由于水文气象监测资料缺乏、预报预警手段有限等特点，山区中小流域的洪水灾害是我国当前受灾面积广、致死率高的自然灾害之一。与此同时，在全球气候变暖和我国快速城镇化的背景下，未来洪水灾害可能会造成更大损失，亟待发展更有效的中小流域洪水预警预报方法。完善的中小河流洪水预报预警系统依赖于实时的降雨观测、高时空分辨率的定量降雨预报以及有效的预报和预警方法。本论文从基于临界雨量的山洪预警、基于GPM卫星实时降雨的实时洪水预报和基于定量降雨预报的短期洪水预报等方面展开研究工作，为中小流域洪水提供了不同预见期的预警预报方法。论文选择我国不同气候区的四个典型小流域，基于GBHM模型原理，分别构建了分布式水文模型并开展了水文模拟。基于观测降雨和模拟的逐时流量过程、土壤水动态，利用频率分析和线性划分算法，确定了各流域不同土壤饱和度下的临界雨量指标。根据实测历史流量和降雨资料，对基于临界雨量的山洪预警方法进行了评估，检验了方法的可靠性。进一步在全国范围，基于GBHM模型，建立了空间分辨率为0.01°的产流过程及土壤水模拟系统，采用SCS无因次单位线法，在全国范围内选择1849个山区小流域计算汇流过程，模拟了1849个小流域2003-2014年的逐时流量过程和土壤饱和度。根据上述模拟结果，分析确定了1849个山区小流域的动态临界雨量指标。为了提高中小流域的实时洪水预报能力，论文评估了GPM卫星实时降水的适用性。在湘江流域的评估结果表明， GPM卫星实时降水产品（IMERG）具有较高精度。IMERG降水产品具有准实时性、时空分辨率高，洪水预报精度较好，在中小流域实时洪水预报中具有应用潜力。论文基于CNN和LSTM神经网络，对欧洲中期天气预报（ECMWF）降雨进行空间降尺度，进一步评估了定量降雨预报在中小流域短期洪水预报中的应用潜力。与ECMWF预报降水相比，在预见期1天至2周内，该方法均能够提高降水预报精度，改善效果虽然随预报期延长而降低，但是对系统性偏差的改善效果仍然显著，从而提高了中小流域短期洪水预报的能力。
Due to the limited capacities of monitoring, forecasting and warning, flash flood has been recognized as one of the most extensive disasters in China that leads to high mortality. In addition, global warming and national urbanization may further increase people’s exposure risk and vulnerability to flash floods in China. Thus, development of the flash floods forecasting and warning is urgently necessary. Reliable warning systems must depend upon the accurate real-time rainfall estimation, quantitative rainfall forecast with high spatial temporal resolution, and effective warning/forecasting methods. Aims to solving the difficulties in the early warning of flash floods, this thesis proposed a new method to determine the rainfall threshold for flood warning and then applied this method to the montanous areas in China, and tested applicability of the GPM satellite precipitation as well as the quantitative precipitation forecast for flood forecasting.This study implements a geomorphology-based hydrological model (GBHM) in four small mountainous catchments in different climate regions over China. Then it proposes a method to determine the rainfall threshold for flood warning by using frequency analysis and binary classification based on long-term hydrological simulation by GBHM. It conducts a comprehensive evaluation of the rainfall threshold and finds that the proposed method produces reasonably accurate flash flood warnings in the study catchments. Furthermore, a national modeling system with resolution of 0.01 degree has been built to simulate the runoff generation and soil water dynamic processes based on GBHM, and the SCS dimensionless unit hydrograph method has been used for flow routing in 1849 selected small catchments over the mountainous region of China. It has simulated the runoff processes and the soil moisture dynamics from 2003 to 2014, based on the simulated results the rainfall thresholds for flash flood warning are derived for the 1849 study catchments.To improve the flood forecasting in small and medium-sized catchments, this thesis applies the latest near real time GPM precipitation in Xiangjiang River Basin. Statistically, near real time IMERG products show high accuracy. IMERG precipitation products have strong timeliness, high spatial-temporal resolution, good accuracy in hydrological simulation, thus is valuable for real-time flood forecasting in small and medium-sized catchments. In addition, it proposes a deep neural network composed of CNN and LSTM to downscale the ECMWF precipitation. Results show considerable improvement of precipitation accuracy compared to the ECMWF-Interim reanalysis and benchmark downscaling approaches including quantile mapping and SVM, etc. The proposed method can provide more accurate precipitation prediction at lead time from 1 day up to 2 weeks. The superiority decreases along forecast lead time, but its improvement of systematic deviation does not fade out. It can also improve the hydrological simulation performance significantly.
