Descriptor : High-resolution hydrometeorological data from a network of headwater catchments in the tropical Andes

This article presents a hydrometeorological dataset from a network of paired instrumented catchments, obtained by participatory monitoring through a partnership of academic and non-governmental institutions. The network consists of 28 headwater catchments (o20 km) covering three major biomes in 9 locations of the tropical Andes. The data consist of precipitation event records at 0.254mm resolution or finer, water level and streamflow time series at 5min intervals, data aggregations at hourly and daily scale, a set of hydrological indices derived from the daily time series, and catchment physiographic descriptors. The catchment network is designed to characterise the impacts of land-use and watershed interventions on the catchment hydrological response, with each catchment representing a typical land use and land cover practice within its location. As such, it aims to support evidence-based decision making on land management, in particular evaluating the effectiveness of catchment interventions, for which hydrometeorological data scarcity is a major bottleneck. The data will also be useful for broader research on Andean ecosystems, and their hydrology and meteorology.


Background & Summary
Tropical mountain environments host some of the most complex, dynamic, and diverse ecosystems 1,2 , but are under severe threats, ranging from local deforestation and soil degradation [3][4][5] to global climate change [6][7][8] . Despite their importance and vulnerability, there are very few observational data to quantify the impacts of these changes 9 . More extensive observations of climatic conditions at high elevations are urgently needed but monitoring in these areas is expensive and logistically challenging 10,11 . Despite many recent monitoring efforts to fill data gaps in high-altitude environments 12 , many mountain areas still lack a strong evidence base to guide decision-making on catchment interventions, such as re-or afforestation, and different conservation and climate adaptation measures. Driven by global agenda that promotes a better use of ecosystem services and natural capital, such interventions are happening at an accelerating pace, yet their hydrological benefits are insufficiently monitored and often based on anecdotal evidence. This is particularly the case for the tropical Andes, where the complex climatic and hydrological characteristics combined with a very dynamic anthropogenic disturbance put severe pressure on water resources 13 .
This combination of high vulnerability, accelerated change, and data scarcity makes it paramount to explore alternative methods of data collection. Among those, the use of participatory and grassroots approaches of information generation is gaining traction 14,15 . Participatory data collection has the potential to enhance public involvement and to fill long-standing data gaps [16][17][18] . It also has the potential to complement another major advance in hydrometeorological monitoring, remotely sensed data [19][20][21][22] , which are still limited by low spatial and temporal resolutions -usually kilometric areal averages and multi-daily-that impede hydrological assessment of small headwater catchments. Satellite precipitation products, for instance, are available at intra-daily timescales, but they involve large uncertainties and inaccuracies at the point-area difference with respect to surface level rain gauges 23,24 that difficult their direct assimilation.
As a response to such a monitoring gap, a partnership of academic and non-governmental institutions set up an initiative for participatory hydrological monitoring in the Andes 11,25 . The network known as Regional Initiative for Hydrological Monitoring of Andean Ecosystems (iMHEA) aims to produce information about the impacts of land-use and watershed interventions on hydrological ecosystem services, and to guide processes of decision making on catchment management. Similar bottom-up initiatives have emerged worldwide creating networks of well-monitored experimental sites 26 , for example, the network of terrestrial environmental observatories in Germany (TERENO) 27 . The iMHEA partnership, established in 2009, uses a hydrological design based on a 'trading-space-for-time' approach [28][29][30][31][32][33] . This concept relies on strengthening the statistical significance of a watershed intervention, such as a land-use signal, by monitoring several paired catchments in a regional setting 5 . The setup of the monitoring sites also allows for robust regionalisation results by covering different ecosystems with diverse physiographic characteristics and contrasting land uses and degrees of conservation/alteration 34 . This paper presents data from 28 catchments from the iMHEA network, representing 16 local stakeholders in 9 sites located in Ecuador, Peru, and Bolivia (Fig. 1). The presented data are composed of precipitation event records at 0.2 mm resolution, water level and streamflow time series at 5 min intervals, data aggregated at hourly and daily scale, as well as a set of hydrological indices derived from the daily time series, and catchment physiographic descriptors and geographic information.
The NGO CONDESAN and Imperial College London co-designed and implemented several iMHEA sites with local stakeholders as part of the Mountain-EVO research project (UK NERC grant NE-K010239-1) "Adaptive governance of mountain ecosystem services for poverty alleviation enabled by environmental virtual observatories" funded under the Ecosystem Services for Poverty Alleviation programme (www.espa.ac.uk) from 2013 to 2017. The project analysed how monitoring and knowledge generation of ecosystem services in mountain regions can be improved and used to support a process of adaptive, polycentric governance focused on poverty alleviation 16,17 . The data reported here fill longstanding hydrological information gaps allowing for evidence-based, robust predictions and cost-benefit comparisons about the effectiveness of watershed interventions. Such an approach has the potential to alleviate data scarcity in other remote mountain environments and provide relevant input for multi-scale decision making -from community level to national governance entities-ultimately improving the management of water resources and hydrological ecosystem services.

Hydrometeorological monitoring design
A monitoring protocol was designed to optimise the value of the resulting dataset to evaluate the impact of catchment interventions which is available in Ochoa-Tocachi et al. (2017, ref. 25). The protocol is based on a pairwise catchment comparison (Fig. 2) to separate the effects of changes in land use or watershed interventions from those of natural variability in climate and physiography [35][36][37] .
An idealised paired catchment experiment consists of two catchments that are collocated or in close proximity to each other, and whose physical and climatic characteristics are as similar as possible 38 . These catchments are monitored for a sufficient time to generate a statistical baseline for comparison. Subsequently, one of the catchments is subject to an intervention while the second remains in the original (reference) state. This makes it possible to evaluate the hydrological impact of the intervention, even if boundary conditions -such as climate-change over time. Although the use of a statistical baseline provides the most robust design, this is generally only achievable under strict experimental control of both catchments, which is often not practical. Therefore, a more straightforward 'control-intervention' setup (i.e. without baseline) is more commonly implemented instead, and has proven useful in published case studies 5,38,39 .
A major advantage of the 'control-intervention' setup is that impacts are evaluated by means of a spatial comparison rather than of changes over time, which reduces the length of the monitoring times, but this comes at the expense of a reduced attribution power. Therefore, a careful analysis and interpretation is needed to avoid any incorrect attribution of hydrological differences that may not be caused by the controlled factor but by other variables 34,40 . On the long-term, nonetheless, as the length of the monitoring time series increases 26 , individual catchments can be evaluated for impacts of climatic changes or vegetation physiology, especially in the evaluation of restoration strategies. The long-term sustainability of the monitoring using the paired catchment design can therefore reveal meaningful insights about hydrological non-stationarity. Such sustainability can be achieved by a robust institutional agreement as highlighted below in this article.
In each catchment, we monitored rainfall and streamflow, and measured or calculated several catchment physiographic characteristics (Data Citation 1). In extension, we applied a catchment water balance equation on the dataset of precipitation and discharge to estimate the combined losses, which account for an estimation of total water consumed by vegetation, evaporated from the surface, infiltrated to deeper soil strata, and stored in the soils. On the long-term, the water balance becomes a good approximation of catchment evapotranspiration. However, deep permeable soils and aquifers may be The monitoring sites cover the major high Andean biomes (Páramo, Jalca, and Puna) in 3 countries (Ecuador, Peru, and Bolivia). The complete iMHEA network consists of 16 sites, including one in the Venezuelan Andes. In this article, we present data from the 9 sites identified in the map by up-pointing triangles. present in some regions 3,41,42 which may challenge the assumption of a closed water balance. In such cases, a careful identification of the catchments has been performed to minimise errors and potential issues in the water balance simplification, or to monitor and consider those elements in the analyses.

Catchment physiography
Surveys of physical features were performed to select catchments and to document the properties that can influence their hydrological response. These properties were derived from geographic and cartographic data 34 , and divided in categories of shape (area, perimeter, compactness, shape factors, circularity, elongation, equivalent rectangle dimensions), drainage (longest drainage path, longest line parallel to the main stream, drainage density, current frequency, concentration time), elevation (maximum, minimum, mean, range, weighted average, percentiles), topography (hypsometric curve, catchment mean slope, stream mean slope, relief ratio, slope variability, topographic index), subsurface (soil types, soil depths, mean hydraulic conductivity, geology), land cover percentages, and land use types.
A summary of the main catchment characteristics is shown in Table 1. The monitoring sites cover three of the most extensive upper Andean biomes 5 , i.e. páramo, jalca, and puna (Figs 3a-c), with some catchments also covered by high Andean forests. Catchments are small, ranging from 0.5 km 2 to 20 km 2 , and located between 2000 m and 5000 m altitude. They feature a typical mountainous topography, with steep and uneven slopes, and oval shapes tending to circular or stretched. Sites are rural and the main land use types are conservation, grazing, afforestation, and cultivation (Figs 3d-f). Vegetation cover is generally natural and mostly consists of tussock and short grasses, shrubs patches with eventual occurrences of native and exotic forest, and small wetlands in the flat valley bottoms. Catchments classified as forested have at least 80% of their area under forest cover. No water abstractions or alterations are present upstream of the streamflow monitoring stations.

Precipitation
Rainfall is the main component of precipitation in Andean catchments 43 . Precipitation was measured using electronic tipping bucket rain gauges (Onset HOBO Data Logging Rain Gauge, Davis Instruments Rain Collector II, or Texas Electronics Collector Rain Gauge), installed at a height above the ground of 1.50 m. Their resolution is of at least 0.254 mm (0.1 in), but typically 0.2 mm, and exceptionally 0.1 mm. At least two rain gauges were located to improve capturing the catchment area and its elevation range, which is known to generate high spatial variability 44,45 . Furthermore, locations were selected for regional  representativeness and to minimise measurement errors such as those caused by wind effects [46][47][48] . Correlation analysis and double mass plots between the rain gauges were used to detect and correct errors and to fill data gaps. Automatic precipitation records were validated using manual rain gauges. In areas with potential snowfall occurrence, the equipment is above the maximum expected snow height accumulation in the ground, and an antifreeze substance has been recommended to melt ice and snow 25 . However, snowfall is rare within the monitored elevation range in the tropical Andes and has not been reported by local monitoring operators.
Tip time stamps were transformed to rainfall volumes depending on the rain gauge resolution. A composite cubic spline interpolation 49,50 was applied to the cumulative rainfall curve to generate rainfall intensities at 1 min resolution. This procedure generates better results than simple tip counting or linear interpolation 51 (Fig. 4). Rainfall data were then aggregated at time intervals matching those of the corresponding discharge in the catchment, and also at 1 hr and 1 day resolutions. From these data, we derived summarising indices including average annual precipitation, wetness index, days with zero rainfall, precipitation of the driest month, coefficient of variation in daily precipitation, and seasonality index 52 . For the calculation of median and maximum rainfall intensities, we used a 5 min scale moving window for storm durations between 5 min and 2 days. As most of the catchments lack a full meteorological station, the wetness index (ratio between annual precipitation and evapotranspiration) was calculated using WorldClim temperature data 53 Table 2. a JTU_04 is a large catchment containing the other three JTU catchments. b PAU_05, PIU_05, and PIU_06 are monitored for rainfall only and, as they lack streamflow data, their areas are not defined. c The discharge station in HUA_02 was initially located downstream from its current position defining an area of 2.71 km 2 ; data from this early period have been normalised during the data processing to use only the current area shown here. Global Water WL16, or Instrumentation Northwest AquiStar PT2X Smart Sensor). Water level was recorded at a regular interval of at least 15 min and generally 5 min, although some catchment data are limited to 30 min intervals. The measurement resolution ranges between 0.1 to 0.5 mm. The high temporal resolution of the recordings allows capturing the rapid hydrological response of small catchments after precipitation events, which can reach peak flows in only a few minutes (Fig. 4).
Compound sharp-crested weirs were used to rely on an established stage-discharge relation 47 . A compound weir combines a V-shaped section for accurate monitoring of minimum flows, and a triangular-rectangular section to contain peak flows. Manual flow velocity measurements were used to cross-check automatic water level records. The Kindsvater-Shen relation 56 was used to transform water level to discharge, and then revised with flow observations. Discharge data were normalised by catchment area (units of mm or l s -1 km -2 ) to enable comparison between catchments and water balance calculations with rainfall. These data were kept at the maximum temporal resolution and averaged at 1 hr and 1 day resolutions, from which a set of hydrological indices was computed.

Derived hydrological indices
The hydrological literature offers a large number of indices (also known as streamflow signatures) that characterise different features of the hydrological regime [57][58][59] . Indices are useful to represent several conservation/alteration conditions in catchments by means of streamflow information. Streamflow data are typically analysed in terms of five major response characteristics: (i) Magnitude (flow at any given time interval and location); (ii) Frequency (how often a flow above a threshold recurs over time); (iii) Duration (period of time over which a specific flow condition exists); (iv) Timing/Predictability (regularity with which a flow of a predefined magnitude occurs); and, (v) Rate of change/Flashiness (how quickly the flow magnitude changes). These characteristics can be subdivided further in low, average, and high flows.
We selected hydrological indices that are independent 37,60 , unambiguous 59,61 , and respond to the contextualised practical issue 34,62 , i.e., those that show evidence of change under watershed interventions, land use changes, or green infrastructure implementation. iMHEA partners are particularly interested in characterising water yield (flow statistics such as minimum, maximum, median, annual mean, long-term mean, driest month flow, and flow percentiles), hydrological regulation (baseflow index, recession constant, discharge range, slope of the flow duration curve, hydrological regulation index, Richards-Baker  flashiness index 63 ), and water balance (dry month discharge ratio, dry month discharge range, average annual discharge, runoff ratio). This list has been further expanded to include selected indices 34,59 of magnitude (skewness in daily flows, coefficients of variation, discharge ranges), high and low flow pulse analysis (magnitude, volume, frequency, and duration), timing (flood occurrence or absence, Julian day of extremes), and flashiness (flow reversals, logarithm of increasing and decreasing flow differences, ratio of days with increasing flow). We developed scripts to derive the several hydrological indices from normalised daily discharge data 34 .
One particular interest in the monitoring of small and seasonal catchments is the characterisation of low flow conditions. The Base Flow Index (BFI), calculated as the ratio between baseflow to total flow, is interpreted as the proportion of river discharge that originates from internal catchment stores 64 . We followed two different procedures to derive baseflow time series and the BFI. The first method follows the UK Flood Estimation Handbook 64 dividing the mean daily flow data into non-overlapping blocks of five days. The method calculates the minima of these consecutive periods and subsequently searches for turning points in this sequence. Daily values of baseflow are calculated by linear interpolation between these turning points and constrained by the original hydrograph when the separated hydrograph exceeds the observed. Baseflow time series are included in the daily flow dataset (Data Citation 1). The second method follows the Boughton two-parameter algorithm 65 with a filter parameter of 0.85 that is fitted subjectively. The twoparameter algorithm provides more consistent results than either a one-parameter algorithm based on the recession constant or than the IHACRES three-parameter algorithm 65 . We identified baseflow recessions as sections of the hydrograph of at least a 7 day duration that were apparently linear (R 2 >0.8) on a logarithmic scale for flows, from which the hydrograph recession constant was calculated. The presented hydrological indices include BFI and recession constants obtained by the two methods.

Institutional agreement for participatory monitoring
The monitoring sites implemented by the iMHEA network aim to achieve optimal complementarity with existing monitoring networks in the Andes 11 and with national authorities of water, environment, hydrology and meteorology 25 . iMHEA's grassroots approach is based on the assumption that civil society-based institutions can contribute with local scale and headwater monitoring to cover remote areas, thus complementing other efforts of data collection. The high spatiotemporal resolution of these data is compatible with the generally long-term and low-spatial density of national networks 17,47 .
The monitoring system includes an institutional agreement between local users of land and water, development organisations or local government offices, academic institutions, and other monitoring networks. Local communities and institutions are responsible for sensor installation, operation,  maintenance, and security and logistical tasks. Quality control and scientific rigour are achieved by research groups and universities who process and interpret the data. Such interaction, often understood as a form of citizen science [16][17][18] , has shown strong potential to tackle data scarcity in remote regions such as the tropical Andes. Observational data from experimental catchments have an essential value for hydrology and water resources management that increases with time 26 . As one of iMHEA's fundamental principle, this value is leveraged when information is shared at multiple levels for diverse audiences with the aim to inform decision-making processes more effectively. The network is interested in improving the management of water resources and hydrological ecosystem services in the Andean region, especially, to support the evaluation of the hydrological-and cost-effectiveness of watershed interventions. This institutional agreement provides an exemplary model to ensure the sustainability of monitoring networks in other contexts as well.

Code availability
The codes used to process the iMHEA data are in the form of freely available MATLAB scripts (Data Citation 1). Calculations were done using MATLAB R2017a (version 9.2) under an Academic License provided by Imperial College London. Revised or improved versions of the code and adapted scripts for R and Python are also available in a GitHub repository 66 .

Data Records
The list of the observatories used to generate the information in the present dataset is shown in Table 2.
In this data descriptor, we present data from 9 sites of the iMHEA network, which account for 28 catchments comprising 67 rain gauges and 25 streamflow stations. Each site has a particular research question being addressed by the monitoring which generally fits under three broad categories: hydrological impacts of human activities, hydrological benefits of watershed interventions, and characterisation of hydrological processes. A summary of the dataset and monitored variables is provided in Table 3.
The data are intended to support decision making at multiple levels, both addressing relevant local questions and supporting regional analyses of water resources in the Andes. The hydrometeorological data as presented in this paper are archived at the public repository Figshare (Data Citation 1). A backup of the data, including the original raw files obtained from each particular monitoring sensor and expansions with further data updates, can be requested to the iMHEA network via the corresponding authors.
The file format is CSV. The naming convention for the sensor data is: iMHEA_osite code>_ocatchment number>_ovariable>_osensor number>_raw.csv The naming convention for the compiled, processed catchment data is: iMHEA_osite code>_ocatchment number>_otemporal resolution>_processed.csv The naming convention for the catchment properties and indices is: iMHEA_Indices_ oindex type>.csv The naming convention for the geographic data is: iMHEA_osite code>_ocatchment number>_ogeodata type> The site codes are defined by the three letters shown in Table 2. The variables are identified by two letters, where the first indicates the type of station ('P' for pluviometric, 'H' for hydrological). The second the type of sensor, ('O' for Onset HOBO, 'D' for Davis Instruments, 'T' for Texas Electronics) for precipitation, and ('O' for Onset HOBO, 'S' for Solinst, 'D' Schlumberger Diver, 'W' for Global Water, and 'I' for Instrumentation Northwest) for discharge measurement. Numbers identify catchments and sensor locations and are used to organise the matrices of physical and climatic characteristics and hydrological indices (Table 3). Geographic data include catchment delineation, sensor location, contour lines, drainage network, lithology, land cover, and digital elevation model.

Regional representativeness and monitoring coverage
To maximise the regional representativeness of the tropical Andes and the potential for impact attribution, catchments were chosen to be hydrologically representative of a single land cover in the surrounding area 34 (Fig. 3). Large catchments usually host various ecosystems or land uses, and therefore observational records may not capture the hydrological signals of individual management practices or land cover. This would make it difficult to attribute a streamflow response to the effect of a specific catchment intervention. As a result, the monitored catchments here are smaller than 20 km 2 and host a homogeneous land use in at least 75% of their areas 25 . However, very small catchments ( o0.2 km 2 ) may have some experimental issues as well. For instance, in those with extensive wetland areas, subsurface water flow can become important 42 , increasing the risk of subsurface flow bypassing the flow gauging station. For this reason, our design avoids zero-order catchments 25 .
The number of stations necessary to approximate catchment rainfall input depends on the catchment area and the expected spatial rainfall variability therein 44 25 . This is not a problem of the data themselves, but of their spatial extrapolation. Additionally, it has been found that tipping bucket rain gauges may underestimate total precipitation by around 15% for low intensity events in high Andean catchments 43 . Therefore, most of the iMHEA monitoring sites are equipped with 3 rain gauges at representative low, middle and high elevation points within the catchments 5 , or at least 2 rain gauges in the smaller areas (Table 4 (available online only)). This also allows validation and correction of measurements if one of the sensors fails, for example, due to battery or memory issues or inlet obstructions.

Data collection and storage
As small mountain catchments feature a rapid hydrological response to precipitation events, the frequency of discharge data monitoring in the iMHEA sites is high (Table 4 (available online only)). The recommended minimum temporal interval for measuring water level is 15 min, but most of the stations record water levels every 5 minutes 25 . Tipping bucket rain gauges typically record data as tip time stamps (events), but a few stations use dataloggers that aggregate data at fixed intervals (Table 4 (available online only)). The exception is site CHA where rainfall and streamflow are both monitored at an interval of 30 min. The high temporal resolution allows identifying impacts on the short-term hydrological regulation, such as streamflow flashiness or hydrograph recession behaviour, that can be overlooked in aggregated indices 5 .
Although the most advanced sensors may have the capacity to store large amounts of data, batteries and data loggers have commonly a limited capacity to function for longer periods in harsh climatic conditions. These conditions and the remote locations reduce the possibility for regular sensor maintenance 25 . As a result, issues such as obstructions in the rain gauge orifices by forest litter or small animals may have occasionally occurred. To avoid the loss of important information, data are retrieved from each sensor approximately once a month.
Quality control of the generated data was performed immediately after the data have been collected in order to identify errors such as data outside the expected measurement range, negative water level values, or potential outliers such as extreme rainfall intensities. Data are curated thoroughly before their posterior processing and analysis as explained below. The high-resolution time series aggregated at different time intervals (5, 15, 60 min, or 1 day) represent a large amount of information in the long term (Figs 4a-c). However, it is imperative for impact assessment to store the finest resolution data to avoid losing important information that could be used in diverse studies and analyses 5,34 .

Precipitation data uncertainty, quality control, and processing
We identify four main uncertainty sources in the monitoring and processing of rainfall data (Table 5): (i) equipment malfunction uncertainty, (ii) measurement errors, (iii) rainfall intensity interpolation, and (iv) spatial interpolation.
Equipment malfunction uncertainty can be investigated by comparing data with and without quality control 67 . Our rainfall data are first curated by differentiating between rainfall tip time stamps and software communication and functioning logs. The latter time stamps are still kept in the raw data files but flagged as 'I' when the sensor is launched, 'D' when data have been downloaded from the sensor, and 'X' when software logs have been removed from the rainfall count. Potentially erroneous tips (e.g, as a result of manipulating the rain gauge) have been flagged with an 'X'. Other anomalies in the data, such as unrealistically high rainfall intensities, have been flagged in the raw files with the letter 'P'.  Measurement uncertainty includes random errors such as turbulent airflow around the gauge 48,51 . Systematic errors, for example, underestimation of real rainfall due to wind loss, wetting loss, and splashin/out, could theoretically be corrected when site-specific information is available 68 . We have found that the Onset HOBO rain gauges are prone to systematic anomaly, in which some time stamp intervals are exactly 1 s from the previous, which is the maximum time resolution of the data logger. Two tips of 0.2 mm with an interval of 1 s would represent a rainfall intensity of 720 mm hr À 1 . According to its manual 69 , the maximum rainfall intensity that this rain gauge can measure is 127 mm hr À 1 , and thus these immediate consecutive events where removed during the data depuration and processing. The standard deviation of the relative error in 1 hr rainfall rates can be approximated as a function of rainfall intensity 51 , which can then be used to estimate point measurement errors assuming a Gaussian distribution with zero mean 67 . By doing this, measurement uncertainty has been found to have only a small effect on the error standard deviation 67 .
Tipping bucket rain gauges aggregate rainfall data in the time domain, which complicates the estimation of rainfall intensities 68 . We use a composite cubic spline interpolation to generate a piecewise continuous curve that is thought to be the most accurate representation of rainfall rates 50,70 (Figs 4d and e), but this is based on several assumptions 43,[49][50][51] . Four threshold intensities were defined: (i) a minimum intensity to separate rainfall events, which depends on the rainfall structure of the studied region (here, 0.2 mm hr À 1 ) 43 ; (ii) a maximum intensity to merge consecutive tips, which is mainly to avoid issues when concatenating consecutive high-and low-intensity intervals (here, 127 mm hr À 1 ) 69 ; (iii) a mean intensity to distribute single tips due to the lack of information to interpolate them (here, 3 mm hr À 1 ) 50 ; and, (iv) a low threshold intensity above which data are accepted to avoid having extremely low rainfall rates (here, 0.1 mm hr À 1 ) 50 . To estimate the 'real' initial and final extremes of rainfall events, we assumed that they occur at half the rainfall rate of the two extreme rainfall tips and considered half partial tip remaining in the bucket from the previous event 49 . The first derivatives are then set to zero at these new defined extremes when applying the cubic spline interpolation. This improves the rainfall rate interpolation at the extremes of rainfall events and is otherwise equivalent to not estimating the initial and end points and setting the second derivatives to zero at the extremes.
Data were aggregated at 1 min resolution using tip counting before proceeding with the interpolation 50 . In the case that independent rainfall events had only 2 data points, a linear interpolation 51 was used instead. Mass-conservation and discontinuities were checked in the interpolated data for each independent rainfall event 50 : (i) when the estimated rainfall rates were negative (produced by a rapid change in the spline slopes between high and low rainfall rates), they were replaced by zero; (ii) rates that were between 0 and the low threshold intensity are replaced by this low threshold intensity; (iii) the resulting bias between the original and the interpolated rainfall totals was corrected by re-scaling the remaining intensities. Then, the cumulative rainfall curve was reconstructed by consecutively adding the corrected rainfall rates. In the rare cases when the bias after the interpolation was unacceptably large (here, >25%, although some studies use a tolerance of as much as 50% (ref. 50)), the cubic spline interpolation was discarded and a linear interpolation was applied instead.
Spatial interpolation errors affect the calculation of catchment average rainfall from the point measurements and depend on rain gauge density, network design, and rainfall spatial variability, which in turn is affected by topography, rainfall intensity, and storm type 44,67 . Interpolation errors in large catchments can be investigated by subsampling from the rain gauge network 67 . Here, we constrain spatial uncertainty by using at least two rain gauges in catchments much smaller than 20 km 2 , which is well above the WMO-recommended station density of 1 gauge 250 km À 2 for mountain environments 47 . Data gaps identified in the time series are flagged with a letter 'V' in the raw files, and monitored periods and percentages of data gaps for each sensor are shown in Table 4 (available online only). Rainfall data within each catchment were contrasted between the several rain gauges using double mass plots. We use the Units: water level (cm), streamflow (l s À 1 km À 2 ). Weir dimensions used to convert water level to discharge for each station in Table 6 Catchment characteristics  double mass plots also to fill data gaps in the rain gauges, but only when the regression coefficient (R 2 ) was larger than 0.99. To calculate catchment average rainfall, we used a simple average of rain gauge data within each catchment. Rainfall data between paired catchments were used to fill any remaining gaps, again only when R 2 >0.99.
Streamflow data uncertainty, quality control, and processing Monitoring and processing of streamflow data are affected by several sources of uncertainty (Table 5), from which we consider the most important to be: (i) uncertainty in the measurement of water level and discharge, (ii) rating curve calibration and extrapolation uncertainty (iii) time step used in the calculation of hydrological indices, (iv) seasonality in the hydrological response, and (v) hydrological index calculation method. Data collected using pressure transducers need temperature and barometric pressure compensation. Daily variations in surface water level have been observed in several studies [71][72][73][74] . While this phenomenon has been attributed to a set of confounding factors 75 , including evapotranspiration of riparian vegetation 76 , it may also be the consequence of an inappropriate monitoring design and operation 73 , for instance, external sensors, loggers, or cables directly exposed to solar radiation or precipitation. To minimise these issues, the barometric and submerged sensors of the absolute pressure transducer systems were installed in temperature conditions that were as similar and as buffered as possible. Vented tube sensors, in principle, do not require additional barometric compensation 74 ; nonetheless, when using them, we ensure the vented tube was free of obstructions, dry and isolated from direct solar radiation to avoid potential issues as recommended [77][78][79] . Uncertainty in the measurement of water level time series can be approximated by sampling from a uniform distribution using the nominal sensor accuracy as the estimated error (Table 6). Previous studies have used a uniform distribution with a mid-range value of ±5 mm (ref. 80) to investigate the propagation of uncertainty from discharge data to the calculation of hydrological indices 67,81 . Discharge uncertainty is typically larger than water level uncertainty and can be difficult to estimate especially during high-flow gaugings 67 .
Water level data were validated and corrected using manual observations during each site visit. The water depth over the weir was measured manually with 1 mm precision and then used to compensate differences with respect to the values collected by the automatic water level sensors. The same convention as for rainfall data was used for water level and discharge data, flagging data in the raw files (i.e., letters 'I', 'D', 'X', 'P', and 'V'). Water level data gaps were only filled for short periods in which the sensors were stopped for data retrieving, which were generally less than a few hours (Table 4 (available online only)). Other data gaps in water level or discharge data were not filled to avoid the manipulation of hydrological data that could result in erroneous attributions of streamflow responses. For instance, a non-linear impact of livestock grazing on the hydrological response time and flow recession curve can be observed in Fig. 4f. A linear interpolation of measurement gaps could affect this non-linear impact and potentially hinder the attribution of observed differences in the hydrological response.
The approximation of the true stage-discharge relation by means of a rating curve is usually a dominant source of uncertainty 67,82 , especially when this relationship changes over time. Processes that affect the validity of the rating curve include changes in roughness, vegetation growth, upstream sediment deposition and erosion, out-of-bank flow ranges, and unsteady flows or backwater effects 81,83 . In the monitored iMHEA catchments, flows are contained within a weir, which constrains the rating curve uncertainty and allows application of a theoretical relation derived from hydraulic principles (the Kindsvater-Shen equation 56 ). Only in a few catchments was this relation compared to direct flow observations. Table 6 shows the weir dimensions used for the Kindsvater-Shen equation in each

Variable
Uncertainty sources Investigation method Reference Precipitation Equipment malfunction Rainfall data with and without quality control 90 Point measurement error Normal distribution with standard deviation as a function of rain rate 48,51 Rainfall intensity interpolation method, rainfall event definition Composite cubic spline interpolation, threshold intensities, and bias correction 43,[49][50][51]70 Spatial interpolation uncertainty Subsampling from the available rain gauges in the network 67 Streamflow Water level measurement uncertainty, barometric and temperature pressure compensation Sampling from uniform or normal distributions with standard error of ±5mm or using sensor nominal accuracy 67,74,75,79,82,91 Rating curve uncertainty, calibration and extrapolation Constraining uncertainty by using a discharge control structure; voting point likelihood; heteroscedastic maximum likelihood model 56,81,83,85,92 Streamflow time discretisation Comparing indices calculated at different time steps 5,92 Seasonality in hydrological response Using total length of time series or divided by season 38,93 Hydrological index calculation method, e.g. baseflow separation method Comparing different algorithms and methods; Correlation analyses; regionalisation 34,59,64,65,81,87  catchment. Several iMHEA sites have started the task of deriving specific stage-discharge curves to improve the accuracy in the calculation of streamflow data from water levels. In the literature, different parameterisations have been tried, with polynomial functions performing slightly better than power functions 83 . A previous study investigating the use of stage-discharge rating curves in non-ideal conditions for short-term projects, in which measurements were generally gathered by non-specialists, found large uncertainties in the calculation of cumulative annual discharge between −13 and +14% (ref. 84). The extremes of the rating curve tend to be more uncertain because of the lower number of low and high flows available as calibration points. Previous studies on quantitative analysis of discharge estimation 67,81-83 have found that high flow extrapolation is the dominant source of uncertainty as the conditions that control the stage-discharge relationship can be different at water levels much higher than average. These errors can introduce much larger uncertainty than river flow measurements and seasonal changes in roughness. Additionally, being the non-linear least squares method the standard technique to calibrate rating curves, its lack of flexibility may introduce heteroscedasticity that needs to be accounted for 85 . In the case of unsteady flow occurrence, a dynamic rating curve approach can be used to provide an indirect discharge measurement based on simultaneous water level measurements at adjacent cross sections 86 , but this has not been implemented in our monitoring protocol 25 and is still not common in national level hydrometeorological monitoring elsewhere 47 .
Uncertainty in the calculation of hydrological indices results from the time discretisation of the flow data. This can be investigated by comparing indices calculated at different time steps (e.g. hourly and daily). For instance, using these hydrometeorological data to evaluate the effects of land-use changes, Ochoa-Tocachi et al. (2016, ref. 5) have found that impacts on the short-term hydrological regulation can remain unnoticed on daily aggregated indices but become evident when using high-resolution time series. Other sources of uncertainty arise from seasonal variations in the hydrological response that can provide evidence of catchment intervention impacts 38 or, in contrast, obscure statistical differences between catchments and between different monitoring periods 34 . The calculation method of particular hydrological indices can also introduce uncertainty in the assessment of catchment interventions. For instance, the baseflow index and its correspondent recession constant were calculated here using two different methods, one proposed by the UK Flood Estimation Handbook 64 and the other is a twoparameter algorithm 65   catchment data are pooled, for example in hydrological modelling or regionalisation, then the propagation of uncertainty from rainfall and streamflow data to any derived synthetic time series, model parameter, or hydrological index, as well as the uncertainty in the implemented methods and model structural errors, need to be considered and quantified 81,87 .

Usage Notes
The iMHEA dataset is intended to analyse the hydrological impacts of human activities in Andean catchments (e.g., land use change), to support the evaluation of hydrological benefits of watershed interventions (e.g., restoration strategies), or to increase understanding of hydrological processes of natural ecosystems and green infrastructure (e.g., water harvesting techniques). Recent applications of these data include evidence-based, robust predictions and cost-benefit comparisons of the effectiveness of different watershed interventions to support institutions in their ex-ante planning, implementation, and ex-post evaluation 5,88 . Such analyses can be extended to hydrological model calibration and regional extrapolation to make predictions in ungauged basins 34 . We do not consider the data to be suitable for trend analysis because of the short length of the time series. The value of the 'trading-space-for-time approach' relies on the comparison of climate conditions and hydrological responses between the paired catchments 5,38 . However, users need to be aware that catchments are unique 89 and their physical characteristics are inherently different, and thus land use and land cover are the main but not the only variance between them. An added value of the monitoring network design is the ability to compare and regionalise results by pooling the catchments together in a regional impact model 34 . In this case, users need to be cautious of the different monitoring periods of sites and catchments which may be influenced by particular regional hydrometeorological drivers.
The iMHEA sites are also available for the implementation of external research projects that can contribute to generate better and more detailed understanding of hydrological processes and impacts of human activities on water quantity and quality, sediment transport, and other ecosystem services. The long-term sustainability in the monitoring of experimental catchments will allow a deeper understanding of seasonality, natural variability, environmental changes, and extreme events such as drought and flooding 26 . The sites have been defined based on their socioeconomic relevance for local and regional stakeholders and can benefit from extending the monitoring to other variables or methods of interest. We invite researchers worldwide to make use of this dataset, and to engage in and complement the monitoring of these and new iMHEA sites.
The data are freely available under the Creative Commons Licence: CC BY 4.0.