Neural network detection of potential thunderstorm zones from remote sensing data

Convection origination on the territory of Eurasia is noted up to 65-67° north latitudes, as noted in the annual reports on thunderstorm registration by Vaisala. Registration of thunderstorms by geostationary satellites in this territory is difficult. Therefore, the question of using satellite data from polar-orbiting spacecraft to monitor zones of convection development is relevant. The objectives of this study were to develop an architecture of algorithms for detection of atmospheric phenomena based on machine learning and neural networks, creation and training of algorithms for thunderstorm detection using ERA5 reanalysis data and further verification of the obtained algorithms and models on satellite sensing data of the MIRS program complex. The technology of selecting probable zones of thunderstorm development from satellite sensing data is a model of probabilistic detection of the presence or absence of atmospheric phenomena. With the help of calculations, the zones in which the atmospheric parameters more or less correspond to the conditions under which the hydrometeorological phenomena can form are highlighted. Calculations are carried out with the help of machine learning technology and neural networks. The paper presents the architecture of an algorithm for thunderstorm detection based on machine learning and neural network technology. A fully connected neural network is used, where the output signal of each neuron is fed as an input signal to all subsequent neurons. The neural network includes forty input neurons, twenty neurons on the first hidden layer, fifteen neurons on the second hidden layer, and two neurons corresponding to the necessary classification - presence/absence of thunderstorm on the output layer. The activation function on the hidden layers is Rectified linear unit (ReLU), the activation function on the output layer is Softmax. The algorithm was trained on the data of reanalysis ERA5 and verified on the data of satellite sounding of program complex MIRS which uses microwave measurements of AMSU/MHS devices of NOAA series, MetOp and ATMS device of Suomi NPP. Based on the results of the analysis of a number of data that were not involved in the model development and training process for May-October 2019-2020, an accuracy of 84 % was obtained. Additionally, tests were conducted on 109 meteorological stations located within the boundaries of 49.9°-60.35° N and 75.68°-88.67° E. For the periods: April-September 2021, the predictability value indicates that 96 % of the events are successfully classified (presence/absence of thunderstorms). The presented algorithm is dynamic. The learning procedure can be reinitialized when enough data is accumulated. This will allow us to take into account the appearance of new extremes of atmospheric characteristics. The use of the presented algorithm and the results of its calculations is promising as an additional tool for operational work.

Keywords

Authors

References