Applying Google Earth Engine to Wildfire Disturbance Detection in the State of Alaska

Size: px

Start display at page:

Download "Applying Google Earth Engine to Wildfire Disturbance Detection in the State of Alaska"

Geoffrey Poole
5 years ago
Views:

University of Tennessee Knoxville, and 3 USDA

1 Applying Google Earth Engine to Wildfire Disturbance Detection in the State of Alaska Forrest M. Hoffman 1,2, Zachary L. Langford 2, Jitendra Kumar 1,2, Steve Norman 3, and William W. Hargrove 3 1 Oak Ridge National Laboratory, 2 University of Tennessee Knoxville, and 3 USDA Forest Service Eastern Forest Environmental Threat Assessment Center 2018 US-IALE Annual Meeting Chicago, Illinois, USA April 11, 2018

Rationale To use remote sensing and meteorology to estimate wildfire areal extents for the 2004 wildfire season in Alaska To test the utility of Google Earth Engine (GEE) to determine the most

2 Rationale To use remote sensing and meteorology to estimate wildfire areal extents for the 2004 wildfire season in Alaska To test the utility of Google Earth Engine (GEE) to determine the most important attributes from MODIS and Daymet Future: Monitor: Locate fires under 1,000 acres (below cutoff for Monitoring Trends in Burn Severity (MTBS) dataset) Predict: Determine preceding meteorological conditions that result in fire susceptibility Area (km 2 ) Numerous forest fires (outlined in red) were burning in the Yukon Flats region of east-central Alaska in mid-june 2004 captured by MODIS sensor. (Image Source: NASA) Monitoring Trends in Burn Severity (MTBS) dataset for Alaska from , with 2004 having the most area (km 2 ) burned.

Study Region: Interior Alaska I Study region is Interior Alaska (Bieniek et al., 2012) I Wildfires are natural and may be increasing in intensity due to climate change (e.g., length of the growing season has increased 45% over the last century (Chapin III et al.

3 Study Region: Interior Alaska I Study region is Interior Alaska (Bieniek et al., 2012) I Wildfires are natural and may be increasing in intensity due to climate change (e.g., length of the growing season has increased 45% over the last century (Chapin III et al., 2014)). Legend I I Monitoring Trends in Burn Severity (MTBS) dataset was used for training classifier with MODIS and Daymet variables. Binary classification: wildfire (1) no wildfire (0) Wildﬁre Extent Land Cover Open Water Perennial Ice/Snow Developed, Open Space Developed, Low Intensity Developed, Medium Intensity Developed, High Intensity Barren Land (Rock/Sand/Clay) Unconsolidated Shore Deciduous Forest Evergreen Forest Mixed Forest Dwarf Scrub Shrub/Scrub Grassland/Herbaceous Sedge/Herbaceous Lichens Moss Pasture/Hay Cultivated Crops Woody Wetlands Emergent Herbaceous Wetlands km

2004 Wildfire Season in Alaska I One of the warmest and driest summers on record.

Historical Lightning Departure from average temperature across Alaska for every year since 1949.

This map is a user generated static output from an Internet mapping site and is for general reference.

4 2004 Wildfire Season in Alaska I One of the warmest and driest summers on record. I Most lightning strikes recorded during summer. I Wildland fires burned the largest area in recorded Alaska history. Historical Lightning Departure from average temperature across Alaska for every year since (Image Source: Alaska Climate Research Center) mi. This map is a user generated static output from an Internet mapping site and is for general reference. Data layers that appear on this map are obtained from many sources. This map is not to be used for navigation. Number of lightnings strikes (6,538) in Alaska from June 5 19, The grand total was over 147,642 strikes (Chapin et al., 2008). Apr on th

Emissivity 8-Day L3 Global 1km) (Vermote, 2015; Wan et al., 2015). Daymet: gridded estimates of daily weather parameters (Thornton et al.

5 Geospatial Datasets We used Google Earth Engine (GEE) for processing images and building models. Two types of datasets were employed: MODIS: MOD09A1 (Surface Reflectance 8-Day L3 Global 500m) and MOD11A2 (Land Surface Temperature and Emissivity 8-Day L3 Global 1km) (Vermote, 2015; Wan et al., 2015). Daymet: gridded estimates of daily weather parameters (Thornton et al., 2017). Daymet V3 average annual minimum temperature for 1980 and 2015 for a subset of the Daymet domain in Alaska and western Canada. MOD09A1 RGB composite from June 17, 2004.

6 Geospatial Datasets Description Resolution Variable GMTED m elevation (m) 225 m slope (%) MOD09A1 500 m at 8 days NDVI 500 m at 8 days EVI 500 m at 8 days SAVI 500 m at 8 days Bands 1 7 ( nm) MOD11A2 1 km at 8 days Daytime LST (Kelvin) Daymet 1 km at daily Daylight period (seconds) 1 km at daily Precipitation (mm) 1 km at daily Snow water equivalent (kg/m 2 ) 1 km at daily Maximum temperature ( C) 1 km at daily Minimum temperature ( C) 1 km at daily Shortwave Radiation (W/m 2 ) 1 km at daily Vapor Pressure (Pa) Daymet and MODIS products were processed from early-april through late-october in 2004.

7 Data Workflow I Perform image processing methods for classification. I Build models with MODIS (288), MODIS/Daymet (456), and Daymet (168) variables and the MTBS dataset. I Right Figure: Google Earth Engine interactive development environment (Gorelick et al., 2017).

8 Image Processing I Increased resolution to 500 m for all datasets, GEE performs nearest neighbor resampling. I Linear interpolation for missing values. I Savitzky-Golay filter was applied to smooth out noise (Chen et al., 2004). I Daymet variables were merged into 8-day averages Alaska: BOUNDARY ak Original Interpolation Savitzky-Golay NDVI ! 0.1 This map portrays fire severity for the fire specified in the map title and summarizes proportions of fire severity classes. These data are produced under the Monitoring Trends in Burn Severity (MTBS) project jointly implemented by the USGS EROS and the USFS RSAC. The MTBS project ascertains the locations of fires based on available fire occurrence information provided by federal and state agencies, and other reliable sources. The MTBS project reserves the right to correct, update or modify geospatial inputs to this map without notification Latitude: 65o 15' 46.8" Longitude: -146 o 56' 34.8" Fire Ignition Date: June 13, 2004 Assessment Type: Extended Pre-Fire Image Date: May 27, 2002 (Landsat 7) Post-Fire Image Date: July 22, 2005 (Landsat 7) May 2004 Jun 2004 Jul 2004 Aug 2004 Sep 2004 Oct 2004 Example image processing workflow applied to a large wildfire, which occurred on June 13, * Areas in either the pre-fire or post-fire reflectance imagery containing clouds, snow, shadows, smoke, significantly sized water bodies, missing lines of image data, etc Miles Acreage of Burn Severity Burn Severity Unburned to Low Low Moderate High Increased Greenness Non-Processing Area Mask* Total Acres 41,952 83, , ,133 19, , ,166 Fire severity for the Boundary fire based on Landsat 7. (Source: USGS and US Forest Service)

9 Random Forest Random Forest: estimator that fits a number of decision tree classifiers on various sub-samples of the dataset and use averaging to improve the predictive accuracy and control over-fitting. We split up dataset based on pixel tiles, with 163 tiles used for training and 73 tiles used for testing. Legend Wildfire Extent Test Tiles Train Tiles 100 x 100 Pixel Tiles Train = 163 Tiles Test = 73 Tiles km

Results: MODIS Validation Metrics Classes 0 1 precision Classification report recall f1-score 1.000 0.975 0.950 0.925 0.900 0.875 0.850 0.825 0.

10 Results: MODIS Validation Metrics Classes 0 1 precision Classification report recall f1-score Precision, recall, and F1-score for non-wildfire class (n=493710) was 0.99, 1.00, and 1.00, respectively. Precision, recall, and F1-score for wildfire class (n=16878) was 0.93, 0.78, and 0.85, respectively. Feature Importance Blue EVI Green LST NDVI NIR Red type SAVI SWIR1 SWIR2 SWIR3 Importance EVI, NDVI, and MODIS Red band contributed the most using the Gini feature importance metric. Elevation, slope, SAVI, and SWIR1 contributed the least. ele slope

11 Results: Daymet Validation Metrics Feature Importance Classes 0 1 precision Classification report recall f1-score Precision, recall, and F1-score for non-wildfire class (n=493710) was 0.99, 1.00, and 0.99, respectively. Precision, recall, and F1-score for wildfire class (n=16878) was 0.62, 0.53, and 0.57, respectively Importance dayl prcp srad swe type Most variables contributed equally for feature importance. Snow water equivalent, minimum temperature, and precipitation were the highest scoring features. tmax tmin vp

Results: MODIS/Daymet Validation Metrics Classes 0 1 precision Classification report recall f1-score 1.000 0.975 0.950 0.925 0.900 0.875 0.850 0.825 0.

12 Results: MODIS/Daymet Validation Metrics Classes 0 1 precision Classification report recall f1-score Precision, recall, and F1-score for non-wildfire class (n=493710) was 0.99, 1.00, and 1.00, respectively. Precision, recall, and F1-score for wildfire class (n=16878) was 0.93, 0.78, and 0.85, respectively. Feature Importance Blue EVI Green LST NDVI NIR Red SAVI SWIR1 SWIR2 SWIR3 dayl ele prcp slope srad swe tmax tmin type Importance Daymet variables (minimum temperature, maximum temperature, shortwave radiation) contributed most for Daymet/MODIS classification. MODIS Green, Red, and SAVI variables contributed the most. vp

Results: Test (Bonanza Creek Wildfire) 2004 Alaska: BONANZA CREEK ak6457815738420040628 MODIS Latitude: 64o 34' 40.8" Longitude: -157 o 23' 02.

This map portrays fire severity for the fire specified in the map title and summarizes proportions of fire severity classes.

The MTBS project ascertains the locations of fires based on available fire occurrence information provided by federal and state agencies, and other reliable sources.

13 Results: Test (Bonanza Creek Wildfire) 2004 Alaska: BONANZA CREEK ak MODIS Latitude: 64o 34' 40.8" Longitude: -157 o 23' 02.4" Fire Ignition Date: June 28, 2004 Assessment Type: Extended Pre-Fire Image Date: July 31, 2002 (Landsat 7) Post-Fire Image Date: June 21, 2005 (Landsat 7)! This map portrays fire severity for the fire specified in the map title and summarizes proportions of fire severity classes. These data are produced under the Monitoring Trends in Burn Severity (MTBS) project jointly implemented by the USGS EROS and the USFS RSAC. The MTBS project ascertains the locations of fires based on available fire occurrence information provided by federal and state agencies, and other reliable sources. The MTBS project reserves the right to correct, update or modify geospatial inputs to this map without notification. * Areas in either the pre-fire or post-fire reflectance imagery containing clouds, snow, shadows, smoke, significantly sized water bodies, missing lines of image data, etc Miles Acreage of Burn Severity Burn Severity Unburned to Low Low Moderate High Increased Greenness Non-Processing Area Mask* Total Acres 33,879 60,676 74,976 10,859 5,231 56, ,845 Daymet Fire severity for the Bonanza Creek wildfire based on Landsat 7. (Source: USGS and US Forest Service) MODIS/Daymet

14 Conclusions MODIS bands and vegetation indices can be used to predict spatial extents of wildfire with good accuracy, and including Daymet does not improve the predictions. MODIS (NIR, SWIR), indices (EVI, NDVI, SAVI), and Daymet variables (minimum temperature, maximum temperature, snow water equivalent, shortwave radiation) are the most important factors determining wildfire extent. Random Forest provides a good approach for determining feature importance. Google Earth Engine provides a powerful platform for processing and analyzing datasets without moving data. Future Work: Would even (fire/no fire) sampling for training provide more balanced prediction accuracy? Can the method be used to predict fire extent across multiple years? Can we use antecedent meteorological conditions (for 3 months or 1 year anomalies) to predict fire susceptibility?

Acknowledgements This research was sponsored by the Climate and Environmental Sciences Division (CESD) of the Biological and Environmental Research (BER) Program in the US Department of Energy Office

15 Acknowledgements This research was sponsored by the Climate and Environmental Sciences Division (CESD) of the Biological and Environmental Research (BER) Program in the US Department of Energy Office of Science and the US Department of Agriculture Forest Service, and the National Science Foundation (AGS ). This research used resources of the Oak Ridge Leadership Computing Facility (OLCF) at Oak Ridge National Laboratory (ORNL), which is managed by UT-Battelle, LLC, for the US Department of Energy under Contract No. DE-AC05-00OR22725.

16 References I P. A. Bieniek, U. S. Bhatt, R. L. Thoman, H. Angeloff, J. Partain, J. Papineau, F. Fritsch, E. Holloway, J. E. Walsh, C. Daly, M. Shulski, G. Hufford, D. F. Hill, S. Calos, and R. Gens. Climate divisions for Alaska based on objective methods. J. Appl. Meteor. Climatol., 51(7): , doi: /jamc-d F. S. Chapin, S. F. Trainor, O. Huntington, A. L. Lovecraft, E. Zavaleta, D. C. Natcher, A. D. McGuire, J. L. Nelson, L. Ray, M. Calef, N. Fresco, H. Huntington, T. S. Rupp, L. DeWilde, and R. L. Naylor. Increasing wildfire in Alaska s boreal forest: Pathways to potential solutions of a wicked problem. BioSci., 58 (6): , doi: /b

17 References II F. S. Chapin III, S. F. Trainor, P. Cochran, H. Huntington, C. J. Markon, M. McCammon, A. D. McGuire, and M. Serreze. Alaska. Technical report, U.S. Geological Survey, Washington, D.C., URL J. Chen, P. Jönsson, M. Tamura, Z. Gu, B. Matsushita, and L. Eklundh. A simple method for reconstructing a high-quality NDVI time-series data set based on the Savitzky-Golay filter. Remote Sens. Environ., 91(3): , ISSN doi: /j.rse N. Gorelick, M. Hancher, M. Dixon, S. Ilyushchenko, D. Thau, and R. Moore. Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sens. Environ., 202:18 27, ISSN doi: /j.rse Big Remotely Sensed Data: tools, applications and experiences.

18 References III P. Thornton, M. Thornton, B. Mayer, Y. Wei, R. Devarakonda, R. Vose, and R. Cook. Daymet: Daily surface weather data on a 1-km grid for North America, version 3, E. Vermote. MOD09A1 MODIS/Terra Surface Reflectance 8-Day L3 Global 500m SIN Grid V006, Z. Wan, S. Hook, and G. Hulley. MOD11A2 MODIS/Terra Land Surface Temperature/Emissivity 8-Day L3 Global 1km SIN Grid V006, 2015.

Utility of National Spatial Data for Conservation Design Projects

Utility of National Spatial Data for Conservation Design Projects Steve Williams Biodiversity and Spatial Information Center North Carolina State University PIF CDW St. Louis, MO April 11, 2006 Types of