River intermittency: mapping and upscaling of water occurrence using unmanned aerial vehicle, Random Forest and remote sensing landscape attributes

Abstract. Although intermittent rivers exist naturally, climate change along with changes in land use and occupation have a direct influence on streamflow permanence. Measurements and modelling in non-perennial rivers are still scarce and yet, essential for prediction and understanding of water scarcity scenarios. Thus, this work aims to map and model the spatio-temporal dynamics of an intermittent river. The study area is the Umbuzeiro River in the Brazilian Semiarid (∼ 100 km), whose spatially coherent streamflow occurs exclusively in the wettest months during the rainy season. We conducted twelve UAV surveys between March and November 2022 in selected river reaches. With the imagery from UAV surveys, we classified river reaches into “Wet”, “Transition”, “Dry” or “Not Determined” with visual inspection of 1.0 m reaches. In order to explain the observed patterns, we analysed 40 candidate predictors based on static and dynamic landscape attributes and grouped them into three Random Forest models based on the different source for dynamic predictors. Among these, altitude, drainage area, distance from dams, and one different dynamic predictor per model proved to be the most informative in Random Forest models. The selected models differ in the source and type of dynamic predictor used to capture the temporal dynamics: (a) series of Sentinel MNDWI; (b) series of Planetscope NDVI; and (c) antecedent precipitation index (30 d). All model variants successfully described river intermittency with an accuracy of around 80 % for both test and training datasets. Models (a) and (b) captured the temporal dynamics in model extrapolation to the whole river. When analysing the spatial distribution of flow intermittency, models (a) and (c) better identified areas more prone to “Wet” or “Transition” classes. This way, model (a) was identified as the most successful in simulating intermittency both temporally and spatially. The use of Sentinel MNDWI in model (a) aggregates enough spatial information, so the model can better simulate water occurrence classes. The findings presented here emphasize the possibility of using this index even in narrow non-perennial rivers, although its performance may vary depending on local hydrological and environmental conditions. The modelling framework developed in this study contributes to a broader understanding of flow intermittency as a spatially complex and highly dynamic process over time. The integration of high-resolution predictors demonstrates a scalable and adaptable approach for mapping wetness conditions in non-perennial rivers using landscape attributes, dam presence, and satellite indices as predictors.