Sampling techniques in high-dimensional spaces for the development of satellite remote sensing database

[1] This study presents various strategies to sample databases from large atmospheric data sets in high-dimensional spaces for satellite remote sensing applications. In particular, two sampling algorithms are examined: the traditional uniform sampling that lists all possible situations and the clustering sampling (K-means) that respects the natural variability probability distribution functions. In order to assess the quality of both sampling methods, the extracted databases are used to extract first guesses for satellite remote sensing schemes. They are also employed as training databases for the calibration of statistical retrieval algorithms. The analysis of these sampling algorithms is illustrated by constructing both a first guess (FG) extraction and a retrieval databases of temperature and water vapor profiles over sea for the Atmospheric Microwave Sounding Unit (AMSU) instrument. The advantages and problems of each sampling approach are thoroughly examined and sensitivity studies are conducted to analyze the impact on the FG extraction and retrieval of various algorithmic parameters such as the distance being used, the size of the databases, or the instrumental noise sensitivity. The K-means clustering algorithm, not yet used for this kind of problems, is very efficient compared to the more traditional uniform sampling approach. It is also shown that it is important to have quasi-automatic and flexible tools that can be used to generate problem-specific databases.