Improvement Of Rain Area Delineation Schemes Based On Machine Learning Algorithms

Two machine learning scheme based on the Gradient Boosting Decision Tree(GBDT)and Random Forest(RF)algorithm is developed to improve the accuracy of rain area delineation for daytime,twilight and nighttime modules using Advanced Himawari Imager-8(AHI-8)geostationary satellite data and United States Geological Survey(USGS)Digital Elevation Model(DEM)data.The Gradient Booting Decision Tree and Random Forest algorithm are able to efficiently manage non-linear relationships among high-dimensional data without being affected by over-fitting problems.The new delineation module utilizes several features related to physical variables,including cloud top heights,cloud top temperatures,cloud water paths,cloud phases,water vapor,temporal changes and orographic variations.The scheme procedure is as follows.First,we perform extensive experiments to optimize module parameters such that the equitable threat score(ETS)reaches its maximum value.Then,the GBDT-based and RF-based modules are trained and classified with optimum parameters.Finally,validation datasets are applied to test the true performance of the GBDT-based modules.Agreement between estimations and observations of the ground-based rain gauges is investigated.The results show that ETS values of the GBDT-based modules are 0.42 for the daytime,0.30 for the twilight period,and 0.32 for the nighttime.The cloud water path and cloud phase features make high contributions to the modules.For the RF-based modules,the ETS values are 0.42 for the daytime,0.29 for the twilight period and 0.31 for the nighttime.Comparisons drawn with two probability-related methods show that our new schemes present great advantages in terms of statistical scores on overall performance.For three periods in general,the GBDT and RF modules have increased 16%,42%and 28%of ETS values and 8%,-12%and 13%of POD values and decreased 15%25%and 12%of FAR values.