Microsoft AI for Good Lab

The full scale of the earthquakes in Venezuela is only beginning to emerge. IOM is working to identify where needs are greatest, helping ensure emergency assistance reaches people affected by the earthquakes as quickly as possible.

🗺️ Connecting AI research across the Global South and Global North Sharing African AI research with the world is essential for building inclusive digital futures. The growing collaboration between the Deep Learning Indaba and the IJCAI International Joint Conferences on Artificial Intelligence Organization (IJCAI) exemplifies this effort — and United Nations University Institute in Macau is proud to be part of it. What began in 2023 with the inaugural IJCAI AI for Good Symposium in South Africa (8 papers published in the proceedings) expanded significantly in 2024 in Dakar. This year, the Indaba publications committee received 175 submissions from 34 countries, highlighting both the scale and diversity of AI research emerging from the continent. Accepted peer-reviewed papers will be published in a special volume of the IJCAI proceedings.  UNU Macau has been proud to help convene and support this progress. Dr. Georgina Curto, PhD, MBA, Senior AI Researcher and Team Lead at UNU Macau, chaired the first symposium in South Africa and co chaired the second edition in Dakar. Her role was instrumental in ensuring that African perspectives are meaningfully connected to global AI dialogues. We remain committed to building bridges that turn local knowledge into global impact. 📕 Read the 2024 proceedings: https://lnkd.in/gkrPWYTC

I am excited to share our latest publication in Marine Mammal Science: Adaptive Acoustic Monitoring for Endangered Cook Inlet Beluga Whales in Complex Soundscapes Monitoring endangered species at scale is challenging, especially when they are rare, inhabit dynamic environments, and are recorded using different technologies over decades. For the endangered Cook Inlet beluga whale population, passive acoustic monitoring has become a critical tool for understanding habitat use and informing conservation decisions, but analyzing these large and evolving datasets remains a significant challenge. In this study, we developed an adaptive, open-source AI framework that combines: 🔹 A two-stage deep learning architecture that separates cetacean signal detection from species classification 🔹 Multi-species classification of beluga, killer whale, and humpback whale vocalizations 🔹 Contrastive audio-language models (CLAP) to efficiently expand annotations for underrepresented species 🔹 Active learning workflow that enables rapid adaptation to new soundscapes, recording systems, and environmental conditions One of the most rewarding outcomes was seeing how these methods improved the detection of rare species occurrences across critical winter habitats that have historically been difficult to monitor. Beyond Cook Inlet, the framework was designed to be transferable to other long-term passive acoustic monitoring programs facing domain shifts, changing hardware, and expanding spatial coverage. This work is a great example of how advances in AI can support real-world conservation challenges while remaining grounded in ecological understanding and expert validation. Many thanks to Manuel Castellote, Rahul Dodhia, Zhongqi Miao, Pablo Arbeláez, Verena Gill, Lori Polasek, Juan M. Lavista Ferres, and all collaborators from Microsoft AI for Good, NOAA Fisheries, the Alaska Department of Fish & Game, and Universidad de los Andes who made this interdisciplinary effort possible. 📖 Read the paper here: https://lnkd.in/ewitZtnB 🐋 Learn more about Cook Inlet beluga whale research and conservation: https://lnkd.in/eFzcjjFW #AIForGood #MarineMammalScience #Bioacoustics #MachineLearning #DeepLearning #ConservationTechnology #PassiveAcousticMonitoring #BelugaWhales #MarineConservation #WildlifeMonitoring #NOAA #ConservationAI

I am excited to share our latest paper, now available on arXiv: Overhead Wildlife Locator (OWL): Benchmarking Weakly Supervised Learning for Aerial Wildlife Surveys Counting wildlife from aerial imagery is one of the most labor-intensive steps in modern conservation. Standard object detectors need bounding-box annotations, which are reported to be up to 7x slower and 3x more expensive to produce than simple point labels. That annotation bottleneck slows down the surveys that wildlife managers depend on. To address this, we built our MegaDetector for overhead animal localization. The framework is called OWL: a weakly supervised density-estimation approach that learns to locate and count animals from point annotations alone. It comes in three model variants, each suited to a different survey regime: ⚡ OWL-C - a fully convolutional model for fast, high-throughput screening 🧩 OWL-T - a Swin-augmented hybrid for cluttered, heterogeneous scenes 🧠 OWL-D - built on a frozen DINOv3 foundation encoder with a DPT-style fusion decoder We benchmarked all three against strong detection baselines (POLO, YOLOv11, and HerdNet) across five public aerial datasets, from sparse fixed-wing savanna surveys to dense UAV paddock imagery. OWL-D set a new state of the art on the Delplanque benchmark (0.934 AP vs. HerdNet's 0.840) and achieved the highest detection accuracy (AP) on four of the five datasets. The most rewarding outcome was seeing OWL hold up in a real deployment. Working with the Alaska Department of Fish and Game, we applied OWL-C to the 2022 Central Arctic caribou census under cross-herd and cross-temporal transfer (trained on 2017 Porcupine Caribou Herd imagery) and reached F1 = 0.965 with a +3.1% population-count error, across nearly 15 gigapixels of survey imagery. To support the community, we're releasing the code, model weights, and annotated patch-level datasets for large-scale caribou aerial surveys (Porcupine Caribou Herd 2017 and Central Arctic Herd 2022). We hope OWL helps other monitoring programs facing the same annotation and scaling challenges. Huge thanks to my co-authors Zhongqi Miao, Bruno Demuro, Caleb Robinson, Rahul Dodhia, Lasha O., Jason Holmberg, Kirk Larsen, Howard Frederick, Nathan J. Pamperin, Pablo Arbeláez, and Juan M. Lavista Ferres. This was a deeply collaborative effort spanning the Microsoft AI for Good Lab, Microsoft AI for Good Lab - Biodiversity, Universidad de los Andes - Colombia (Cinfonia), the Alaska Department of Fish and Game, Conservation X Labs, and the Tanzania Wildlife Research Institute. 📄 Read the paper: https://lnkd.in/eWCR48dF 💻 Code, model weights, and datasets: https://lnkd.in/eT-Pb7Vu #AIForGood #MegaDetector #Overhead #RemoteSensing #WildlifeMonitoring #MachineLearning #DeepLearning #ComputerVision #ConservationTechnology #AerialSurveys #FoundationModels #Caribou #ConservationAI #Photogrammetry

Taylor Geospatial and partners including NASA Harvest have released the first global agricultural field boundary map at 10-meter resolution, covering more than 3 billion field units across 241 countries and territories. Field boundaries are a critical unit for agricultural analysis. They help researchers estimate yields, monitor water use, assess risk, and compare farming systems. Yet in many parts of the world, reliable field boundary data has been unavailable or difficult to obtain. Built from Sentinel-2 imagery, the Fields of The World (FTW) release provides a new global field boundary dataset along with tools that allow users to generate field boundaries from recent satellite imagery through the FTW Explorer App, API, and QGIS plugin. The result is a new resource for field-scale agricultural analysis, particularly in regions where reliable agricultural maps previously did not exist. Learn more about Fields of the World! 👉 https://lnkd.in/eGVmbX-h