The 2010 CDL was released in January 2011 @ 30 meters resolution.No farmer reported data is revealed, nor can it be derived in the publicly releasable Cropland Data Layer product.
The CDL program began in earnest in 1997 with the ability to deliver geospatial content annually to customers who were interested in annual crop land cover updates. Prior to the creation of the CDL product, estimates were provided in tabular format, with pictures/outputs depicting the results.The CDL can be considered a “Census by Satellite”, as it is a comprehensive land use classification covering an entire state, and uses ortho-rectified imagery, to accurately locate and identify field crops.The CDL utilizes a comprehensive and robust archive of AWiFS satellite imagery from the Foreign Ag Service, Landsat TM provided by the U.S. Geological Survey, MODIS dataThe following ground truth is held as confidential: The NASS June Ag Survey and the Farm Service Agency/Common Land Unit. This data is not provided or shared with anyone.The Cropland Data Layer (CDL) is now operational providing in-season estimates for decision support in our NASS Field Offices and Agricultural Statistics Board. Estimates are delivered multiple times during the growing season, helping improve agency estimates. The CDL program strives to cover all NASS speculative program crops of Corn, Soybeans, Wheat, and Cotton in crop year 2009, providing improved acreage estimates throughout the growing season as more farmer reported and satellite data are utilized. Crop year 2009 marked a first with a national release of the 48 conterminous states. Coverage of additional 21 states was funded by the US EPA.The CDL is a publically releasable crop specific land cover classification that focuses primarily on mapping cultivated fields and providing an update on the agriculture landscape. The 2010 CDL product was released in January 2011, coincident with the release of the new NASS CropScape Geo Portal @ nassgeodata.gmu.edu/CropScape. Users can now interact with the data online.
To appreciate the CDL, it helps to zoom in to the field level. Here is Craighead County Arkansas. In this map you can see the rice and soybeans on the left side and the corn and cotton on the right.
These are the major inputs into the production of the CDL Program.Medium resolution satellites like Landsat and AWiFS are used as a primary input.The Farm Service Agency – Common Land Unit depicts planted fields that are proportionately sampled as input to the classifier.The National Land Cover Dataset along with the National Elevation, Percent Forest Canopy, Percent Imperviousness are ancillary input layers sampled over the non-ag domain.The NASS June Ag Survey are enumerated segments that are utilized to build acreage estimates using statistical modeling.
The first critical component of any remote sensing program is ground truth. NASS is very fortunate to have available administrative data from USDA FSA for agriculture ground truth. Farmers who participate in government programs and crop insurance report their crops to FSA. And starting in 2006 this data is available by geo-location. Non-ag land ground truth comes from USGS National Land Cover Dataset. This identifies, the cities, forests, water and other non-ag land.Agricultural categoriesUse Farm Service Agency (FSA) farmer reported “578” program crop tied to Common Land Unit (CLU) polygon dataEarly in the season this information is thin< 10% representation of cropsMore complete as season evolves20, 30, 40%, or more coverage….Non-agricultural categoriese.g. water, woodlands, grasslands, developed…Ignore the cropland categorized areas of the National Land Cover Data Set (NLCD)Not perfect but needed because not everything is agriculture across the landscape
U.S. National Agricultural Land Cover Monitoring Rick Mueller USDA/National Agricultural Statistics Service IEEE IGARSS July 28, 2011
14 states – winter wheat 14 – corn & soybeans 15 – rice, cotton & peanuts 17 – all small grains 24 – all crops 27 – operational Cropland Data Layer 2010 in-season production @ 56m 48 states post season @ 30m
2010 Cropland Data Layer Inputs Satellite Imagery - AWiFS & Landsat TM USDA Farm Service Agency/Common Land Unit NLCD & Derivative products NASS June Agricultural Survey
Ground Truth – Land Cover Agriculture Ground Truth Non-Agriculture Ground Truth Provided by Farm Service Agency Identifies known fields and crops Divide known fields into 2 sets 70% used for training software 30% used for validating results U.S. Geological Survey National Land Cover Dataset Identifies urban infrastructure and non-agriculture land cover Forest, grass, water, cities
Satellite Data with Farm Service Agency Common Land Unit (CLU) Polygons
Satellite Data with Farm Service Agency CLUs Overlay Corn - Soybean - Winter Wheat - Alfalfa -
Kansas 2010 CDL Input Layers AWiFS AWiFS AWiFS TM TM TM TM DEM Canopy Impervious MODIS Scenes of data actually used: 24 AWiFS, 13 Landsat TM, 2 MODIS NDVI, DEM, Canopy, and Impervious
Classification Methodology Overview “Stack” Landsat, Landsat-like data, and ancillary data layers within a raster GIS Sample spatially through stack from known ground truth from FSA (ag. categories) and NLCD (non-ag. categories) “Data-mine” those samples using Boosted Classification Tree Analysis to derive best fitting decision rules Apply derived decision rules back to entire input data stack to create full scene classification Agricultural ground truth (via the USDA Farm Service Agency) Non-agricultural ground truth (using the National Land Cover Dataset as a proxy) (dependent data) Manages and visualizes datasets Output “Cropland Data Layer” Rulequest See5.0 Derives decision tree-based classification rules Imagery stack (independent data) Generated rule set
CDL Accuracy Assessment Each classification tested against independent set of ground truth data to determine overall and within class accuracies Example classification subset Example validation subset
Accuracy Statistics Producer’s Accuracy: relates to the probability that a ground truth pixel will be correctly mapped and measures errors of omission. Errors of Omission: occur when a pixel is excluded from the correct category User’s Accuracy: indicates the probability that a pixel from the classification actually matches the ground truth data and measures errors of commission Errors of Commission: occur when a pixel is included in an incorrect category
Regression-based Acreage Estimator Acreage not just about counting pixels Simple Linear Regression Regression used to relate categorized pixel counts to the ground reference data (X) – Cropland Data Layer (CDL) classified acres (Y) – June Agricultural Survey (JAS) reported acres Outlier segment detection - removal from regression analysis Using regression results in estimates reduces error rates over using JAS alone Estimate 17 crops in 39 states
Washington - June Winter Wheat – Stratum 11 R² Jun – 0.769
Washington - September Winter Wheat – Stratum 11 R² Jun – 0.769 Sep – 0.883
Washington - October Winter Wheat – Stratum 11 R² Jun – 0.769 Sep – 0.883 Oct – 0.890
Cropland Data Layer Summary Operational Program Timely estimate delivery Measureable statistical error Set national/regional/county acreage estimates Components AWiFS/Landsat imagery Farm Service Agency/Common Land Unit USGS NLCD/ancillary layers June Agricultural Survey Leverage CDL program paramount to other NASS geospatial activities Partnerships with cooperating agencies critical for success Heavy reliance on satellites and information technology Distribution CropScape Portal NRCS Data Gateway