Validating NISAR’s cropland mapping approach and the USDA/NASS Cropland Data Layer against ground truth data in a fragmented urban agricultural region

Authors: Kraatz, Simon; LAMB, B - Us Geological Survey (USGS); HIVELY, W - Us Geological Survey (USGS); Jennewein, Jyoti; Gao, Feng; Cosh, Michael; SIQUEIRA, P - University Of Massachusetts

Submitted to: Sensors
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 10/18/2023
Publication Date: 10/20/2023
Citation: Kraatz, S.G., Lamb, B.T., Hively, W.D., Jennewein, J.S., Gao, F.N., Cosh, M.H., Siqueira, P. 2023. Validating NISAR’s cropland mapping approach and the USDA/NASS Cropland Data Layer against ground truth data in a fragmented urban agricultural region. Sensors. 23(20). https://doi.org/10.3390/s23208595.
DOI: https://doi.org/10.3390/s23208595

Interpretive Summary: Cropland mapping (crop/non-crop) is important for applications relating to monitoring agricultural health, management practices, food security and decision-making. USDA/NASS Cropland Data Layer (CDL) is probably the most widely used dataset for this purpose over CONUS, but there are important limitations. The CDL, like many other modern cropland products are derived from optical sensors are therefore limited by clouds cover. Also, these mapping approaches rely on labor intensive model development each year (modeling, including training, validation, and calibration). The presented approach differs from optical approaches in mainly two ways: (1) radar data is used, allowing for observations during nighttime and cloudy conditions, and (2) it utilizes an inexpensive approach for cropland mapping based on the coefficient of variation. This approach will also be used in the upcoming NASA-ISRO radar mission but are limited to using relatively shorter wavelength C-band radar due to lack of L-band radar at biweekly or better repeat. We study crop/non-crop identification of the approaches over a USDA Long Term Agricultural Research (LTAR) at the Beltsville Agricultural Research Center (BARC), located in Maryland, USA. BARC has a long and detailed record of ground truth data, allowing accurate evaluation of cropland products. The study extends from 2017 to 2021 and encompasses 54 agricultural production fields, and also tests these products over 39 non-crop fields consisting of forest and built-up areas. Results show that the CA outperforms the CDL at this site, showing the utility of this inexpensive approach for cropland mapping.

Technical Abstract: A general limitation of landcover mapping accuracy assessments is availability of ground-truth data, necessitating the use of proxy datasets that offer consistent large-scale data (Contiguous United States) at sub-field-scale resolutions. The USDA/NASS Cropland Data Layer (CDL) is a popular agricultural land cover dataset due to its high accuracy (>80%) and resolution (30 m), and inclusions of many land cover and crop types. However, because the CDL is derived from satellite imagery and has resulting uncertainties, comparisons to available in situ data are necessary for verifying classification performance. This study compares the cropland mapping accuracies (crop/non-crop) of an optical approach (CDL) and the radar-based crop area (CA) approach to be used for the upcoming NASA-ISRO Synthetic Aperture Radar (NISAR) L- and S-band mission but using Sentinel-1 C-band data. CDL and CA performance was evaluated against a ground truth dataset including 54 agricultural production and research fields located at USDA’s Beltsville Ag-ricultural Research Center (BARC) in Maryland USA. To better evaluate non-crop mapping ac-curacy, we also evaluated CA performance over non-crop sites consisting of twenty-six built-up and thirteen forest sites at BARC. We note that the CDL and CA have good pixel-wise agreement be-tween them (87%). However, when comparing to ground truth data, the 2017-2021 mean accu-racies were 77% and 96% for crop, 100% and 94% for built-up, and 100% and 100% for forest, for the CDL and CA, respectively. The overall accuracy was 86% for the CDL and 96% for CA, due to the CDL underdetecting crop cover at BARC. We also note that annual accuracy levels varied less for CA (from 91% to 98%) than for the CDL (from 79% to 93%). This study demonstrates that computationally inexpensive radar-based cropland maps can give accurate results over complex landscapes with accuracies similar to or better than optical approaches, and that radar CA mapping can be competitive with optical approaches.