Spatiotemporal assessment of event-scale rainfall erosivity (EsRE): a modelling approach based on artificial intelligence and satellite information

Soil and water conservation is a pressing global challenge, exacerbated by land use changes, cover transitions, and extreme rainfall events. Rainfall plays a pivotal role in soil erosion due to its ability to detach and transport soil particles. Understanding its spatial and temporal variability is critical for devising effective conservation strategies. Rainfall erosivity studies typically focus on the RUSLE R-factor to quantify monthly and annual trends. However, these broader timescales may overlook event-based dynamics captured by metrics like EI30, which considers storm kinetic energy and maximum 30-minute rainfall intensity. Such event-based analyses are especially important in regions with irregular rainfall patterns or increasing extreme weather events. Developing countries lack spatial representativeness of sub-hourly monitoring stations, demanding modelling strategies to address these data limitations. Relying only on the traditional R-factor may underestimate the soil loss quantification in regions experiencing intense land use transition (e.g., deforestation and urban expansion) and significant changes in rainfall patterns due to climate change. Such underestimations can have severe consequences in the management of river and reservoir silting, loss of agricultural land, and increased landslide risks. In this context, technological solutions based on Artificial Intelligence (AI) and satellite data present a promising alternative for addressing these challenges in countries like Brazil. In this study, we proposed a general framework to estimate the EsRE in a spatiotemporal fashion by leveraging the strengths of AI and the temporal and spatial availability of satellite-derived rainfall data, such as CHIRPS (available since 1981). The traditional and well-stablished event-based seasonal model is used as baseline model. As case study, both the AI-based model and the seasonal model used daily rainfall and the fortnight of the event as inputs to estimate the EsRE (retrieved from monitoring stations available since 2015 in the metropolitan region of Belo Horizonte, Brazil). The predictive performance of the models was evaluated using R², Nash-Sutcliffe Efficiency (NSE), and Pbias. Additionally, a bootstrap approach was employed to assess the uncertainty of the models. To evaluate the local applicability of the proposed model, we analyzed the impacts of the improved EsRE on the erosion process in the peri-urban Ibirité watershed, which faces intensified erosion and reservoir siltation. The AI-based model outperformed the traditional model, achieving R², NSE, and Pbias of 0.73, 0.73, and 2.4%, respectively, compared to R², NSE, and Pbias of 0.62, 0.62, and 0.13% for the traditional model. The observed uncertainty on EsRE simulation was expected due to the parsimonious model and problem complexity. Nevertheless, the AI-based model was able to track the overall spatiotemporal pattern, shedding light on the potential of this approach in modelling EsRE in poorly monitored regions. When replacing rainfall observations with CHIRPS-retrieved rainfall, uncertainty of EsRE increased, which was improved after bias correction. The bias correction addressed challenges related to intense rainfall events, such as convective storms, which are often underrepresented in satellite-derived data. The AI-based model improved the spatiotemporal analysis in the Ibirité watershed. Future studies should test other AI-based structures, input variables, and calibration strategies to decrease uncertainties on EsRE simulations.