Workflow follows the flowchart shown in Fig. . There are three main steps: (1) UAV data collection and mapping, (2) model training, and (3) application to the whole river. Step 1 encompasses the river course visual mapping. Then, we conduct UAV surveys multiple times in river reaches, producing high-resolution orthomosaics. We use the images to classify 1.0 m reaches in water occurrence classes, which are used as ground-truth data for model training and evaluation. In step 2, input data from predictor variables is used during the training of Random Forest classification models in order to obtain the same class as in the observed data (UAV-derived classification). Training includes the use of static and dynamic landscape attributes as candidate predictors to explain the observed patterns of water occurrence. In a recursive process, we compute the importance of the candidate predictors (or features). We test three models based on predictor subgroups and select the most important predictors in each model. The Random Forest models are retrained with the selected predictors in order to obtain the best models. In step 3, the models are applied to the entire river. Analyses were performed using custom R scripts and datasets archived on Zenodo .

Initially, data are collected in the monitored river reaches and characterized in terms of water occurrence (see Sect. ). Then these observations are used together with candidate predictors (Table ) from various sources in the training of Random Forest classification models. This kind of models take into account decision trees derived from resampling the calibration dataset . Each tree leads to a decision based on that particular calibration set, but a majority vote is taken on the final observations considering all trees. This way, dependence on calibration sets is potentially reduced. In the present study we implement the Random Forest classification model using the R package “randomForest” .

The importance of predictors is assessed by the increase in mean square errors during training when a predictor is randomly permuted within the tree. The “rfe” function (recursive feature elimination tool) in the R package “caret” selects the most suitable predictors . Recursive Feature Elimination (RFE) is an algorithm that applies a backward selection process to find the optimal combination of features and thereby selects the most important predictors .

Through the elimination of unimportant predictors, we test whether it is possible to use fewer predictors to model intermittency. For this purpose, models with different data sources are tested in three groups of candidate predictors: (a) Model using Sentinel predictors to capture temporal dynamics (spectral indices and LULC data); (b) Model using Planetscope indices as dynamic predictors; and (c) Model using hydrological data as the only source of temporal dynamics. The specific predictors employed in each model are summarized in .

Figure  shows the spatio-temporal evolution of the four classes in the monitored reaches. As expected, the “Wet” class is more present in the most downstream reach and its share diminishes towards the dry season. This class represents the puddles and pools formed along the river. The “Wet” class mapping provides us with information on the distribution and connection of these puddles. Figure  illustrates that there are longer, more connected patches in the downstream reach and that they take longer to dry.

The “Transition” class is present especially in reaches with algae or herbaceous vegetation which make it difficult to distinguish between “Wet” and “Dry” patches. Macrophytes are known to impact surface water detection from remote sensing imagery, as in . In our classification, the impact is observed due to the presence of low vegetation in transition zones between “Wet” and “Dry”. We use this as a new class, so we can observe its patterns. Given its nature of occurrence, it is a very heterogeneous class.

“Transition” is more present in the upstream reach alongside the “Not Determined” class. Reaches are classified as “Not Determined” especially due to dense tree canopies which restrict visibility of the riverbed. Since most of the Caatinga vegetation is deciduous, this biome is nearly leafless during the dry season. The denser and full canopy is why the “Not Determined” class can be found more frequently during the rainy season and in narrower river stretches.

The “Dry” class is more present during the dry season, since the riverbed dries out almost completely and becomes visible due to leafless tree canopies. The decreased leaf area index during the dry season has already been explored in hydrological studies about this the study area .

In general terms, we can say that “Wet” and “Dry” classes refer to open portions of the river, while “Transition” and “Not Determined” classes refer to areas with restricted visibility. The latter classes also represent narrower portions.

High-resolution UAV imagery enabled detailed visual classification of water occurrence classes. However, visual interpretation inherently brings uncertainty to the ground-truth generation. Ambiguous surfaces, such as distinguishing between “Wet” and “Transition” conditions, may introduce classification errors. To mitigate this uncertainty, ground-truth data was generated using the most frequent class within each modelling unit. This aggregation reduced spatial resolution of the analysis by summarizing 1 m classified segments into 100 m reaches. However, we also reduced the influence of local misclassification and increased robustness of the ground-truth representation. Field observations during the UAV campaigns also supported the interpretation of imagery and helped the classification.

Overall accuracy in relation to the number of predictors is shown in Fig. a for each model. Through recursive feature elimination, we can observe that there is almost no gain in using all candidate predictors. The five most important predictors are sufficient to accurately predict river intermittency for all three models. Often times, the use of all candidate predictors adds more noise than important information to our predictions, selecting the most important predictors removes that noise and maintains enough information to classify the models .

The top 10 most important predictors for each model are shown in Fig. c. Model training determines their specific powers. Figure c displays the mean decrease accuracy (MDA) of each and highlights the selected predictors for each model. Variables with higher MDA values were selected because they have a greater impact on the predictive performance of the model. MDA reflects how much the accuracy of the model decreases when a given variable is excluded; thus, a higher MDA indicates that the variable provides important information for distinguishing between classes. By selecting variables with the highest MDA, we focused on those that most strongly influence the model’s ability to predict spatial patterns of intermittency. This approach helps identify predictors most relevant to distinguishing our water occurrence classes. Among the not selected predictors, the observed values for mean decrease accuracy is around 20 %. This corroborates with the maximum observed overall accuracy shown in Fig. a of around 80 % and means that in general 20 % of accuracy is lost when less important predictors are removed.

The same four predictors were selected for all three models: mean elevation, drainage area, distance from last dam, and distance to next dam. The difference for each model was the dynamic predictors. It was selected one spectral index for each of the remote sensing models: model (a) Sentinel MNDWI and (b) Planetscope NDVI. For model (c), antecedent precipitation index (30 d) was selected. The similarities among models show consistency on the relative importance of predictors.

Model performance patterns align with known physical drivers of river intermittency. For example, river reaches with small contributing areas were more frequently classified as dry. Conversely, regions with bigger contributing area tend to sustain surface water for longer periods . This way, our models display the expected relationships between intermittency and landscape attributes (Fig. b); landscape attributes, such as mean altitude and maximum drainage area of the modelling units, are consistently selected as important predictors.

Catchment area or drainage area are sometimes identified as very important variables to classify streams as temporary or perennial . Altitude is another important variable to identify intermittent streams . Drainage area integrates upstream contributing catchment area and, by extension, potential water inputs to the channel network. Reaches with larger drainage areas are therefore more likely to sustain surface water for longer periods than headwater reaches with small contributing areas. This relationship is demonstrated in process-based studies of connectivity and its drivers in non-perennial rivers, which show that drainage area explains patterns and magnitude of flowing conditions . This is also observed during periods of river network contraction, when headwater streams are pruned in dry years .

Altitude may further modulate river intermittency through its influence on catchment morphology and hydrological processes. Because our study focuses on a single river, altitude shows a consistent downstream gradient and scales with drainage area, with lower elevations corresponding to larger contributing areas. This relationship may differ when comparing multiple river reaches at similar altitudes across the catchment, where contributing areas might vary. In our study, the observed patterns of water presence along the river network are consistent with the influence of both drainage area and altitude and likely reflect their role in controlling flow permanence.

We note that the “distance from the last dam” is the most important predictor in all models. The reason for this may be the fact that even if there is a dam further ahead (“distance to the next dam”), a reach may be dry or wet due to other factors. Immediately after dams, however, reaches have a higher probability of being dry even in more favourable conditions (e.g. lower portion of catchment and during the rainy season).

This consistent selection of “distance to the next dam” and “distance from the last dam” as top predictors highlights the strong influence of anthropogenic flow regulation on surface water distribution. While the relevance of dams in altering flow regimes is widely acknowledged, our findings provide a fine-scale perspective on how even small impoundments can generate abrupt shifts in flow permanence along the network . also showed that small dams individually and cumulatively alter lotic ecosystems. This is particularly relevant in semi-arid regions, where it is common to use small-scale storage and retention of surface water. However, as river networks become more intermittent due to climate-driven changes, this can be a reality in all climates .

Although predictors related to landscape attributes are well-established in hydrology, they gain new value here through their interaction with other spatial variables at finer resolution. For example, areas with high accumulation but located downstream of dams often remain dry, suggesting that topography alone does not explain surface wetness in human-modified catchments.

Dynamic predictors play an important role in observing temporal variation within a hydrological year. The spectral indices for each of the remote sensing models were previously employed to detect water. Sentinel MNDWI was used to study open-surface water in large water bodies and perennial rivers . Planetscope NDVI helped to successfully access LULC change detection and water class detection . This shows that characteristics captured by these indices are important to classify reach intermittency.

We also acknowledge that other important attributes may have been omitted from our analysis. For example, we did not consider the role of groundwater in river intermittency, despite it being a well-established control on water presence along the riverbed . In many catchments, groundwater levels tend to converge toward the river channel in downstream reaches, contributing to spatial gradients in intermittency that may not be directly observable from surface-based remote sensing alone . In our study area, detailed investigations at a small scale (12 km²) have shown that runoff is mainly governed by precipitation, with negligible baseflow due to the short and highly concentrated events . In contrast, analysis conducted at a broader scale (20 000 km²) have shown different roles of groundwater on streamflow depending on the period of hydrological season . However, groundwater-level observations, hydrogeological information, and coupled surface–subsurface modelling outputs were not available for our study. Future work would therefore benefit from integrating such data to better disentangle the respective roles of climatic forcing, catchment properties, and groundwater storage in shaping non-perennial river behaviour.

Water accumulation in intermittent river seems to be governed by processes captured by landscape attributes (including human-made interferences) and satellite indices. The consistent selection of these predictors across models reinforces them as controls and indicators to water retention.

Our analysis considered the distance of dams as a proxy for potential human influence on water occurrence. However, detailed information on dam characteristics, such as reservoir capacity, was not available. These characteristics likely modulate the magnitude and spatial extent of dam impacts on downstream water presence and should therefore be considered in future studies.

The out of bag errors (OOB) for all models are exhibited in Fig. b. We present OOB values for models (a), (b), and (c) trained with all candidate predictors and trained only with the selected variables shown in Fig. c. Among them, model (b) performs slightly worse, while models (a) with selected predictors and model (c) with all predictors achieved the best performances. These findings agree with Fig. a: both graphs make clear that the use of all predictors does not help to predict water occurrence classes in general. The benchmark classifier (BC) is based on a simple estimate of the dominant class from the overall dataset of observed classes. In our study, the “Not Determined” class is present in 43 % of observations, so the benchmark is 57 %. All trained models outperform the BC estimate.

Among the models, similar accuracy and balanced accuracy metrics are obtained (Table ). Balanced accuracy accounts for class imbalance by averaging sensitivity (true positive rate) and specificity (true negative rate), providing a more reliable assessment when classes are unevenly distributed. Values are reported separately for training and testing datasets. In general, test scores are slightly higher than training scores, maybe due the smaller dataset. Since the partitioning of the observation dataset is random, it might also be due to our subset selection.

The classes have different balanced accuracy scores; yet for all models, “Transition” is the most difficult class to predict. This result is expected since this class has the smallest number of observations. The “Wet” class receives the highest balanced accuracy scores of all models, followed by the “Dry” class. The overall performance of “Wet” and “Dry” classes can be explained by their more homogeneous observations, whereas more heterogeneous observations can be expected from the “Transition” or “Not Determined” class. For “Not Determined” reaches, we take into account the lack of riverbed visibility, but this condition includes very heterogeneous vegetation types, for example.

An example of model performance can be observed in detail in Fig. . In it, we use one of the monitored reaches to demonstrate the comparison between observed and predicted classes for each modelling unit. Overall accuracy was 80 % for train and 78 % for the test data for this model (Table ). Figures for other reaches and all models can be found in Appendix .

The models are applied to the whole river with all available data, enabling the evaluation of model performance in spatial and temporal distribution. The proportion of each class in the whole river is shown in Fig.  for each model showing the temporal distribution for each class. Temporal variations along the year can also be observed in comparison to monthly precipitation. Model (c) is the least sensitive to temporal variation specially in the “Wet” class. The “Dry” and “Not Determined” classes vary along the year, but not in a consistent way. February and March, for example, are two of the months with the highest proportion of “Dry”, and they are normally part of the rainy season .

In 2022, March is the rainiest month, still only a very small increase is observed for the “Wet” and “Transition” classes. The pattern of drying predicted by Model (c) may reflect lagged hydrological responses not captured by the antecedent precipitation index for that month. In contrast, the prediction for November coincided with isolated rainfall events that triggered surface water reappearance, which may be better captured by the antecedent rainfall.

Since Model (c) only uses rainfall data as dynamic predictor, it was expected to follow the seasonal behavior of water occurrence. However, the poor performance of this model in its application to the whole river indicates that it cannot extrapolate local-trained conditions considering only the landscape attributes together with precipitation.

For both models (a) and (b), there is a gradual decrease of “Wet” and “Not Determined” classes as the dry season progresses. While the “Wet” class gives information about the occurrence of puddles, the “Not Determined” class gives an idea of how dense the vegetation is. Since we are dealing with deciduous forest, a higher proportion of this class means that vegetation has leaves and is covering the riverbed. The “Dry” class proportion increases during the dry season.

The better plausibility of models (a) and (b) is due to more detailed and distributed information added to the models. This suggests that satellite-driven variables may better capture ecologically meaningful signals of intermittency, possibly due to their ability to represent spectral landscape responses. It is also possible to observe the complementary convergence of vegetation and water-related satellite indices and landscape attributes, which reinforce each other as physical controls (or responses) related to water retention.

Although the simplified model (c) performs satisfactorily in the test data, it does not provide a good temporal extrapolation of the river drying dynamic. This way, the dynamic predictor of model (c) (antecedent precipitation index for the last 30 d) is not sufficient to accompany the temporal river dynamics when training reaches are extrapolated. Other figures showing the spatial and temporal performance for all models can be found in Appendix .

The models are applied to the whole river, and the number of days with either “Wet” or “Transition” classes are calculated for each spatial unit: they indicate that there is “water in the riverbed” (WIR). In Fig. , it is possible to see the spots with more WIR-observations. Due to data availability, the highest possible numbers display different variations depending on the model: for models (a) and (b), it ranges from 0 to 12 or 13 observations, and for model (c) it differs from 0 to 365 observations with either of the two classes. Because of this, the number of observations is normalized and expressed as percentages. It is possible to observe that similar spots in all three models are prone to wetter conditions.

Regarding the spatial distribution of areas with higher WIR, the graphic shows to its extreme right that only models (a) and (c) identify the Benguê reservoir as a place with frequent “Wet” or “Transition” classes, for example. This is important because the reservoir is not part of the model training, but still a key factor to assess streamflow in the catchment. This way, models (a) and (c) outperform model (b), when they identify important reservoirs as areas prone to pounding. Model (a) is even more specific in this respect, and indicates mainly areas in the lowest part of the basin. The identification of areas prone to wetter conditions is very important even in the smallest of scales because they can be key areas for river ecology, for instance .

Analyzing our findings, we see that model (a) seems more plausible than the others, as it is able to extrapolate and produces smaller OOB errors. During training, model (b) performs worse than the others and generates greater OOB errors. It is also less conceivable than the others regarding spatial distribution of areas prone to wetter conditions. Model (c) presents small OOB errors and good accuracy for training and test datasets. However, the temporal analyses of its extrapolation to the whole river shows that this model can not predict river dynamics. Model (a), on the other hand, extrapolates well as was proved by both our temporal and spatial analysis.

Considering that model (a) modelled the drying and rewetting dynamics of the riverbed and spotted areas prone to water occurrence, it is shown as suitable for application on narrow rivers to observe classes of water occurrence. This way, we can model how intermittent is the river or how it changes from one year to another. The mapping of intermittency of water occurrence throughout the river in non-perennial rivers is important to analyse the migration and resilience of species, to understand their habitat and to conduct studies on intermittent river dynamics .

Evaluating our work, we begin by pointing out that we opt for a better observation quality versus quantity. Instead of mapping the entire riverbed (high quantity of observed areas) with UAV flights, for example, we decide to map and observe certain reaches in detail multiple times (better quality). This way, we add a higher time resolution and favour the ability to model temporal dynamics, instead of observing more areas. In addition, most of the extension of the river flows through private properties, so access is limited. Yet access is important because we need to be at an acceptable distance from the UAV platform during the flight. Take-off and landing should also be closer to the actual flight area. This way, we save battery and get better area coverage per flight.

In order to provide a general assessment of classification, the processing of each UAV image includes orthomosaic generation for the visual distribution of river reaches into intermittency classes. Through this orthomosaic generation we acquire one scene per monitored reach. Each reach is on average 5 km long and divided in 1.0 m reaches for classification. In total, 12 flights are performed and so we obtain a total of 60 000 reaches to be visually categorized for observation classes.

Although high-resolution UAV imagery enabled detailed visual classification of flow permanence classes, the method relied on expert interpretation without direct hydrological measurements such as streamflow or water level data. This introduces potential subjectivity, particularly in distinguishing between “Wet” and “Transition” conditions. Field observations were conducted during the UAV campaigns and helped inform classification, but were limited to qualitative assessments of ponded water, as no surface flow was observed during the study year.

We based our analysis on data from a single hydrological year. While this approach allowed for the capture of short-term hydroclimatic variability, it limits the model’s capacity to generalize across years with different rainfall patterns. Future studies including multi-year data and alternative temporal windows could help address this limitation.

Additionally, we acknowledge the potential value of integrating other dynamic hydrological variables, such as soil moisture and evapotranspiration, into predictive models of flow intermittency. These variables are relevant for ecohydrological modeling, particularly in dryland environments. However, in the context of this study, such data were unavailable at the spatial and temporal resolutions required to support fine-scale modeling.

The use of either wing-based or copter platforms have different advantages and disadvantages: while wing-based UAV has an average coverage of 150 ha per flight, the copter UAV mean coverage is only 20 ha per flight. For example, we cover the whole downstream reach with only two flights of the wing-based platform, whereas we need six flights for coverage of the same river reach with the copter platform. This is why we try to exclude unnecessary areas and turns as much as possible and capture only the riverbed area. We do our best to keep flight altitude stable between platforms at around 100 m. It could be increased in order to have faster coverage, but this decreases final ground resolution. Our wing-based platform also has multispectral cameras that would be useful for spectral indices based on UAV data; however, the lack of light uniformity during flights makes uniform reflectance correction and the use of spectral indices not possible.

Moreover, we use the spatial definition of the modelling units for both the response variable and most landscape-based predictors. Arguably, water occurrence in a river reach may be affected by landscape property beyond the 100 m radius. Furthermore, we do not take into consideration if non-overlapping units performed the same way or if the circular shape of the areas influences the outcome.

The results of our modelling approach is given as one class per modelling unit, i.e. categorical response variable. Instead, predicting the fraction of each class within the modelling unit would yield a more differentiated picture. What is more, implementing some proxies for spatial auto-correlation (e.g. intermittency state of neighbouring units) could potentially improve the spatial coherence of the predicted patterns.

As for our choice of predictors, we recognize that spatial resolution and ease of access vary greatly among them. The selected predictors are consistent among models; but great importance is given to dam identification. Both “distance from” and “distance to” dams ranked as highly important in model performance, and they already represent half the number of static predictors. In our study area, we identify many different types of damming structures – ranging from small rural weirs to larger reservoirs – future studies could benefit from classifying them into distinct functional or structural categories.

The manual mapping of small dams enabled a more realistic representation of water retention in the basin. However, dam mapping may also limit applicability of our model to other regions as it requires considerable manual effort and, potentially, familiarity with the study area. That said, the methodology is adaptable: similar analyses can be replicated using global datasets such as the Global Surface Water Explorer , which provides historical water occurrence based on Landsat imagery.

Finally, although Sentinel-2 data used in model (a) yielded strong results in our study area, the performance of its spectral indices – particularly MNDWI – may decline in more heavily forested or topographically complex catchments. In these cases, alternative data sources or higher resolution sensors may be needed to accurately identify surface water. However, it may be that dense vegetation can also serve as an indirect indicator of groundwater presence or surface wetness, and these segments were conservatively labeled as “Not Determined” in our classification.