Toward an Operational Land Surface Temperature Algorithm for GOES

Abstract

or most land surface temperature (LST) regression algorithms, a set of optimized coefficients is determined by manual separation of the different subdivisions of atmospheric and surface conditions. In this study, a machine-learning technique, the regression tree (RT) technique, is introduced with the aim of automatically finding these subranges and the thresholds for the stratification of regression coefficients. The use of RT techniques in LST retrieval has the potential to contribute to the determination of optimal regression relationships under different conditions. Because of the lack of split-window channels for the Geostationary Operational Environmental Satellite (GOES) M?Q series (GOES-12?GOES-15, plus GOES-Q), a dual-window LST algorithm was developed by combining the infrared 11-?m channel with the shortwave-infrared (SWIR) 3.9-?m channel, which presents lower atmospheric absorption than does the infrared split-window channels (11 and 12 ?m). The RT technique was introduced to derive the regression models under different conditions. The algorithms were used to derive the LST product from GOES observations and were evaluated against the 2004 Surface Radiation budget network. The results indicate that the RT technique outperforms the traditional regression method.