s. He demonstrates the increasing density of Earth-observing satellites over time and the diverse types of data they collect (RGB, infrared, thermal, radar). He also covers modern sensors beyond satellites, including drones, airplane trackers, and even cameras used for crowd analysis, highlighting how these myriad data sources contribute to more detailed and frequent mapping.
Tim Wallace presents various applications of sensor data and AI. He shows how flight tracker data was used to define search areas for the Germanwings flight, and how Sentinel 5P satellite data maps atmospheric pollutants like nitrogen dioxide, revealing industrial, shipping, and urban patterns. He then illustrates how these data are used to create compelling visualizations, such as algal blooms in Lake Erie and time-lapse imagery of the Larsen Ice Shelf breakup. He also shares a personal example of mapping Maine using only satellite-derived data, demonstrating the power of these technologies to create detailed and artistic maps.
Wallace emphasizes that AI, despite its capabilities, is not infallible and is only as good as its implementation and human oversight. He humorously demonstrates AI's limitations using Photoshop's content-aware fill and scale features on a map of New York City, resulting in distorted or nonsensical landscapes. He also points out the presence of AI in daily life (e.g., Gmail's 'important' flagging) and discusses potential biases in AI, especially in geographical contexts where 'things look' different globally. He highlights the need for human experts to review and correct AI outputs, citing his New York Times project mapping buildings across the US where human intervention was crucial to refine Microsoft's machine learning data.
Wallace highlights Descartes Labs' work on mapping trees using lidar data to train algorithms, enabling precise tree mapping from traditional aerial imagery. He demonstrates this with maps of Washington D.C., Boston, and New York. He then transitions to wildfire detection, comparing traditional fire towers with current satellite monitoring via MODIS and VIIRS satellites. He introduces the revolutionary GOES-16 satellite, which provides images every 5 minutes, allowing for near real-time wildfire tracking and analysis like the Camp Fire. Descartes Labs is developing an AI-powered fire detection system that can alert officials within 10 minutes of ignition by comparing predicted images with real-time satellite data. He reflects on how machine learning offers a 'Holy Grail' for cartography by enabling highly detailed and accurate mapping.
Wallace shares additional examples of Descartes Labs' AI applications, such as mapping surface-to-air missile sites in North Korea, power substations, wind turbines, and solar panels globally, all based on unique visual patterns. He then opens the floor to questions, addressing topics like the minimum size/heat for fire detection (which depends on both), the challenges and opportunities for AI in archaeology (e.g., shipwreck detection), tools and platforms for mapping (Python, Jupiter notebooks for ML, various GIS software for visualization), and the persistent issue of cloud cover in satellite imagery. He also discusses the accuracy of AI algorithms and the importance of human expertise in fact-checking these maps, especially with open-source data. The lecture concludes with a discussion on potential applications of AI in conflict monitoring and the accessibility of data for both public and private entities.