Deriving Precipitation and Latent Heating Profiles from NASA TROPICS Smallsat through a Two-Step Ensemble Convolution Neural Network Method

Abstract. This study develops and implements a two-step ensemble convolutional neural network (CNN) framework to retrieve precipitation and latent heating (LH) profiles from passive-microwave observations from NASA’s Time-Resolved Observations of Precipitation structure and storm Intensity with a Constellation of Smallsats (TROPICS) mission. In the first step, a precipitation CNN is trained using collocated TROPICS brightness temperatures (Tb) and Global Precipitation Measurement (GPM) combined radar-radiometer (2BCMB) precipitation estimates. Ensemble CNN training shows that retrieval skill is controlled by interactions among a set of choices for hyperparameters, including optimizer settings, architecture depth, and loss weighting. Orbital-scale global evaluation is also required for final model selection because precipitation-intensity distributions differ across tropical environments. Relative to the baseline TROPICS Precipitation Retrieval and Profiling Scheme (PRPS), the CNN retrieval reproduces the large-scale climatological structure of GPM precipitation more accurately in terms of correlations and biases. The CNN produces greater frequencies of heavier precipitation than PRPS; however, the heaviest precipitation frequencies remain underrepresented in the CNN due to the limited number of training samples. In the second step, the same TROPICS Tb patches are used to retrieve GPM Convective–Stratiform Heating (CSH) LH profiles from 1 to 12 km altitude, while the first-step CNN precipitation retrieval provides an additional physical constraint through a precipitation–LH budget linkage. The resulting LH climatology compares favorably with GPM CSH products in both horizontal maps and zonal vertical structure. The CNN reproduces the broad tropical heating maxima, their mid-tropospheric placement, and their meridional extent over land and ocean, with the best agreement generally occurring in the 5–7.5 km layer. These results demonstrate that, despite limited single-footprint information content, a two-step ensemble CNN framework can retrieve physically realistic precipitation and LH structure from TROPICS observations.