Main content area

A new optimized algorithm for automating endmember pixel selection in the SEBAL and METRIC models

Bhattarai, Nishan, Quackenbush, Lindi J., Im, Jungho, Shaw, Stephen B.
Remote sensing of environment 2017 v.196 pp. 178-192
Landsat, algorithms, automation, energy balance, evapotranspiration, heat transfer, labor, models, remote sensing, statistical analysis, surface temperature, time series analysis, Florida, Oklahoma
In typical surface energy balance (SEB) models such as surface energy balance for land (SEBAL) and mapping evapotranspiration (ET) at high resolution with internalized calibration (METRIC), calibration of sensible heat flux (H) requires identification of endmember (i.e. hot and cold) pixels. Such pixel selection is typically done manually, which makes it labor intensive to apply SEBAL or METRIC over large spatial areas or long time series. In this paper, we introduce a new automated approach that uses an exhaustive search algorithm (ESA) to identify endmember pixels for use in these models. The fully automated models were applied on 134 near cloud-free Landsat images with each image covering one of four flux measurement sites covering a distinct land cover type in humid Florida or relatively drier Oklahoma. Observed land surface temperatures (Ts) at automatically identified hot and cold pixels were within 0.25% of manually identified pixels (both coefficient of determination, R2, and Nash-Sutcliffe efficiency, NSE, ≥0.98, and root mean squared error, RMSE, ≤1.31K). The new fully automated model performed better and demonstrated better consistency than an existing semi-automated method that used a statistical approach to subset coldest and hottest pixels within an image. Daily ET estimates derived from the automated SEBAL and METRIC models were in good agreement with their manual counterparts (e.g., NSE≥0.94 and RMSE≤0.35mmday−1). Automated and manual pixel selection resulted in similar estimates of observed ET across all sites. The proposed automated pixel selection approach greatly reduces time demands (e.g. approximately one image per hour vs. hundreds of image per hour) for applying SEBAL and METRIC and allows for their more widespread and frequent use. This automation can also reduce potential bias introduced by an inexperienced operator and extends the domain of the models to new users.