Research On Deep Learning Interpretation Methods And Their Application In Precipitation Prediction

Over the past decade,research and application of deep learning techniques have received a lot of attention and achieved remarkable results in many fields.However,the decision-making process of the models should be trustworthy in applications related to health,safety,and livelihood issues.The black-box feature of the deep models is one of the main factors that hinder their application,and the interpretation can improve the credibility and guide the improvement of the deep model.To address the problems of many optional variables,poor model performance and low model credibility in climate prediction,this thesis investigates the interpretation of deep learning models and applies them to specific climate science tasks(especially precipitation prediction)to solve the corresponding problems.The main research includes:(1)We propose an input variable selection for convolutional neural network(CNN)model.Firstly,a contribution measure of input variables is defined by interpreting feature importance in CNN.Then a greedy elimination algorithm is implemented for input variable selection.The proposed algorithm is applied to the statistical downscaling(SD)problem of monthlyscale precipitation in South China.The experimental results demonstrate that the proposed algorithm can reduce the 20 initial input variables to less than half and reduce the number of floating-point operations by more than1/5 while improving the model accuracy.The main input variables selected are also shown to be the key predictors of precipitation formation.(2)We present a locally-connected and multilayer feature fusion CNN model(CNN-LC-Fusion)base on the feature importance interpretation of existing SD CNN model.CNN-LC-Fusion is more suitable for large scale SD tasks,and is applied to a SD task of daily-scale precipitation in China.The experimental results demonstrate that CNN-LC-Fusion achieves much better accuracy than the common fully-connected CNN model in all the given evaluation metrics with only 1/44 of the number of parameters.The results of the explanatory analysis indicate that CNN-LC-Fusion solves the problem of possible input redundant information of the common fullyconnected CNN.(3)We propose a simple but effective scheme that achieves model pruning and input variable selection based on the interpretation of CNN convolutional kernel matrix.A globally comparable importance of input channels among all convolutional layers of CNN is defined.The proposed scheme is applied to the SD problem of daily precipitation in the Yangtze River basin,and the experimental results demonstrate that the proposed channel pruning scheme can reduce the number of channels in a given CNN model by about 2/3,reduce the number of floating-point operations in the model by more than 4/5,and improve the model accuracy.The pruning result also successfully selects 6 water vapor-related predictors from 20 initial variables that mainly affect precipitation.(4)We define the feature importance and feature contribution in CNN model by feature gradient and the product between feature gradient and feature itself,then use them for feature interpretation of statistical CNN model.The concepts are used to investigate the decision basis of daily precipitation probability and amount prediction in CNN-based SD models in China.The results reveal multiple interpretable statistical relationships between inputs and outputs,as well as the differences in the decision basis of the model in different regions and seasons.The results of the study can further improve the credibility of the deep model in precipitation prediction problems.To sum it up,this thesis investigates the interpretation method of deep learning models,realizes the selection of input variables,guides the design of better CNN structures,proposes the simplification of CNN structures and input variable selection based on channel pruning,and reveals the main basis for CNN decision procedure.The studies are applied to several applications related to SD of precipitation to solve specific problems.