Satellite Residual Two‐Stage Convolutional Neural Network Heating (R2CH) Algorithm for Retrieving Latent Heat Between the Precipitation Top and the Cloud Top

Latent heat release is associated with phase changes in water. Current physical‐based satellite latent heating (LH) retrieval methods mainly depend on precipitation radar observations, which detect hydrometeors only below the precipitation top height (PTH), thereby limiting the LH estimation to the inner precipitation region. However, phase transitions between the PTH and the cloud top height (CTH) also release non‐negligible LH, which accounts for up to 9.74% of the total based on statistical results. Neglecting this component can lead to systematic underestimation of LH. To address this issue, this study develops a Residual Two‐stage Convolutional Neural Network Heating (R2CH) algorithm to retrieve LH associated with non‐precipitating clouds above the PTH. The method adopts a residual learning strategy within a two‐stage framework, where a convolutional framework refines LH retrieval using CTH, PTH, air temperature, and vertical derivation of precipitation rates from satellite observation. Compared with Weather Research and Forecasting (WRF) model simulations, R2CH effectively captures LH from non‐precipitating clouds above the PTH and reproduces both horizontal and vertical structures. Relative to the Convolutional Neural Network Heating (CNNH) and two official GPM LH products, the R2CH retrieval based on GPM and Himawari‐8 observations addresses the underestimation of LH above the precipitation tops from CNNH and becomes more consistent with the two reference products at that layer. Nevertheless, the LH above PTH retrieved by R2CH is generally weaker than that from the two other products. More validation studies are needed to determine which method is more accurate.