***NOTE ABOUT THE UNBUFFERED VALIDATION ACCURACY TABLES BEGINNING IN 2016: The training and validation data used to create and accuracy assess the CDL has traditionally been based on ground truth data that is buffered inward 30 meters. This was done 1) because satellite imagery (as well as the polygon reference data) in the past was not georeferenced to the same precision as now (i.e. everything "stacked" less perfectly), 2) to eliminate from training spectrally-mixed pixels at land cover boundaries, and 3) to be spatially conservative during the era when coarser 56 meter AWiFS satellite imagery was incorporated. Ultimately, all of these scenarios created "blurry" edge pixels through the seasonal time series which it was found if ignored from training in the classification helped improve the quality of CDL. However, the accuracy assessment portion of the analysis also used buffered data meaning those same edge pixels were not assessed fully with the rest of the classification. This would be inconsequential if those edge pixels were similar in nature to the rest of the scene but they are not- they tend to be more difficult to classify correctly. Thus, the accuracy assessments as have been presented are inflated somewhat. Beginning with the 2016 CDL season we are creating CDL accuracy assessments using unbuffered validation data. These "unbuffered" accuracy metrics will now reflect the accuracy of field edges which have not been represented previously. Beginning with the 2016 CDLs we published both the traditional "buffered" accuracy metrics and the new "unbuffered" accuracy assessments. The purpose of publishing both versions is to provide a benchmark for users interested in comparing the different validation methods. For the 2019 CDL season we are now only publishing the unbuffered accuracy only publishing the unbuffered accuracy assessments within the official metadata files and offer the full "unbuffered" error matrices for download on the FAQs webpage. Both metadata and FAQs are accessible at <https://www.nass.usda.gov/Research_and_Science/Cropland/SARS1a.php>. We plan to continue producing these unbuffered accuracy assessments for future CDLs. However, there are no plans to create these unbuffered accuracy assessments for past years. It should be noted that accuracy assessment is challenging and the CDL group has always strived to provide robust metrics of usability to the land cover community. This admission of modestly inflated accuracy measures does not render past assessments useless. They were all done consistently so comparison across years and/or states is still valid. Yet, by providing both scenarios for 2016 gives guidance on the bias. If the following table does not display properly, then please visit this internet site <https://www.nass.usda.gov/Research_and_Science/Cropland/metadata/meta.php> to view the original metadata file.
USDA, National Agricultural Statistics Service, 2019 North Dakota Cropland Data Layer
STATEWIDE AGRICULTURAL ACCURACY REPORT
Crop-specific covers only *Correct Accuracy Error Kappa
------------------------- ------- -------- ------ -----
OVERALL ACCURACY** 532,180 80.1% 19.9% 0.764
Cover Attribute *Correct Producer's Omission User's Commission Cond'l
Type Code Pixels Accuracy Error Kappa Accuracy Error Kappa
---- ---- ------ -------- ----- ----- -------- ----- -----
Corn 1 69,156 88.7% 11.3% 0.877 92.0% 8.0% 0.913
Sorghum 4 370 30.4% 69.6% 0.303 52.5% 47.5% 0.524
Soybeans 5 122,748 92.1% 7.9% 0.908 90.7% 9.3% 0.892
Sunflower 6 12,939 85.9% 14.1% 0.856 89.9% 10.1% 0.897
Pop or Orn Corn 13 0 0.0% 100.0% 0.000 n/a n/a n/a
Barley 21 5,935 46.9% 53.1% 0.464 73.8% 26.2% 0.735
Durum Wheat 22 14,423 63.0% 37.0% 0.623 76.3% 23.7% 0.757
Spring Wheat 23 181,197 92.9% 7.1% 0.910 88.1% 11.9% 0.851
Winter Wheat 24 1,328 55.9% 44.1% 0.558 82.0% 18.0% 0.819
Rye 27 656 49.8% 50.2% 0.497 74.6% 25.4% 0.746
Oats 28 2,283 29.6% 70.4% 0.293 54.4% 45.6% 0.541
Millet 29 682 26.3% 73.7% 0.262 57.6% 42.4% 0.574
Canola 31 45,593 92.0% 8.0% 0.916 95.3% 4.7% 0.950
Flaxseed 32 4,794 60.9% 39.1% 0.607 80.2% 19.8% 0.800
Safflower 33 100 37.5% 62.5% 0.374 81.3% 18.7% 0.813
Mustard 35 151 42.5% 57.5% 0.425 92.1% 7.9% 0.921
Alfalfa 36 10,287 56.2% 43.8% 0.555 66.0% 34.0% 0.654
Other Hay/Non Alfalfa 37 17,722 40.3% 59.7% 0.384 58.2% 41.8% 0.563
Buckwheat 39 104 43.7% 56.3% 0.437 74.3% 25.7% 0.743
Sugarbeets 41 4,017 91.2% 8.8% 0.912 95.7% 4.3% 0.956
Dry Beans 42 12,639 82.0% 18.0% 0.817 87.5% 12.5% 0.873
Potatoes 43 1,294 69.6% 30.4% 0.696 94.7% 5.3% 0.946
Other Crops 44 90 24.9% 75.1% 0.249 70.3% 29.7% 0.703
Onions 49 0 0.0% 100.0% 0.000 n/a n/a n/a
Chick Peas 51 982 70.3% 29.7% 0.703 87.7% 12.3% 0.877
Lentils 52 2,460 76.4% 23.6% 0.763 89.3% 10.7% 0.893
Peas 53 11,244 85.1% 14.9% 0.849 89.1% 10.9% 0.889
Clover/Wildflowers 58 16 9.9% 90.1% 0.099 32.7% 67.3% 0.326
Sod/Grass Seed 59 6 8.0% 92.0% 0.080 50.0% 50.0% 0.500
Fallow/Idle Cropland 61 8,931 67.7% 32.3% 0.673 77.8% 22.2% 0.775
Triticale 205 24 12.4% 87.6% 0.124 48.0% 52.0% 0.480
Radishes 246 9 52.9% 47.1% 0.529 100.0% 0.0% 1.000
*Correct Pixels represents the total number of independent validation pixels correctly identified in the error matrix.
**The Overall Accuracy represents only the FSA row crops and annual fruit and vegetables (codes 1-61, 66-80, 92 and 200-255).
FSA-sampled grass and pasture. Non-agricultural and NLCD-sampled categories (codes 62-65, 81-91 and 93-199) are not included in the Overall Accuracy.
The accuracy of the non-agricultural land cover classes within the Cropland Data Layer is entirely dependent upon the USGS, National Land Cover Database (NLCD 2016). Thus, the USDA, NASS recommends that users consider the NLCD for studies involving non-agricultural land cover. For more information on the accuracy of the NLCD please reference <https://www.mrlc.gov/>.
Attribute_Accuracy_Value:
Classification accuracy is generally 85% to 95% correct for the major crop-specific land cover categories. See the 'Attribute Accuracy Report' section of this metadata file for the detailed accuracy report.
Attribute_Accuracy_Explanation:
The strength and emphasis of the CDL is crop-specific land cover categories. The accuracy of the CDL non-agricultural land cover classes is entirely dependent upon the USGS, National Land Cover Database (NLCD 2016). Thus, the USDA, NASS recommends that users consider the NLCD for studies involving non-agricultural land cover.
These definitions of accuracy statistics were derived from the following book: Congalton, Russell G. and Kass Green. Assessing the Accuracy of Remotely Sensed Data: Principles and Practices. Boca Raton, Florida: CRC Press, Inc. 1999. The 'Producer's Accuracy' is calculated for each cover type in the ground truth and indicates the probability that a ground truth pixel will be correctly mapped (across all cover types) and measures 'errors of omission'. An 'Omission Error' occurs when a pixel is excluded from the category to which it belongs in the validation dataset. The 'User's Accuracy' indicates the probability that a pixel from the CDL classification actually matches the ground truth data and measures 'errors of commission'. The 'Commission Error' represent when a pixel is included in an incorrect category according to the validation data. It is important to take into consideration errors of omission and commission. For example, if you classify every pixel in a scene to 'wheat', then you have 100% Producer's Accuracy for the wheat category and 0% Omission Error. However, you would also have a very high error of commission as all other crop types would be included in the incorrect category. The 'Kappa' is a measure of agreement based on the difference between the actual agreement in the error matrix (i.e., the agreement between the remotely sensed classification and the reference data as indicated by the major diagonal) and the chance agreement which is indicated by the row and column totals. The 'Conditional Kappa Coefficient' is the agreement for an individual category within the entire error matrix.