Optimizing Satellite-Based Latent Heating Rate Profiling Using a Convolutional Neural Network Heating (CNNH) Algorithm

Precise spatial distribution of latent heat released during precipitation formation is crucial to enhance weather forecasting and climate prediction accuracy. This study introduces an innovative convolutional neural network heating (CNNH) algorithm. The algorithm incorporates the vertical gradient of precipitation rate and air temperature at different altitudes as key inputs. It combines additional information from adjacent vertical layers and neighboring horizontal areas. To mitigate possible misjudgments of negative heating, a punishing mechanism was introduced with a latent heating (LH)-structure loss function within the optimization framework. By employing the adaptive differential evolution (ADE) algorithm, the optimal configuration of the network structure optimizes accurate LH retrieval. To evaluate the CNNH algorithm’s efficacy, a self-consistency check using weather research and forecasting (WRF) model simulation data was conducted. Furthermore, an inter-comparison study with four other algorithms using real satellite observations was undertaken. Evaluations found that the CNNH algorithm could precisely retrieve the primary characteristics of the horizontal and vertical structure, along with the temporal evolution process and statistical information of WRF simulated LH, in eastern China in August 2017. It effectively addressed the issues of overestimation of near-surface cooling and mixing layer heating, thereby outperforming selected artificial intelligence (AI) and physics-based LH algorithms. The intercomparison study between LHCNNH with four other published LH products based on the same GPM observations reveals that CNNH got comparable performance among them. However, specific differences among these algorithms highlight considerable uncertainties in multiple LH satellite remote sensing products and underscore the necessity for further improvements in satellite LH algorithms.