Description: Remote sensing air temperature mostly relies on linear algorithms that produce significantly variable results depending on various weather conditions. Recently, a novel nonlinear algorithm based on support vector machine (SVM) was reported with improved prediction accuracy by using multiple types of data including satellite and unmanned weather station, land coverage imagery, digital elevation model, astronomy, and calendar. To further improve the accuracy and consistence, this paper reports a selective arithmetic mean (SAM) approach for optimization of a previously reported SVM algorithm for area-wide near surface air temperature remote sensing using satellite and other types of data. Using Guangxi province as the study area, the results show that this new SAM approach significantly improved the overall retrieving quality over the previously reported simple arithmetic mean approach. The SAM approach has high tolerance to cloud, ground vegetation, and vertical and spatial spectrum variations, with superb prediction errors (absolute error, AE) and root mean square errors concentrated around 0.7 and 0.8 °C, respectively. The prediction error patterns with different atmosphere water content, enhanced vegetation index, and spatial spectrum were similar under all examined conditions. After SAM operations, the prediction error patterns showed a deep gap near a set error threshold ${boldsymbol delta} _{i}$ , especially near δ 0 (δ 0 ± 0.2) in every examined situation. SAM also produces significantly lower errors at AE ≥ δ 0 ≥ 0. The SVM model with SAM optimization minimizes the shortcomings of the classical temperature remote sensing technologies and is suitable for area-wide retrieving under natural conditions. Four modeli- g principles are summarized for building superb models.