Estimation of the leaf area index from Landsat over the contiguous US

Authors: KANG, Y. - University Of Wisconsin; OZDOGAN, M. - University Of Wisconsin; Gao, Feng; Anderson, Martha; White, William - Alex; YANG, Y. - US Department Of Agriculture (USDA); YANG, Y. - US Department Of Agriculture (USDA); ERICKSON, T. - Google

Submitted to: Remote Sensing of Environment
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 2/27/2021
Publication Date: 3/16/2021
Publication URL: https://handle.nal.usda.gov/10113/7313360
Citation: Kang, Y., Ozdogan, M., Gao, F.N., Anderson, M.C., White, W.A., Yang, Y., Yang, Y., Erickson, T. 2021. Estimation of the leaf area index from Landsat over the contiguous US. Remote Sensing of Environment. 258:112383. https://doi.org/10.1016/j.rse.2021.112383.
DOI: https://doi.org/10.1016/j.rse.2021.112383

Interpretive Summary: Leaf area index (LAI) is a key biophysical parameter used for vegetation growth monitoring and water resource management. Current LAI data products derived from satellite remote sensing imageries are typically generated at coarse spatial resolution (0.25 to 1 km), which is often too coarse for many agricultural and hydrological applications at field scales. This paper proposes an approach to map LAI at 30-m resolution using Landsat images and the Moderate Resolution Imaging Spectroradiometer (MODIS) LAI for the Contiguous US (CONUS). The algorithm was built on the Google Earth Engine, which allows for the generation of long-term 30-m LAI records from the 1980s using Landsat. Results show good agreements with the ground LAI measurements over various landscapes. This paper provides a feasible method for producing LAI for CONUS at field scales, which is essential for large-area crop conditions and water use monitoring.

Technical Abstract: Leaf Area Index (LAI) is a fundamental variable for quantifying vegetation dynamics and is an essential input to many land surface and atmospheric models. Long-term LAI maps are typically generated with satellite images at coarse spatial resolution (0.25 to 1 km), which is often inadequate to resolve the spatial heterogeneity for many agricultural and hydrological applications. In this paper, we propose an approach to map LAI at 30-meter resolution based on Landsat images for the Contiguous US (CONUS). The algorithm was driven by 1.6 million high-quality and well-balanced samples derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) LAI and Landsat surface reflectance (Landsat 5, 7, and 8) products from 2006 to 2018. The samples were screened for spatial homogeneity within the 500-meter MODIS pixel footprint. Based on these samples, we trained separate random forest regressors for LAI estimation using Landsat surface reflectance and sensor viewing geometries for eight biomes from the National Land Cover Database (NLCD) and three Landsat sensors (Landsat 5, Landsat 7, and Landsat 8). The Mean Absolute Error (MAE) between aggregated Landsat LAI and MODIS LAI ranged from 0.15 to 0.79 for different biomes. The Landsat LAI algorithm was validated using ground measurements in eight study sites located in forests, grassland, and cropland landscapes. The RMSE was between 0.52 and 0.91 m2/m2, and the MAE ranges from 0.43 to 0.69 m2/m2. In addition, we found that the Landsat spectral signals of MODIS LAI retrievals with and without saturation showed substantial overlap (i.e., spectral overlap), which led to a trade-off in the accuracy between medium (3 to 4 m2/m2) and high LAI (4 – 6 m2/m2, i.e. saturated) values. A balanced sample design between saturated and unsaturated MODIS LAI is thus required to achieve the optimal overall accuracy. The estimation bias, introduced by the spectral overlap, can be further reduced using a novel noise detection technique based on the random forest classifier. The proposed algorithm was built on the Google Earth Engine, which allows for the generation of long-term high-resolution LAI records from the 1980s using Landsat images. Our findings also highlight the need for in-depth analyses of the impact of sample imbalance on regression-based modeling of remote sensing data.