Several methods have been tried to estimate air temperature using satellite imagery. In this paper, the results of two machine learning algorithms, Support Vector Machine and Random Forest, are compared with Multivariate Linear Regression, TVX and Ordinary kriging. Several geographic, remote sensing and time variables are used as predictors. The validation is carried out using four different statistics on a daily basis allowing the use of ANOVA to compare the results. The main conclusion is that Random Forest with residual kriging produces the best results (R$^2$=0.612 $\pm$ 0.019, NSE=0.578 $\pm$ 0.025, RMSE=1.068 $\pm$ 0.027, PBIAS=-0.172 $\pm$ 0.046), whereas TVX produces the least accurate results. The environmental conditions in the study area are not really suited to TVX, moreover this method only takes into account satellite data. On the other hand, regression methods (Support Vector Machine, Random Forest and Multivariate Linear Regression) use several parameters that are easily calculated from a Digital Elevation Model, adding very little difficulty to the use of satellite data alone. The most important variables in the Random Forest Model were satellite temperature, potential irradiation and cdayt, a cosine transformation of the julian day.

The aim of this study is to evaluate three different strategies to improve classification accuracy in a highly fragmented semiarid area. i) Using different classification algorithms: Maximum Likelihood, Random Forest, Support Vector Machines and Sequential Maximum a Posteriori, with parameter optimisation in the second and third cases; ii) using different feature sets: spectral features, spectral and textural features, and spectral, textural and terrain features; and iii) using different image-sets: winter, spring, summer, autumn, winter+summer, winter+ spring+summer; and a four seasons combination. A 3-way ANOVA is used to discern which of these approaches and their interactions significantly increases accuracy. Tukey-Kramer contrast using a heteroscedasticity-consistent estimation of the kappa covariances matrix was used to check for significant differences in accuracy. The experiment was carried out with Landsat TM, ETM, and OLI images corresponding to the period 2000-2015. A combination of four images was the best way to improve accuracy. Maximum Likelihood, Random Forest and Support Vector Machines do not significantly increase accuracy when textural information is added, but do so when terrain features are taken into account. On the other hand, Sequential Maximum a Posteriori increases accuracy when textural features are used, but reduces accuracy substantially when terrain features are included. Random Forest using the three feature subsets and Sequential Maximum a Posteriori with spectral and textural features had the largest kappa values, around 0.9.