Calculation of Ice Cloud Bulk Scattering Properties Based on Deep Reinforcement Learning

The potential of satellite observations at millimeter and submillimeter wavelengths for retrieving cloud ice has been demonstrated with data from airborne instruments. A major limitation in inversions of ice hydrometeors is the lack of accurate calculations of the bulk scattering parameters of ice clouds. The most uncertain aspect of modeling the millimeter- and submillimeter-wave scattering properties of ice hydrometeors lies in their shape. To address this issue, instead of calculating the bulk scattering properties of ice clouds by integrating single-shaped ice hydrometeors, this letter proposes the construction of an enhanced neural network to directly fit the bulk scattering parameters of ice clouds. The network parameters are optimized using actual satellite observations, such as Advanced Technology Microwave Sounder (ATMS) through deep reinforcement learning (DRL). Finally, a validation experiment is conducted by comparing the calculated brightness temperature (BT) with the actual ATMS-observed BT. The results demonstrate that a 21.3% reduction in the average error of BT calculated by the enhanced network compared to that obtained from the nonspherical particle model.