Revealing the potential of spectral and textural predictor variables in a neural network-based rainfall retrieval technique

ABSTRACT Estimating rainfall areas and rates from geostationary satellite images has the opportunity of both, a high spatial and a high temporal resolution which cannot be achieved by other satellite-based systems until now. Most recent retrieval techniques are solely based on spectral channels of the satellites. These retrievals can be classified as ‘purely pixel-based’ because no information about the neighbourhood pixels is included. Assuming that precipitation is highly correlated with cloud processes and therefore with cloud texture, textural information derived from the neighbourhood of a pixel might give valuable information about the cloud type and hence about a respective probability of the rainfall rate. To study the potential of textural variables to improve optical rainfall retrieval techniques, rainfall areas and rainfall rates were estimated over Germany for the year 2010 using a neural network approach. In addition to the spectral predictor variables from Meteosat Second Generation (MSG), different Grey Level Co-occurance Matrix (GLCM) based textural variables were calculated from all MSG channels. Using recursive feature selection, models were trained and their performance was compared to spectral-only models. Contrary to the expectations, the performance of the models did not increase when textural information was included.