Land and Sea Surface Temperature Estimation by Support Vector Regression

Abstract

Land and sea surface temperatures are important quantities for many hydrological and meteorological models and satellite infrared remote sensing represents an effective way to map them on global and regional scales. A supervised approach, based on support vector regression, has recently been developed to estimate surface temperature from satellite images. Such a strategy requires the user to set several internal parameters. Moreover, in order to integrate the resulting estimates into hydrological or meteorological data-assimilation schemes, a further important input is the statistics of the regression error affecting each pixel. In this chapter, we propose a method to automatically set the input parameters and two techniques to model the statistics of pixelwise regression error. Parameter optimization is achieved by numerically minimizing a generalization-error bound that can be computed by using only the training set, and error modelling is performed by integrating a Bayesian support vector regression approach with maximum-likelihood or confidence-interval supervised parameter estimators. Some parts of this chapter have been derived from papers (Moser and Serpico

Estimation of Air Surface Temperature From Remote Sensing Images and Pixelwise Modeling of the Estimation Uncertainty Through Support Vector Machines

“…In general, if the region around the measurement station is thermally diverse, the in situ measurement might not be representative of the overall thermal behavior of the whole resolution cell. This is a common issue when supervised regression is applied to bio/geophysical parameter retrieval from coarse resolution imagery (e.g., [8], [18], [37], [42], [43]). The possible difference between the overall temperature of the pixel area and the point value measured on the station can be considered as an additive noise term that affects each training sample.…”

“…In the applications to LST and SST, this "Powell-span-bound" (PSB) method has been demonstrated to be capable of providing estimation accuracies very similar to those of grid searches with remarkably shorter computation times. Details of this can be found in [17] and [18].…”

“…A pattern recognition approach was recently proposed by the authors to estimate LST and sea surface temperature (SST) from satellite TIR data [17], [18]. It is a supervised approach based on support vector machines (SVMs) [19] and compared to classical split-window techniques (SWTs) for LST or SST [20], it usually exhibits higher estimation accuracy, although at the cost of longer computation times.…”

“…It is a supervised approach based on support vector machines (SVMs) [19] and compared to classical split-window techniques (SWTs) for LST or SST [20], it usually exhibits higher estimation accuracy, although at the cost of longer computation times. Unlike SWT, this SVM-based approach also allows the pixelwise statistics of the estimation error to be modeled, thus endowing each LST/SST estimate map with an accompanying map that expresses the pointwise uncertainty associated with the retrieved temperature [18], [21]. An alternate nonlinear approach that also provides pixelwise characterization of estimation uncertainty is Gaussian process regression (GPR) that is based on a Bayesian nonparametric formulation through multidimensional Gaussian processes whose covariance structures are defined by suitable kernel functions [22].…”

“…In this paper, a method for the estimation of air temperature near the Earth's surface from satellite TIR images is proposed by extending the aforementioned SVM-based methods developed in [17], [18], and [21] for LST and SST, integrating them with case-specific techniques aimed at computing periodic (e.g., monthly) statistics of air temperature, and experimentally validating the results in this new applicative framework. SVMs rather than GPR have been used because of their sparsity properties [19], which favor robustness to overfitting and which minimize the computational burden associated with the application of a trained estimator to unknown samples.…”