Downloaded 30 times
More Related Content
Similar to AI for Earth: Analyzing Global Data with Azure
More from Microsoft Tech Community
AI for Earth: Analyzing Global Data with Azure
- 2.
- 4. AI is agame changer for sustainability. AI is changing many facets of life; it’s the next industrial revolution. AI innovations amplify human ingenuity and can increase the quantity and quality of data we gather about the Earth—and do it exponentially faster.
- 6. AGRICULTURE In order tofeed the world’s rapidly growing population, farmers must produce more food, on less arable land, and with lower environmental impact. WATER In less than two decades, demand for fresh water (for human consumption, agriculture and hygiene) is projected to dramatically outpace supply. BIODIVERSITY Species are going extinct beyond the natural rate by orders of magnitude, driving the decay of key ecosystem services, like pollination, that humans depend upon. CLIMATE CHANGE An increasingly variable climate, extreme weather events, rising sea levels, higher global temperatures, and increased ocean acidity threaten human health, infrastructure, and the natural systems we rely on for life itself. AI for Earth http://www.microsoft.com/AIforEarth
- 7. Commitments Increase Access We areproviding seed grants so researchers and organizations can gain access to cloud and AI computing resources. Fuel Innovation We are accelerating the pace of innovation by showcasing lighthouse projects, publishing research, and collaborating with others to expand and grow initial projects. Provide Education We are providing educational resources to make sure organizations know what is available, how to use it, and how it can meet their specific needs.
- 9. Challenge Creating high-quality landcover maps with today’s high-resolution imagery is a resource-intensive process. It can take up to a year to become available to the public, and is time-consuming to sort, manage, and classify images into land cover data. Solution Microsoft’s goal is to help make the land cover mapping process faster, easier, and more accessible to everyone. We want to enable on-the-fly land cover mapping, and allow conservation groups to create high-resolution land cover maps with their own data; develop accurate, actionable insights more quickly; and easily transfer maps and insights to other organizations. Land Cover Mapping Giving organizations a faster, more effective, and lower cost land cover mapping tool to help them better analyze, monitor, and manage natural resources.
- 12. manually created ahigh-resolution land cover map for precision conservation of the Chesapeake watershed 100k mi2 Area of watershed to map 2TB File size of imagery to classify 18 months Time to create map By the time the land cover map was completed in December 2016, it was already out of date, and an update would be time-intensive and costly.
- 13. Land Classification Model Convolutional Network Architecture 23layer U-Net Test ImagesLabeled Training Images Chesapeake Conservancy Dataset Land Classification Model Land Cover Map Working Platform: Geo AI Virtual Machine Dataset: 120k mi2 of imagery at 1-meter resolution, split in half geographically into train and test sets
- 14. AI for Earth Neuralnet implementation U-Net model - 23 layers; 314 373 parameters - implemented in Cognitive Toolkit (CNTK) - trained for 30 hrs on Tesla K80 GPU - Input: 4-channel image patch (typically 256×256) - Output: predicted probabilities of 4 output classes at each pixel (excluding boundary) Dataset: ~120,000 mi2 (3×1011 pixels) of imagery at 1m resolution, split into train & test sets Evaluation takes ~1min for 5×5-mi on single GPU Trained model is packaged and readable by CNTK evaluation APIs (Python, C++, C#, etc.) Training & evaluation easily parallelized across multiple GPUs and machines with Cognitive Toolkit distributed API (training time down to one hour on a cluster of 128 GPUs)
- 15. Land Classification Modelin Action Oakland, Michigan Aerial photo 1m resolution, input data Existing land cover map Created 7 years ago, out of date Land classification model Classifying on the fly, and detects new roundabout Land classification model Show mix of probabilities across land cover types
- 17.
- 20. { "Barren/impervious": 0.026041666666666668, "Herbaceous": 0.3568996853298611, "NoData": 0.0, "Trees": 0.18366156684027779, "Water": 0.43339708116319442 }
- 21.
- 23. Project Premonition Detecting pathogensbefore they cause outbreaks by turning mosquitoes into devices that collect data from animals in the environment. Challenge Nearly 75% of emerging infectious diseases originate from animals, yet it’s difficult to track and monitor these diseases in order to prevent outbreaks. Early data is laborious, and is often collected by hand from potential disease sources in the environment, and sample analyses are time- consuming. Public health organizations need data as early as possible to predict disease spread and plan timely responses. Solution Microsoft and its partners are turning mosquitoes into field biologists by combining smart hardware, machine learning, and advanced cloud-based data analytics. When mosquitoes bite animals, they draw a few microliters of blood containing DNA that identifies the types of animals that were bitten. With advanced machine learning, this can help early detection of vector-borne diseases such as Zika and West Nile.
- 25. Project Premonition smarttraps attract insects into chambers using CO2 and white LED light. Low-cost & power sensors and actuators detect insect presence and species from wingbeat patterns with ~90% accuracy, and can selectively trap species of interest. Traps have been successfully tested in 87 experiments with 22k mosquito landings Preserves each specimen in unique cell for efficient labeling & data integrity Fan-less design allows natural flight & probing behavior to be observed
- 26.
- 27.
- 29. Microsoft Confidential –to be shared under NDA onlyMicrosoft Confidential to be shared under NDA only Azure Maps AT GENERAL AVAILABILITY Time Zones The ability to query for a time zone Note: Additional services will be added to the offering in the future Maps The ability to fetch a visual rendition of map data Routing The ability to calculate a route from point A to B or n points, and receive step by step directions Search and Geocoding The ability to find places, addresses, businesses, POIs etc. Traffic The ability to show dynamic traffic and incident information Map Control A web control mechanism for developers to more easily integrate mapping capabilities into their applications
- 30. Microsoft Confidential –to be shared under NDA only choose Azure Maps? WHY In-vehicle use licensing rights End-to-end location aware solutions Functionality in over 30 languages Commercial routing Best in class traffic data Build on and integrated into Azure! Powered by leading map supplier TomTom Key reasons for customers to opt for Azure Maps for their geospatial needs Learn more at azure.com/maps
- 32. A Z UR E M A C H I N E L E A R N I N G A C C E L E R A T E D B Y P R O J E C T B R A I N W A V E Real-time AI at cloud scale with industry-leading performance and lowest cost Models are easy to create and deploy into Azure Write once, deploy anywhere – to intelligent cloud or edge http://aka.ms/aml-real-time-ai Record-setting DNN performance with accelerated ResNet50 • Record speed: Object classification on FPGA in <1.8 ms per image • Lowest cost: Only 21 cents per million images during preview More accelerated models coming soon
- 33. Data Build TrainDeploy Intelligent Apps AC TIONINTELLIGENC EDATA Stored on Azure Premium Storage Azure Machine Learning 10 01 Using Project Brainwave for Land Use Mapping Land Classification Model ResNet-50 NAIP Data 20TB, 200M images Visual Studio Tools for AI Geo AI Data Science Virtual Machine Azure Batch AI Ultra-fast Inferencing using FPGAs Satellite Images for the Entire US
- 37.
- 43.
- 45.
- 46.
- 47. AI for Earth AIfor EarthLearn more at microsoft.com/aiforearth
- 48.