Publications

Zhang, JZ; Okin, GS; Zhou, B (2019). Assimilating optical satellite remote sensing images and field data to predict surface indicators in the Western US: Assessing error in satellite predictions based on large geographical datasets with the use of machine learning. REMOTE SENSING OF ENVIRONMENT, 233, UNSP 111382.

Abstract
Indicators of vegetation composition, vegetation structure, bare ground cover, and gap size in drylands potentially gives information about the condition of ecosystems, in part because they are strongly related to factors such as erosion, wildlife habitat characteristics, and the suitability for some land uses. Field data collection based on points does not produce spatially continuous information about surface indicators and cannot cover vast geographic areas. Remote sensing is possibly a labor- and time-saving method to estimate important biophysical indicators of vegetation and surface condition at both temporal and spatial scales impossible with field methods. Regression models based on machine learning algorithms, such as random forest (RF), can build relationships between field and remotely sensed data, while also providing error estimates. In this study, field data including over 15,000 points from the Assessment, Inventory, and Monitoring (AIM) and Landscape Monitoring Framework (LMF) programs on Bureau of Land Management (BLM) lands throughout the Western U.S., Moderate Resolution Imaging Spectroradiometer (MODIS) bidirectional reflectance distribution function (BRDF) parameters, MODIS nadir BRDF-adjusted reflectance (NBAR), and Landsat 8 Operational Land Imager (OLI) surface reflectance products with ancillary data were used as predictor variables in a k-fold cross-validation approach to RF modeling. RF regression models were built to predict fourteen indicators of vegetation cover and height, as well as bare gap parameters. The RF model estimates exhibited good correlations with independent samples, with a low bias and a low RMSE. External cross-validation showed good agreement with out-of-bag (OOB) errors produced by RF and also allowed mapping prediction uncertainty. Predicted distribution maps of the surface indicators were produced by using these relationships across the arid and semiarid Western U.S. The bias and RMSE distribution maps show that the sample insufficiency and unevenly pattern of sample strongly impact the accuracy of the RF regression and prediction. The results from this study clearly show the utility of RF as a means to estimate multiple dryland surface indicators from remotely sensed data, and the reliability of the OOB errors in assessing the accuracy of the predictions.

DOI:
10.1016/j.rse.2019.111382

ISSN:
0034-4257