A COMPUTER VISION AND MACHINE LEARNING SYSTEM FOR BIRD AND BAT DETECTION AND FORECASTING

Transcription

1 A COMPUTER VISION AND MACHINE LEARNING SYSTEM FOR BIRD AND BAT DETECTION AND FORECASTING Russell Conard Wind Wildlife Research Meeting X December 2-5, 2014 Broomfield, CO

2 INTRODUCTION Presenting for Engagement Project Is Ongoing and Results Are Preliminary Current Research Is Highly Collaborative with Government and Private Sector Partners Focused Foremost on the Science December 1,

3 OBJECTIVES Applying Computer Vision to Remote Wildlife Sensing Identify Birds by Species in Video Remotely Operate Computer Vision Sensor Networks Build Predictive Models for Avian and Bat Behavior in High-Dimensional Spaces Multispectral Vision Data Meteorological Data Bat Acoustic Detections Deploy Marinised System Offshore December 1,

4 TIMELINE Indiana University Computer Science Research Introduced to Wind Energy by Anna Peterson, PhD (Also Presented at NWCC Research Meeting X) Field Collected Training Data on Raptors 2012 Proved Feasibility in the Laboratory and Field Commercialization in Wind Energy Unsuccessful DoD Environmental Programs Adversely Impacted by Sequestration December 1,

5 TIMELINE Future 2014 Won Department of Energy Office of Wind and Water Power Technologies Small Business Innovation Research Grant Began First Field Deployment at NRG Yield s ALTA Wind Energy Center Through a Partnership with Terra-Gen Power Ongoing Deliver Field Results and Forecasting Develop Marinised System for Offshore Deployment Deploy Automated Onshore and Offshore Wildlife Tracking System December 1,

6 CONTEXT Soaring Innovation, Grounded in Science Species Identification Matters Some Species Are More Equal Than Others Deterrence Has a Species Component Curtailment Is too Expensive to Perform Indiscriminately Eagle Enforcement Is Driving New Technology Projects of Increasing Scale Difficult to Survey Extensively Offshore Data Collection Is Tremendously Difficult and Expensive December 1,

7 TECHNOLOGICAL SCOPE Focus on the Software Hardware and Data Collection Machine Vision Cameras Data Capture Bat Acoustics Computing Servers GigE Vision Protocol Ornicept Software Meteorological Sensors Housings Bat Acoustic Recognition Bird Identification Predictive Models Risk Classification December 1,

8 OVERVIEW OF THE APPROACH SOAR: Specific Optical Avian Recognition December 1,

9 TRAINING DATA December 1,

10 SOAR METHOD Fine-Grain Object Recognition Motion Detection Eliminate Background Suppress Clouds, Wind Turbines, and Insects Target Extraction Detect Bird Blobs Extract Aligned Birds Species Recognition Novel Feature Extraction Methods Novel Machine Learning Algorithms for Recognition Post-Processing Calculate Altitude, Velocity, and Trajectory Export Human Auditable Dataset December 1,

11 SPECIES RECOGNITION RESULTS Preliminary Bald Eagle Model Results December 1,

12 RAPTOR BEHAVIOR PREDICTION A Systems Approach Incorporates Meteorological Data and Species Data Predictive Classification Based on Support Vector Machines Current Model Based on Migration Data Training Data Obtained Through Partnerships US Fish & Wildlife Service Holiday Beach Bird Observatory Hawk Migration Association of North America Hawk Mountain Sanctuary Future Model Based on Field Collected Data December 1,

13 MIGRATION FORECAST Sharp-shinned Hawk (Accipiter striatus) Accipiter Example Raw Data Shown Data Is Binned for Interpretation Based on Population Thresholds Yields More Accurate Forecasts December 1,

14 MIGRATION FORECAST Red-tailed Hawk (Buteo jamaicensis) Buteo Example December 1,

15 FORECASTING IMPLICATIONS A Foundational Model Outperforms Commonly Accepted Models (e.g. Hussell & Farmer; Bird Studies Canada) for Modeling Raptor Migration Currently Trained for Migrating Raptors Given Sufficient Data, Could Be Trained for Other Species Given Sufficient Data, Could Account for Behaviors Adapting Model to Micro-Forecasting Within Regions and Sites December 1,

16 CURRENT RESEARCH Multiple Experiments One System Fully Operational Research Platform at the ALTA Wind Energy Center Multispectral Recognition Recognition Algorithms Development Visible, Near Infrared, and Thermal Long-Range Detection Ultra High Resolution Recognition Advanced Background Modeling and Target Differentiation Sensor Integration On-site Meteorological Data Collection Concurrent Acoustic Data Collection Site Modeling Incorporates Data from All Sensors Models Wildlife Use Patterns Geospatially and Temporally December 1,

17 LONG-TERM IMPLICATIONS Actionable Data from Software Wind Project Siting Automated Avian Point Counts Risk Modeling Based on Site Parameters Threat Detection Species Specific Threat Classification Automated and Auditable Reduced Curtailment Costs Species Data for Decision-Making Forecast Curtailment for Birds and Bats in Advance December 1,

18 PARTNERS AND SUPPORT Funding Partner Research Partner Expertise and Technology Partners Contributors December 1,