A method for estimating air temperature at 2-m height from satellite data and for modeling the estimation uncertainty

An algorithm has been developed to map air temperature at 2-m height, based on satellite images and on in situ temperature measurements collected by micrometeorological stations. The method is based on support vector machines (SVMs) and also provides a pixelwise model for the statistics of the estimation error. The algorithm has been implemented in C++.

Methodology

Air temperature at 2-m height above ground surface is usually evaluated through in situ measurements, which provide only spatially sparse observations. A spatially continuous retrieval is made possible by using satellite images acquired in the infrared wavelength range. However, satellite-based air-temperature estimates are usually endowed only with global error statistics (e.g., a unique value of root-mean square error for a given satellite product), whereas a pointwise characterization of the estimation error is also important.

The team of the University of Genoa (Unige) involved in ENDORSE has developed an algorithm to jointly estimate air temperature from satellite data and also model the pixelwise statistics of the estimation error. The algorithm combines support vector machines for air-temperature regression, the minimization of a generalization bound (the so-called “span bound” functional) for parameter optimization, and nonstationary multidimensional stochastic processes for pixelwise statistical modeling. The method is fed with satellite infrared images and with a set of in situ temperature measurements acquired in the area of interest. Its application is feasible in all areas where networks of micrometeorological stations collecting (also) air temperature are deployed. The method has been developed by extending previous techniques proposed and validated in the context of different applications, i.e., land-surface and sea-surface temperature retrieval from satellite data ([1] [2]).

The method has been experimentally assessed with MODIS and MSG images of PACA. The accuracy of the air-temperature estimates, compared with in situ data, is similar to the one obtained by state-of-the-art techniques. However, unlike such previous techniques, the proposed method also provides, for each air-temperature map, also an accompanying map that expresses the uncertainty in the air-temperature estimation at each pixel. The comparison with ground data suggests this pixelwise characterization of the error statistics to be accurate.

The submissions of this work to an international conference and possibly to an international journal are planned for the next months.

References

[1] Moser and Serpico, IEEE Transactions on Geoscience and Remote Sensing, 47:909-921, 2009

[2] Moser and Serpico, IEEE Geoscience and Remote Sensing Letters, 6:448-452, 2009