The Advanced Technology Microwave Sounder (ATMS) is a cross-track scanning microwave (MW) radiometer useful to retrieve Temperature (T) and Water Vapor (WV) atmospheric vertical profiles. Using spatial and temporal coincidences between ATMS observations and two different datasets of vertical T and WV, a global training dataset for machine learning purpose was built for the whole 2016. For each ATMS Brightness Temperatures acquisition, 32 levels of T (between 10 and 1000 hPa), and 23 levels of WV (between 200 and 1000 hPa), are used to train an algorithm based on Random Forests regression technique. A single RF was trained for each level and atmospheric variable, using the evaluation on the Out of Bag error to optimize the number of random selection of the input variables at each node splitting step, the number of trees in each forest and the minimum leaf size parameter, to avoid overfitting problem and obtain an accurate retrieval. Considering that the sounding below the precipitation level becomes unreliable, the precipitation-affected observations were removed from the training dataset by means of a pre-screening test based on BT. The results show an overall ability of the algorithm to retrieve T and WV vertical profiles in line with expectations.
Retrieval of temperature and water vapor vertical profile from atms measurements with random forests technique
Di Paola Francesco;Cersosimo Angela;Cimini Domenico;Gallucci Donatello;Gentile Sabrina;Geraldi Edoardo;Nilo Saverio T;Ricciardelli Elisabetta;Romano Filomena;Viggiano Mariassunta
2018
Abstract
The Advanced Technology Microwave Sounder (ATMS) is a cross-track scanning microwave (MW) radiometer useful to retrieve Temperature (T) and Water Vapor (WV) atmospheric vertical profiles. Using spatial and temporal coincidences between ATMS observations and two different datasets of vertical T and WV, a global training dataset for machine learning purpose was built for the whole 2016. For each ATMS Brightness Temperatures acquisition, 32 levels of T (between 10 and 1000 hPa), and 23 levels of WV (between 200 and 1000 hPa), are used to train an algorithm based on Random Forests regression technique. A single RF was trained for each level and atmospheric variable, using the evaluation on the Out of Bag error to optimize the number of random selection of the input variables at each node splitting step, the number of trees in each forest and the minimum leaf size parameter, to avoid overfitting problem and obtain an accurate retrieval. Considering that the sounding below the precipitation level becomes unreliable, the precipitation-affected observations were removed from the training dataset by means of a pre-screening test based on BT. The results show an overall ability of the algorithm to retrieve T and WV vertical profiles in line with expectations.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.