Under global climate change, urban flooding occurs frequently, leading to huge economic losses and human casualties. Extreme rainfall is one of the direct and key causes of urban flooding, and accurate rainfall estimates at high spatiotemporal resolution are of great significance for real-time urban flood forecasting. Using existing rainfall intensity measurement technologies, including ground rainfall gauges, ground-based radar, and satellite remote sensing, it is challenging to obtain estimates of the desired quality and resolution. However, an approach based on processing distributed surveillance camera network imagery through machine learning algorithms to estimate rainfall intensities shows considerable promise. Here, we present a novel approach that first extracts raindrop information from the surveillance camera images (rather than using the raw imagery directly), followed by the use of convolutional neural networks to estimate rainfall intensity from the resulting raindrop information. Evaluation of the approach on 12 rainfall events under both daytime and nighttime conditions shows that generalization ability, and especially nighttime predictive performance, is significantly improved. This represents an important step toward achieving real-time, high spatiotemporal resolution, measurement of urban rainfall at relatively low cost.

Process-based hydrological models seek to represent the dominant hydrological processes in a catchment. However, due to unavoidable incompleteness of knowledge, the construction of “fidelius” process-based models depends largely on expert judgment. We present a systematic approach that treats models as hierarchical assemblages of hypotheses (conservation principles, system architecture, process parameterization equations, and parameter specification), which enables investigating how the hierarchy of model development decisions impacts model fidelity. Each model development step provides information that progressively changes our uncertainty (increases, decreases, or alters) regarding the input-state-output behavior of the system. Following the principle of maximum entropy, we introduce the concept of “minimally restrictive process parameterization equations—MR-PPEs,” which enables us to enhance the flexibility with which system processes can be represented, and to thereby investigate the important role that the system architectural hypothesis (discretization of the system into subsystem elements) plays in determining model behavior. We illustrate and explore these concepts with synthetic and real-data studies, using models constructed from simple generic buckets as building blocks, thereby paving the way for more-detailed investigations using sophisticated process-based hydrological models. We also discuss how proposed MR-PPEs can bridge the gap between current process-based modeling and machine learning. Finally, we suggest the need for model calibration to evolve from a search over “parameter spaces” to a search over “function spaces.”.