Luca Sartore- USDA NASS
Title: Crop classification and uncertainty assessment at 10 m resolution using Google Earth Engine
Title: Crop classification and uncertainty assessment at 10 m resolution using Google Earth Engine
Abstract:
Abstract:
NASS produces the Cropland Data Layer (CDL) using remote sensing data to provide an independent source of information for its acreage estimation program. Currently, moderate resolution satellite data (30 m) are acquired throughout the entire summer growing season. These data are then used at the end of the growing season to produce the CDL as higher resolution data pose several computational challenges. Google Earth Engine (GEE) is a web-platform that provides a comprehensive collection of publicly available remote sensing datasets for scientific analysis and visualization. This presentation discusses the challenges and the methodology used to produce early season crop classifications and related uncertainties. JavaScript programs are developed interactively in GEE allowing more programming flexibility and capabilities, particularly in producing produce suitable results where computational resources are limited for very large areas. Examples produced at 10 m resolution, using only images acquired between April and June in the state of Illinois, are evaluated with ground reference data based on Farm Service Agency (FSA). The overall accuracy of early season crop classification at 10 m resolution from 2017 to 2019 varies between 79% and 84%. Useful relationships between local uncertainties (based on the entropy of the predictive distributions) and state-level accuracies (based on classification errors) can provide an additional source of information in modelling acreage.
NASS produces the Cropland Data Layer (CDL) using remote sensing data to provide an independent source of information for its acreage estimation program. Currently, moderate resolution satellite data (30 m) are acquired throughout the entire summer growing season. These data are then used at the end of the growing season to produce the CDL as higher resolution data pose several computational challenges. Google Earth Engine (GEE) is a web-platform that provides a comprehensive collection of publicly available remote sensing datasets for scientific analysis and visualization. This presentation discusses the challenges and the methodology used to produce early season crop classifications and related uncertainties. JavaScript programs are developed interactively in GEE allowing more programming flexibility and capabilities, particularly in producing produce suitable results where computational resources are limited for very large areas. Examples produced at 10 m resolution, using only images acquired between April and June in the state of Illinois, are evaluated with ground reference data based on Farm Service Agency (FSA). The overall accuracy of early season crop classification at 10 m resolution from 2017 to 2019 varies between 79% and 84%. Useful relationships between local uncertainties (based on the entropy of the predictive distributions) and state-level accuracies (based on classification errors) can provide an additional source of information in modelling acreage.