Recently, many researchers have used machine learning in remote sensing applications. One of the biggest advantages of machine learning algorithms is to model non-linear relationship between a dependent variable and independent variables without any distributional assumption.
In this study, we have explored the potential use of two supervised machine learnings (decision tree and random forest) in rainfall estimation using dual-polarization radar variables. The radar variables simulated by drop size distribution at ground disdrometer are used to train machine learning algorithms. Several different configurations for machine learning algorithms are considered with different sets of dependent and independent variables. The tuning of random forest model is also tested. It is revealed that specific differential phase is the most important variable to predict rainfall rate while differential reflectivity is the most important to explain residual. The models are evaluated by 10-fold cross validation. It is shown that the machine learning algorithms outperform traditional Z-R relationship. The best model is the random forest model using residual with classified training set.
Further, we have applied the model to Mountain Myeonbong S-band dual-polarization radar data and validated with rain gauges in Automatic Weather Stations. The results show that residuals have spatial variability. It means that residuals from Z-R relationship have structure in space. Convective areas have negative residuals while weak and stratiform areas have positive residuals. The rainfall rates for all the pixels have adjusted with estimated residuals. Adjusted rainfall rates show good agreement with rain gauge especially at high rainfall rate because they correct the rainfall rate by residual.