Interpretable Machine Learning (IML) has rapidly advanced in recent years, offering new opportunities to improve our understanding of the complex Earth system. IML goes beyond conventional machine learning by not only making predictions but also seeking to elucidate the reasoning behind those predictions. The combination of predictive power and enhanced transparency makes IML a promising approach for uncovering relationships in data that may be overlooked by traditional analysis. Despite its potential, the broader implications for the field have yet to be fully appreciated. Meanwhile, the rapid proliferation of IML, still in its early stages, has been accompanied by instances of careless application. In response to these challenges, this paper focuses on how IML can effectively and appropriately aid geoscientists in advancing process understanding—areas that are often underexplored in more technical discussions of IML. Specifically, we identify pragmatic application scenarios for IML in typical geoscientific studies, such as quantifying relationships in specific contexts, generating hypotheses about potential mechanisms, and evaluating process‐based models. Moreover, we present a general and practical workflow for using IML to address specific research questions. In particular, we identify several critical and common pitfalls in the use of IML that can lead to misleading conclusions, and propose corresponding good practices. Our goal is to facilitate a broader, yet more careful and thoughtful integration of IML into Earth science research, positioning it as a valuable data science tool capable of enhancing our current understanding of the Earth system.

Flood hydrologic response is influenced by rainfall structure (i.e., variability in space and time). How this structure shapes flood frequency is unknown, and flood frequency analyses typically neglect or simplify potentially important aspects of rainfall variability. This study seeks to understand how rainfall structure impacts flood frequency. We use stochastic storm transposition combined with a 15‐year record of hourly, 4‐km² radar rainfall to generate 10,000 synthetic extreme rain events. These events are resampled into four “scenarios” with differing spatial and temporal resolutions, which are used as input to a distributed hydrologic model. Analysis of variance is used to identify the proportions of total flood peak variability attributable to spatial and to temporal rainfall variability under two antecedent soil moisture conditions. We simulate peak discharges for recurrence intervals of 2 to 500 years for 1,343 subwatersheds ranging in size from 16 to 4,400 km² in Turkey River in the Midwestern United States, which is situated in a typically humid continental climactic region. Antecedent soil moisture modulates the role of rainfall structure in simulated flood response, particularly for more frequent events and large watershed scales. Rainfall spatial structure is more important than temporal structure for drainage areas larger than approximately 2,000 km² (approximately 200 km²) for wet (dry) initial soil conditions, especially when soils are dry, while the reverse is true for smaller subwatersheds. The results appear to be related to the differing propensities for surface and subsurface runoff production as a function of basin scale, event magnitude, and soil saturation. Our results suggest that hydrologic model‐based flood frequency analyses, and particularly efforts attempting to spanning a range of scales, must carefully consider rainfall structure.

Estimating river flood risks under climate change is challenging, largely due to the interacting and combined influences of various flood-generating drivers. However, a more detailed quantitative analysis of such compounding effects and the implications of their interplay remains underexplored on a large scale. Here, we use explainable machine learning to disentangle compounding effects between drivers and quantify their importance for different flood magnitudes across thousands of catchments worldwide. Our findings demonstrate the ubiquity of compounding effects in many floods. Their importance often increases with flood magnitude, but the strength of this increase varies on the basis of catchment conditions. Traditional flood analysis might underestimate extreme flood hazards in catchments where the contribution of compounding effects strongly varies with flood magnitude. Overall, our study highlights the need to carefully incorporate compounding effects in flood risk assessment to improve estimates of extreme floods.

In order to provide urban flood early warning effectively, two support vector machine (SVM) models, using a numerical model as data producer, were developed to forecast the flood alert and the maximum flood depth, respectively. An application in the urban area of Jinlong River Basin, Hangzhou, China, showed the superiority of the proposed models. Statistical results based on the comparison between the results from SVM models and numerical model, proved that the SVM models could provide accurate forecasts for estimating the urban flood. For all the rainfall events tested with an identical desktop, the SVM models only took 2.1 milliseconds while the numerical model took 25 hours. Therefore, the SVM model demonstrates its potential as a valuable tool to improve emergency responses to alleviate the loss of lives and property due to urban flood.