A Novel Nonlinear Algorithm for Area-Wide Near Surface Air Temperature Retrieval

This paper reports a novel nonlinear algorithm for retrieving near surface air temperature over a large area using support vector machines with satellite remote sensing and other types of data. The steps include the following. 1) Establish the 1st sub model learning dataset and validation dataset, then obtain the 2nd tofth sub model learning datasets and validation datasets, using unmanned weather station data and predefined influential variables. 2) Retrieve Ta of the target area. 3) Correct the generated Ta images based on prediction errors using the inverse distance weighting interpolation. The novelty of this algorithm is to apply multiple sources of remote sensing data combined with data of unmanned weather stations, topography, ground cover, DEM, and astronomy and calendar rules. The results indicated that the model has high accuracy, reliability, and generalization ability. Factors such as cloudiness, ground vegetation, and water vapor show little interference, so the model seems suitable for large area retrieving under natural conditions. The required high-performance computation was achieved by a CPU + GPU isomery and synergy parallel computation system that improved computing speed by more than 1000-fold, with easily extendable computing capability. We found that the current algorithm is superior to seven major split-window algorithms and their best combined algorithms based on prediction errors, root-meansquare errors, and the percentage of data points with <;3°C absolute error. Our SVM approach overcomes shortcomings of classical temperature remote sensing technologies, and is the first report of such application.