Integration of shapley additive explanations with random forest model for quantitative precipitation estimation of mesoscale convective systems

Abstract

Mesoscale convective cloud systems have a small horizontal scale and a short lifetime, which brings great challenges to quantitative precipitation estimation (QPE) by satellite remote sensing. Combining machine learning models and geostationary satellite spectral information is an effective method for the QPE of mesoscale convective cloud, while the interpretability of machine learning model outputs remains unclear. In this study, based on Himawari-8 data, high-density automatic weather station observations, and reanalysis data over the North China Plain, a random forest (RF) machine learning model of satellite-based QPE was established and verified. The interpretation of the output of the RF model of satellite-based QPE was further explored by using the Shapley Additive Explanations (SHAP) algorithm. Results showed that the correlation coefficient between the predicted and observed precipitation intensity of the RF model was .64, with a root-mean-square error of .27 mm/h. The importance ranking obtained by SHAP model is completely consistent with the outputs of random forest importance function. This SHAP method can display the importance ranking of global features with positive/negative contribution values (e.g., current precipitation, column water vapor/black body temperature, cloud base height), and can visualize the marginal contribution values of local features under interaction. Therefore, combining the RF and SHAP methods provides a valuable way to interpret the output of machine learning models for satellite-based QPE, as well as an important basis for the selection of input variables for satellite-based QPE.

Ozone Concentration Estimation and Meteorological Impact Quantification in the Beijing‐Tianjin‐Hebei Region Based on Machine Learning Models

“…Moreover, the SHAP model, serving as an explanatory tool for machine learning models, has found extensive application in various studies involving PM 2.5 (Hou et al, 2022), O 3 (Ahmad et al, 2022;Cheng et al, 2023;Ghahremanloo et al, 2023;Nelson et al, 2023), precipitation (He et al, 2023;Li et al, 2023;Lin et al, 2023), and wind speed (Santos et al, 2023) investigations.…”

Impact of Updating Vegetation Information on Land Surface Model Performance

“…The most cross-correlated ones are: hydraulic conductivity and total soil depth; minimum stomatal resistance and soil moisture stress function; EVI and aridity; EVI and temperature; and the differences in high and low vegetation cover. Although most pairs of factors show correlations lower than 0.2, we accept collinearities as a caveat in our analysis since independence is not the usual case in Earth system sciences (W. Li et al, 2021;Silva et al, 2022;He et al, 2023;Wadoux et al, 2023).…”