Retrieval of total precipitable water from INSAT-3D Imager observations using deep neural network

Total precipitable water (TPW) is a key parameter for monitoring the development of the various weather phenomenon, viz., deep convective systems, thunderstorms, etc. Moreover, various nowcasting applications such as heavy rainfall alerts, cloudbursts, thunderstorms, etc. require high spatio-temporal TPW information, which can be achieved with the satellite-based observations from geostationary platform. Therefore, the present study deals with the estimation of TPW from observations in the water vapour (WV) absorption (6.5-7.0 lm) and thermal infrared (TIR) channels (TIR1 (10.3-11.2 lm); TIR2 (11.5-12.5 lm)) of Imager onboard the third generation Indian National Satellite (INSAT-3D) deployed in a geostationary orbit, using the machine-learning technique. The deep neural network (DNN) technique is utilized here to estimate TPW from INSAT-3D Imager observations over the Indian subcontinent region. Herein, the observations from three channels (WV, TIR1 & TIR2) of INSAT-3D Imager for three years from 2018 to 2020 are collocated with TPW from European Center for Medium range Weather Forecast (ECMWF) reanalysis (ERA5) available at 0.25 degrees uniform spatial resolution to prepare a matchup dataset. The input vector consists of Julian day, geolocation (latitude/longitude), brightness temperatures of WV, TIR1 and TIR2 channels of INSAT-3D Imager, surface emissivity of all three channels and the satellite zenith angles. The ERA5 TPW is the output (predictand) of the DNN. For the training of the DNN, 70 % of the total collocated matchups are selected randomly, and the rest 30 % are used for the testing. The DNN is developed for land and ocean surfaces separately, due to a larger variation in the emissivity over land. The assessment of the developed DNN over both surfaces show the similar statistics. The quality evaluation of the training and testing dataset shows the negligible biases and root-mean-squared error (RMSE) of X 3.5 mm over the land and X 4.5 mm over the ocean, respectively, in the predicted TPW when compared with ERA5 TPW. However, the bias (RMSE) in the predicted TPW is observed-0.3 (4.4) mm and 0.4 (4.8) mm, respectively over the land and ocean surfaces with reference to the ERA5 TPW for the independent testing of the developed DNN, which was performed on the observations of INSAT-3D for 2021. Additionally, the comparison with radiosondes' TPW demonstrates the 6.7 mm RMSE with bias of-0.7 mm for 2021. (c) 2024 COSPAR. Published by Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.