River linking in India: Downstream impacts on water discharge and suspended sediment transport to deltas

To expand agricultural production and address water scarcity, India is moving forward with a large-scale civil engineering project to connect 44 rivers via a vast network of canals (Joshi, 2013; Bagla, 2014). The National River Linking Project, or NRLP, aims to increase irrigated area by 350,000 km² and improve food security and clean water access. India receives between 50 and 90% of its annual precipitation during the summer monsoon, during which time water is abundant and floods are common. The NRLP will store and redistribute this water in an effort to reduce temporal and spatial inconsistencies in supply. The project is intended to address the substantial challenges of food production and clean water access that India will face in the coming century – challenges that are being confronted globally as countries face rising temperatures, and increasing populations with stressed water supplies (Wallace, 2000; Battisti and Naylor, 2009; Hanjra and Qureshi, 2010). Interbasin water transfer systems are a common solution to water scarcity, and the NRLP is the largest of many new diversion schemes proposed or underway in China, Brazil, and Central Africa (Zhang, 2009; Lemoalle et al., 2012). However, large-scale river diversion projects such as the NRLP can result in far-reaching consequences for downstream river discharge and delta maintenance. For example, due to damming, diversions, and increased water usage, the Colorado, Nile, Indus and Yellow (Huanghe) rivers discharge little to no sediment today, whereas they previously accounted for 10% of the global sediment flux to the ocean (Syvitski and Milliman, 2007).

The NRLP in its current form was designed in the early 1980s, when the Indian government established the National Water Development Agency (NWDA) to manage implementation of the project. The project stalled between the 1980s and the 2010s, but it was renewed in 2012 following a Supreme Court order to proceed. The first canal was completed in September 2015; project completion is expected by mid-century. If fully realized, the NRLP will consist of 29 canals running a total of 9,600 km in length. Nine of the canals would each transfer more than 10 km³ y^–1 of water. This transfer capacity is greater than that of the current largest canal in the world (the Central Route of the South–North Water Transfer in China, which transfers 9.5 km³ y–¹; Zhang, 2009). The full system, depicted in Figure 1, would transfer fifty to one-hundred times the volume of the largest interbasin water transfer system in the United States and will likely constitute the largest construction project in human history.

Given the far-reaching social, economic, and environmental impacts of the NRLP, there have been innumerable calls for scientific research into nearly every aspect of the proposal (Bandyopadhyay and Perveen, 2003; Mizra, 2008; Amarasinghe and Sharma, 2008; Joshi, 2013). Sediment transport changes that may result from the NRLP have been highlighted as a significant knowledge gap (Bandyopadhyay and Perveen, 2008). Reduced sediment supply has been implicated in the sinking of many major deltas, putting vital agricultural regions, industrial areas, sea ports, diverse wetlands, and population megacenters at risk (Syvitski et al., 2009; Tessler et al., 2015). Deltas downstream of NRLP transfers include the Ganga-Brahmaputra, Mahanadi, Godavari, Krishna and Kaveri deltas. These five deltas together host a population of more than 160 million people (Higgins, 2015). All five deltas currently suffer from reduced aggradation that, together with land subsidence and global sea-level rise, place the deltas at risk of substantial elevation loss relative to sea level (Syvitski et al., 2009; ADB 2011; Tessler et al., 2015). It is therefore essential to establish how NRLP transfers may affect sediment transport to these vulnerable coastal areas.

Three fundamental processes resulting from the NRLP have the potential to affect sediment transport. The first is an increase in reservoir trapping, which has already reduced sediment transport globally by 30% (Vörösmarty et al., 2003) and in India by more than 70% (Gupta et al., 2012). New reservoirs associated with the NRLP will further impound sediment, particularly in the relatively pristine Himalayan rivers. A second impact is expected to result from the storage of high flows during the wet season and their slow release during the dry season. This will affect sediment transport due to the non-linear relationship between water discharge and sediment transport capacity. Finally, increased water utilization within the basin will decrease the average water discharge of rivers, reducing their year-round transport capacity, as sediment load is a function of water discharge for rivers that are not supply-limited (Syvitski et al., 2005; Walling, 2008).

Despite the importance of these processes, efforts to quantify their impacts within the NRLP have been severely hampered by a lack of data on the specific transfer volumes and operations of the project as well as the existing conditions in the rivers, particularly for the Himalayan (northern) half of the system. The Indian government restricts access to hydrologic data for the Ganga and Brahmaputra basins, and freely available data is sparse and incomplete. Prior to this work, even water discharge changes had not been tabulated for rivers affected by multiple canals.

This paper addresses these challenges and evaluates potential NRLP-derived changes to mean monthly water and sediment discharge and delta sedimentation for the Mahanadi, Godavari, Krishna, Kaveri, and Ganga-Brahmaputra deltas. Additionally, changes to mean annual water discharge are established for 29 rivers. Displaced population estimates are updated to the current project timeline, and changes to the existing watersheds of the subcontinental river network are mapped. This work is achieved via construction of the first comprehensive database of NRLP elements, which includes data for the full Himalayan component of the project. A novel graph database of the river network is developed to track and calculate flow. Finally, impacts of the NRLP are discussed within the context of human and environmental requirements and future climate change.

A total of 43 dams and barrages and 29 link canals are planned in the NRLP (Figure 1). Where possible, infrastructure data including locations and operating schedules for dams and canals have been taken directly from NWDA documents and reports (see annotated reference list, Supplementary Material). Table 1 gives the NWDA name for each canal. For the Peninsular component, NWDA summary statements and Feasibility Reports with operating schedules are available for canals 9, 10, and 16–29 and their associated dams, with the exception of a missing summary statement for link canal 29 and a missing Feasibility Report for link canal 25. The Feasibility Report for link canal 16 is unavailable due to recent negotiations that have moved the location of the primary reservoir and revised its operating schedules. A new summary statement has been released, but a new Feasibility Report is not yet available. For link canals 10, 28, and 29, NWDA Detailed Project Reports are available with additional information on the canals and associated dams. Newer NWDA publications were selected over older publications for inclusion in the dataset.

For the Himalayan (northern) components, NWDA data are much sparser. A pre-feasibility report is available for link canal 6 and an NWDA in-house bulletin gives some information regarding the locations of dams in link canal 1. Otherwise, data on the northern canals and dams have not been released. It was therefore necessary to supplement NWDA material with data from other sources. To that end, more than 500 documents from the public domain were reviewed for potential data on NRLP infrastructure. Seventy-five sources were ultimately included in the database, which together include the most current operating specifications for the components (see Supplementary Material). For the Himalayan canals, the most comprehensive source of information came in the form of three NWDA charts obtained by Gourdji (2005; 2008) that together depict the Himalayan canals’ full transfer, evaporation, utilization, and outfall volumes.

Exact latitudes and longitudes were available for approximately half of the dams and none of the link canals. Geo-location of each canal and structure was therefore facilitated by 30 NWDA schematic maps showing the approximate locations of dams, barrages, and water transfer structures in the project. These maps were digitized and georeferenced in a GIS framework (Figure 2). Country and state borders, water bodies, cities/place names, and existing structures provided geo-referencing points. Locations were then refined using text-based descriptions in the database sources. For example, no NWDA data is provided for the Chisapani and re-regulating dams associated with link canal 5, but a World Bank study of potential dam sites describes the recommended site:

Place names such as these were compared to Open Street Map, India-WRIS (Water Resources Information System of India) dam databases, Google Maps and DigitalGlobe/Google Earth satellite imagery, SANDRP publications, and news articles to refine structure locations and link canal routes. Where existing irrigation canals or smaller rivers were to be utilized along the link canal routes, these were extracted from Open Street Maps and joined with the routes.

For the 29 link canals, the following data were compiled: locations (GIS shapefiles of route traces), annual water transfer volumes, operating schedules, en route irrigation, en route domestic and industrial usage, and transmission losses (seepage and evaporation) (Table 2; and supplementary canals database). For the 43 dam structures, the following data were compiled: location (latitudes and longitudes), operating specifications (days or months of operation), full reservoir level (water elevation in meters), maximum (gross) storage, minimum (dead) storage, submergence area, forested submergence area, and estimated displaced population (Table 3 and supplementary dams database).

NWDA estimates of displaced population for most dams in the peninsular (southern) half of the NRLP are based on census data collected between 1986 and 1993 and are therefore out-of-date for a projected construction date of 2020 or later. For dams in the Himalayas and in much of the Ganga basin, no population displacement estimates are available. We therefore estimate displaced populations for each dam location as follows. For locations where population displacement estimates have been made public, we calculate a population growth rate factor from the Center for International Earth Science Information Network (CIESIN)’s (2016) Gridded Population of the World dataset, version 4. CIESIN population data are collected at the most detailed spatial resolution available, which for India is at the district level (10 – 100 km spatial scale). At each dam location, the CIESIN (2016) projected population density for the year 2020 is divided by the population density from the year 2000 to obtain the 20–year growth rate (Figure 3). The 20–year growth rate is then linearly scaled to the number of years between the reported census estimate and the year 2020 in order to estimate population displacement for a 2020 construction scenario. For locations where no population displacement was available, the 2020 population densities for the dams’ locations are used in conjunction with the submerged area to estimate displacement. Submerged areas were taken directly from NWDA documents and other sources (see supplementary databases for complete reference list). Uncertainty in the obtained displaced population estimates relates to both the uncertainty of the original census estimates as well as the uncertainty and consistency of the projected population density for 2020. However, we still regard this analysis as a significant improvement over the outdated reported data because of the relatively large population growth rates, which reflect an average population increase of 61% since the original census estimates.

Hydrological data is available for southern Indian rivers via the Central Water Commission (CWC) Integrated Hydrological Data Books. Published approximately every three years, the data books contain water discharge, sediment discharge, basin properties, and water quality metrics for most of the major river basins in India. These data, along with NWDA reports and peer-reviewed research on individual rivers, were used to establish current mean annual flows for most of the rivers in the Peninsular (southern) component of the NRLP.

Fewer data are available for rivers in the Himalayan component of the project. Data from the Ganga and Brahmaputra basins are classified by the Indian government and are not included in the Integrated Hydrological Data Books. Therefore, estimates from the World Bank-supported Strategic Basin Assessment of the Ganges Basin are used for the current mean annual discharge values for the tributaries of the Ganga. The Strategic Basin Assessment is a multi-year effort by the Institute of Water Modelling (IWM) in Bangladesh to develop a shared knowledge base and analytical framework for the Ganga basin. In the Strategic Basin Assessment, a precipitation-driven numerical model was used to produce estimates of mean annual discharge for the Ganga and all its major tributaries. These estimates include operating reservoirs and irrigation withdrawals throughout the basin (IWM, 2010). For the tributaries of the Brahmaputra, mean annual discharge values are compiled from individual published sources. For the main trunks of both the Brahmaputra and the Ganga, daily discharge data were obtained from the government of Bangladesh and processed to establish both monthly and annual mean flows. These data are described in the following section.

Mean monthly water discharge and suspended sediment data were obtained for six rivers leading into the five selected deltas. Gauging station locations for these datasets are given in Table 4. Monthly data was required for these rivers in order to assess the impacts of monsoon-season storage and dry-season transfers or releases. Given the non-linear relationship between water discharge and sediment transport capacity, reducing peak flows is expected to alter sediment transport capacity even if dry-season flows are increased proportionally.

Daily Ganga water discharge data from Hardinge Bridge and daily Brahmaputra discharge data from Bahadurabad Gauging Station were provided courtesy of the Regional Integrated Multi-Hazard Early Warning System for Africa and Asia (RIMES Project), originally from the Government of Bangladesh. Both gauging stations are located just inside the border with Bangladesh on their respective rivers, before the confluence of the two rivers. The data provided for the Ganga covered the period 1950–2011 but contains gaps in the dry season for later years. The period 1975–2000 is analyzed to provide the longest possible record with complete dry-season data beginning after the construction of Farakka Barrage, which was completed in 1975 and began diverting significant flows from the Ganga to the Hooghly for the purpose of flushing sediment near the port of Kolkata. The data provided for the Brahmaputra cover the period 1956–2011, also with gaps in recent years. The period 1975–2005 is analyzed to provide a 30–year record extending to the most recent possible year with complete dry-season data. Inter-annual standard deviations for each month are computed directly from the data.

Daily Godavari and Krishna discharge data from Polavaram and Vijayawada Gauging Stations were provided courtesy of the Office of the Chief Engineer, Krishna and Godavari Basin Organization. Both Polavaram and Vijayawada are the terminal gauging stations for these basins and are located just upstream of the deltas. As with the Ganga and Brahmaputra rivers, a record of at least 30 years is desirable for assessing true mean conditions in the presence of high inter-annual variability. For the Krishna river, however, a strong decreasing trend in discharge over the last 50 years renders long-term averages meaningless with respect to current conditions. Between the 1960s and today, mean annual water discharge of the Krishna fell from 62 km³ y^–1 to just 12 km³ y^–1, of which 95% comes during the monsoon season (Rao et al., 2010). The Krishna now runs dry before reaching the delta more than four months of the year. The period 2000–2015 was therefore selected to provide the longest possible period over which river discharge remained relatively stable at current conditions. Conversely, the Godavari river has shown no statistically significant change in mean annual water discharge over the last 50 years (Rao et al., 2010). A 30–year record covering the period 1980–2010 is analyzed for the Godavari. For both the Krishna and Godavari rivers, inter-annual standard deviations are computed directly from the data.

Mahanadi mean monthly water discharge was taken from Central Water Commission data at Tikarpara Gauging Station (CWC, 2015; Table 4). Tikarpara is the most downstream gauging station in the Mahanadi basin and is located approximately 90 km upstream from the delta apex. CWC (2015) values are a mean over the period 2000–2010. The shorter period of record was selected to minimize the influence of the Mahanadi’s trend in mean annual water discharge, which has decreased by approximately 0.3 km³ per year since the 1970s (Panda et al., 2013). Inter-annual standard deviations were not available from the CWC; instead, inter-seasonal standard deviations are taken from Panda et al. (2013). The seasons are pre-monsoon, monsoon, and post-monsoon and cover the period 1972–2005. Monsoon data in Panda et al. (2013) is further divided into June, July, August, and September. Standard deviations for these months were taken directly from Panda et al. (2013). For the remaining months, the pre-monsoon standard deviation from Panda et al. (2013) is divided by √5 for the five months January–May, and the post-monsoon standard deviation from Panda et al. (2013) is divided by √3 for the three months October–December.

For the Kaveri, monthly mean water discharge values from the Upper Anaicut (Mokkombu Dam) over the period 2001–2010 were obtained from the Asian Development Bank’s (2011) report. Upper Anaicut is located 2 km upstream from the apex of the Kaveri delta, where the river splits into multiple large distributary channels including the Kollidam. Inter-annual standard deviations are computed directly from the monthly data.

Suspended sediment concentration data were obtained for the selected six rivers. A sediment rating curve was fitted to the data with the form:

a widely-applied form for the relation between water discharge Q in m³/sec and suspended sediment discharge Q_s in kg/s, with two empirically-derived constants for individual rivers (Islam, 1999; Syvitski et al., 2005; Walling, 2008). 95% confidence intervals were calculated following Montgomery and Runger (2011) using a student’s t-distribution (2α) in conjunction with the standard deviation of Q_s and the sum of squares of deviations of data points from their sample mean.

For the Krishna and Godavari rivers, daily suspended sediment discharge data were available along with the daily water discharge data courtesy of the Office of the Chief Engineer, Krishna and Godavari Basin Organization. The same periods of analysis were selected for sediment as for water discharge. For the Ganga and Brahmaputra rivers, sediment discharge data and pre-fitted rating curves are given for Hardinge Bridge over the period 1980–1995 and Bahadurabad Gauging Station over the short period 1989–1994 in Islam et al. (1999). Mahanadi data from the Tikarpara Gauging Station over the period 2002–2012 and Kaveri data from the Musiri Gauging Station were obtained from CWC (2012, 2015). Musiri is approximately 12 km upstream of the Upper Anaicut gauging station from which water discharge data are presented. For both the Mahanadi and Kaveri rivers, suspended sediment measurements are available only as seasonal averages: monsoon, non-monsoon and annual. A sediment rating was fit to the seasonally averaged points.

Figure 4 shows the suspended sediment data and rating curves for each of the six rivers. The rating relationships obtained are as follows: Q_s = .0071Q^1.51 for the Ganga (R² = 0.68), Q_s = .0047Q^1.56 for the Brahmaputra (R² = 0.78), Q_s = .0006Q^1.71 for the Mahanadi (R² = 0.93), Q_s = .00008Q^1.96 for the Godavari (R² = 0.94), Q_s = .0019Q^1.40 for the Krishna (R² = 0.95), and Q_s = .0082Q^1.26 for the Kaveri (R² = 0.81).

Under the NRLP, suspended sediment discharge to deltas will be affected by changes in river water discharge and by increased sediment trapping in new reservoirs. The exact change to water discharge will depend upon canal operating schedules, utilization changes within the basin, and transmission losses (seepage and evaporation) within the canals. The new water flows will then convey new suspended loads as approximated by the individual sediment rating curve estimated from observed data for each river. Reservoir trapping will apply in the new regime. The following sections describe how we model these processes.

To calculate water discharge changes resulting from the NRLP, a graph database was constructed from the text-based datasets (Figure 5). The database handles only those changes associated with the NRLP; we assume current regulations and transfers by existing infrastructure are reflected in gauging station data. In the graph database, canals are represented as edges, with properties of flow direction, annual transfer volume, and en route usage and loss values compiled from primary sources. Confluences, dams, barrages, and canal outfall locations are represented as nodes. Each dam node is controlled by an impoundment schedule, which removes water from the river, and a transfer schedule, which releases monthly portions of the total annual transfer volume into the canal. Losses within the canals follow the same monthly proportions. Transfers from each dam can take one of three schedules: uniform (regular transfers throughout the year); specified (specific transfer volumes per month specified by NWDA reports); or monsoon (no transfers during the monsoon months, uniform transfers during the dry season months). Impoundment schedules are the same for each dam: zero impoundment in the dry season (December–June), with the full transfer volume impounded during the remaining months. In the Krishna basin only, it was necessary to increase dry-season impoundment to 5% of the full transfer volume per month for all simulations to meet NWDA-specified transfer volumes. In all other cases, operating schedules were derived from NWDA sources and reports.

Although the graph database is not a hydrological model, it has several advantages for performing water transfer calculations. Given flow directions and operating parameters, it is capable of handling the massive complexity of a looping, interconnected, and temporally staggered system of this size. The database can be queried to examine exposure and connectivity along individual reaches. The free software can be modified by any user, with subsets of the project simulated, operating parameters modified, and new dams or canals added as the project evolves.

The graph database was used to calculate changes to mean monthly water discharge at the apex of each of the six deltas of interest. Next, sediment rating curves were applied to mean monthly water discharge values under current conditions and under the NRLP (Figure 4). Change in annual suspended sediment load was then calculated from the monthly totals. In this procedure, we make the simplifying assumption that no siltation occurs within the canals. Uncertainty calculations for each part of the sediment-change calculations are described in the section “Quantifying uncertainty,” below.

Reservoir trapping is calculated following the method of Vörösmarty et al. (2003), which uses a modified version of Brune’s (1953) empirically-derived rule for determining the trapping efficiency of reservoirs. Trapping efficiency quantifies the proportion of incoming sediment permanently retained within a reservoir. Vörösmarty et al.’s (2003) method allows an estimate of basin-wide trapping efficiency for basins with multiple reservoirs. It is applied here only for new reservoirs associated with the NRLP, and we assume that existing reservoir trapping is reflected in the current sediment rating curves.

For a single basin divided into m regulated sub-basins, the approximate residence time of water in the reservoirs of regulated sub-basin j is given by:

where n is the number of reservoirs in sub-basin j, V_i is the operational volume at reservoir i, and Q_j is the water discharge at the mouth of sub-basin j. The trapping efficiency for regulated sub-basin j is then given by

The trapping efficiency of the full basin is given by

where Q_m is the water discharge at the river mouth (Vörösmarty et al., 2003).

For the Krishna, Godavari and Mahanadi rivers, all newly planned reservoirs are located on the main stems, with the most downstream reservoir located near or at the delta apex. Each of these rivers was therefore treated as one regulated sub-basin, with reservoir volumes V_i given in Table 3, mean annual river discharge Q_j taken at the river mouth, and basin-wide trapping efficiency for NRLP reservoirs given by Eq. 4. Reservoir volumes V_i are multiplied by a utilization factor of 0.33 to approximate annual average storage, assuming maximum storage occurs only in monsoon months. For the Kaveri river, no new reservoirs are proposed in the NRLP.

For the Brahmaputra basin, there are two regulated sub-basins associated with the NRLP: the Sankosh and the Manas. Mean annual water discharge Q_j for the Sankosh and Manas are 17.3 and 32.0 km³ y^–1, respectively. For the Ganga basin, there are seven regulated sub-basins associated with the NRLP: the Kosi, Gandak, Ghaghara, Sarda, Ken and Son, each with one reservoir, and the Chambal with three reservoirs. Barrages and re-regulating dams were not included in trapping efficiency calculations. Mean annual water discharge Q_j at the dam sites are derived from primary references and are 49.9 km³ y^–1 for the Kosi, 32.7 km³ y^–1 for the Gandak, 43.8 km³ y^–1 for the Ghaghara, 13.6 km³ y^–1 for the Sarda, 12 km³ y^–1 for the Ken, and 13.2 km³ y^–1 for the Kosi (see supplementary dams database and annotated reference list for references.) For the Chambal, mean annual water discharge is 18 km³ y^–1. Mean annual water discharge for the full Brahmaputra and Ganga basins at their river mouths (Q_m) are 682.6 km³ y^–1 and 356.0 km³ y^–1, respectively. As with the previous rivers, reservoir volumes are multiplied by a utilization factor of 0.33.

From these values, basin-wide trapping efficiencies associated with new NRLP reservoirs (TE_bas) were established (Eq. 5). We note that these values do not represent the full trapping efficiencies within the basin, but only the additional trapping related to the construction of NRLP reservoirs. We make the assumption that trapping by existing reservoirs is already represented within the sediment rating curve. Trapping proportions were applied to the rating-curve derived suspended sediment changes to arrive at a total change in suspended sediment load for each river entering a delta.

Uncertainty for the trapping efficiency of each reservoir was derived from the bounding envelopes of Brune (1953), in which trapping efficiency is shown to vary by approximately ±10% from a de-silting to a semi-dry reservoir, or from a coarse-sediment to a fine-sediment reservoir, as highlighted in Verstraeten and Poeson (2000). It was conservatively assumed that errors in trapping efficiency measurements are not independent; a worst-case scenario estimate of trapping uncertainty was made by propagating a ±10% uncertainty for each reservoir through the basin-wide calculation.

In calculating uncertainty on sediment load changes resulting from the NRLP, we assume full construction of the NRLP as proposed. We assume that canals do not fundamentally alter the sediment rating curves once emplaced (beyond trapping efficiency changes from new reservoirs), and we further assume that no sediment storage occurs within the canals. We estimate sediment load changes for current mean water discharge conditions only. The change in sediment load brought about by the NRLP must be calculated independently for each flow condition, as sediment load scales non-linearly with water discharge. For a given flow condition, the uncertainty on the magnitude of the sediment-load shift from current to NRLP conditions comes from two sources: (1) uncertainty on the Q-Qs rating curve relationship (Figure 4), and (2) uncertainty on the trapping efficiency of new NRLP reservoirs.

For each river in Figure 4, the best-fit Q-Q_s relation is given, along with an envelope encompassing the range of possible Q-Q_s relations with 95% confidence. Within this envelope, the highest-exponent and lowest-exponent relation encompassing the data within uncertainty bounds can be found (i.e., highest-exponent in linear space, or highest-slope in log-log space as depicted in Figure 4.) These relations give the highest and lowest sensitivity to water discharge changes, where the highest-slope relation is the most sensitive and the lowest-slope relation the least sensitive to changes in water discharge. For each scenario, a 95% confidence interval for trapping efficiency is then applied to the NRLP load estimates. Both sensitivity scenarios and trapping efficiency uncertainty are included when giving 95% confidence intervals on the percent change in total annual load.

Change in total annual suspended sediment load for each delta was converted to potential change in delta aggradation rate using delta area and an assumed sediment bulk density of = 1.5 g cm^–3. In this approximation, one-third of the total suspended sediment load is deposited on the subaerial delta, with two-thirds transported to the subaqueous or to the ocean. This proportion is considered reasonable for river systems with extensive monsoonal flooding and has been corroborated by long-term sediment budget reconstructions for the Ganga-Brahmaputra Delta and other large Asian river systems (Islam et al., 1999; Goodbred and Kuehl, 1999; Liu et al., 2009). Change in aggradation rate for each delta was therefore calculated as follows:

where ΔD is change in delta aggradation rate (mm y^–1), A is delta area (km²), ρ is sediment bulk density (g cm^–3), 1/3 is the fraction of sediment load deposited on the subaerial delta surface, m is month of year, s_m is seconds per month, Q_{s,m current} and Q_{s,m NRLP} are mean monthly suspended sediment discharge values for month m under current and NRLP conditions (kg s^–1), and 10^–6 is a unit conversion factor with units of (g mm km² kg^–1 cm^–3). Subaerial delta areas are based on the delta boundaries defined by Tessler et al. (2015). Area for the Ganga-Brahmaputra Delta includes the embanked tidal deltaplain, which is not subject to aggradation except in the case of levee breaches (e.g. Auerbach et al., 2015). Insufficient data were available to calculate changes to bed load transport or the bed load contribution to delta aggradation.

Changes to existing watersheds were assessed for six major rivers: the Ganga at Hardinge Bridge, and the Hooghly, Godavari, Krishna, Penna, and Kaveri Rivers at their mouths. The Mahanadi and the Brahmaputra were not included, as they are ‘donors’ to other rivers only. New contributing runoff regions for the other rivers were delineated using the Terrain Analysis Using Digital Elevation Models (TauDEM) package applied to the HYDRO1K dataset, a U.S. Geological Survey hydrologically-corrected 30 arc-second digital elevation model of the world. Watersheds were delineated for the offtake points of each canal and assigned as new contributing runoff regions for the rivers into which the canals outfall. The graph database described previously was used to evaluate connectivity of the network and determine exposure to new runoff regions.

Next, the percent of water exiting the basin that will be sourced from the new contributing areas was calculated. In most cases, unless operating schedules suggested otherwise, it was conservatively assumed that water entering a reservoir did not mix with the existing river water if it was to be immediately withdrawn for further transport down the system. Where water offtakes occur downstream of water deliveries, it was assumed that new and old water were removed proportionally.

An example of the calculation for the Godavari River is illustrated in Figure 5. We assume that the Godavari starts with its current mean annual water discharge of 89.7 km³ y^–1, as all canals meet the river downstream of its confluences with major tributaries. The river loses 16.4 km³ y^–1 to link canal 19, 4.4 km³ y^–1 to link canal 18, and 5.3 km³ y^–1 to link canal 17, making its discharge just 63.6 km³ y^–1 as it nears the river mouth. At the delta, however, link canal 16 outfalls at Polavaram, carrying 18 km³ y^–1 sourced from the Mahanadi, the Ganga and tributaries of the Brahmaputra via Farakka Barrage. The total discharge jumps back to 81.6 km³ y^–1, making the change in mean annual discharge at the Godavari only 8 km³ y^–1, or 10% of flow. However, 22% of flow (18 km³ y^–1) is now sourced from the watersheds of the Mahanadi, Ganga and Brahmaputra tributaries.

Figure 6 shows changes to mean annual water discharge for 29 rivers. The Ganga is projected to experience the largest reduction in absolute water discharge, primarily due to intra-basin transfers and usage. The discharge of the Ganga at Hardinge Bridge is projected to decrease by 87 km³ y^–1 (–24% of total annual flow), of which approximately 12 km³ is transferred out of the basin via link canals 6, 7, and 8, and 75 km³ is utilized or lost to evaporation within the Ganga basin. Two tributaries within the Ganga are projected to increase in mean annual discharge: the Betwa river by 3% and the Yamuna by 31%. All other tributaries are projected to decrease in mean annual discharge, with the largest decreases seen in the Kosi (23%), Ghaghara (55%) and Gandak (68%) Rivers.

The Brahmaputra, which will contribute the largest volume of water to other river basins via withdrawals from its tributaries, is projected to experience a decrease of 43 km³ y^–1 (–6%), of which 5 km³ will be used or lost within the basin and 38 km³ will be transferred to southern basins. These transfers will reduce the mean annual water discharge of the Brahmaputra by only 6% but will decimate the tributaries themselves: the Manas and Sankosh are projected to lose more than 70% of their mean annual water discharge and the Raidak more than 50%, with the Torsa and the Jhaldak losing 16% and 20%, respectively (Table 5).

Southern rivers of India are the primary beneficiaries of the NRLP; however, only the Hooghly, Penna and Kaveri are projected to experience an increase in mean annual water discharge given full implementaton of the project. The Hooghly is projected to experience an increase of 22%, the Penna of 450%, and the Kaveri 33% over their current mean annual discharge. The remainder of the southern rivers are projected to either remain stable or decrease, with the largest percent decreases occuring in the relatively smaller western rivers: the Achankovil (–56%), Damanganga (–28%), Par (–32%), Auranga (–18%), Ambica (–24%), and Purna (–32%).

Mean monthly water discharge under current and NRLP conditions was established for six rivers that feed into five major deltas: the Ganga, Brahmaputra, Mahanadi, Godavari, Krishna, and Kaveri (Figure 7). The same NRLP-derived water discharge changes are shown against mean conditions (black curves), with one standard deviation of inter-annual variability given in blue (high-flow) and red (low-flow).

Though the rivers vary in mean annual discharge by up to two orders of magnitude, their hydrographs follow a similar pattern of low flow during the dry season and high flow during the monsoon-dominated summer months. The arrival of monsoon discharge is delayed by approximately one month for southern rivers relative to northern rivers. Inter-annual variability is high for all rivers, reflecting the strong variability in the strength of the monsoon. However, inter-annual variability is particularly high for the Krishna and Kaveri rivers. This is not entirely due to natural processes; it is partly caused by strongly varying irrigation withdrawals over the last decade. Data in this study show that both rivers ran dry for most (Krishna) or all (Kaveri) of the dry-season months over the periods of record.

Under the NRLP (dashed lines, Figure 7), monsoon-season discharge is reduced for all but the Kaveri river. This is primarily due to water impoundment in reservoirs. In general, changes to mean monthly flow will be on the order of those already seen in current inter-annual variability. However, particularly for the Ganga, future high-flow years will become similar to current mean-flow years; future mean-flow years will become similar to current low-flow years; and, if NRLP withdrawal schedules are not modified during low-flow years, future low-flow years will see unprecedented lows that fall outside of the current envelope of natural variability.

For the Ganga, water discharge will decrease from July–December, with the largest decrease occurring between July and October as water is impounded in Himalayan dams. The Brahmaputra will experience a similar decrease in water discharge from July to November, with its flow unaffected the rest of the year. As the water from the Brahmaputra is transferred to southern rivers during the dry season, some rivers will see an increase in water discharge during those months: the Kaveri will improve from no-flow to scant flow (51 m³ s^–1) during the dry season, and the Krishna will see its mean monthly water discharge increase from approximately 50 m³ s^–1 to approximately 250 m³ s^–1 for most of the dry-season months This will bring dry-season water discharge back to pre-1970s levels (prior to major dam construction) (GRDC, 2017), though with sediment trapping by post-1970s dams now in place (Rao et al., 2013).

Figure 8 shows the corresponding changes to monthly suspended sediment discharge calculated using the sediment ratings curves for individual rivers (Figure 4) and the estimated trapping efficiency of new reservoirs. Black lines in Figure 8 give results using the best-fit rating curve, while orange and blue lines give a 95% confidence interval using the highest- and lowest-sensitivity rating curves that encompass the data. Estimated trapping efficiency increases are 36% for the Ganga, 6% for the Brahmaputra, 45% for the Mahanadi, 79% for the Godavari, 74% for the Krishna, and 0% for the Kaveri, where no new reservoirs are proposed. We note that these values do not represent the full trapping efficiencies of the basin, but only the additional trapping related to proposed NRLP reservoirs. Shading on estimated NRLP conditions is derived from uncertainty on trapping efficiency changes.

As with water discharge, the Ganga is projected to experience a year-round drop in its suspended load, with peak sediment discharge falling from 58,000 kg s^–1 to 23,000 kg s^–1 in the month of September. Similarly, peak suspended sediment discharge in the Brahmaputra will drop in same month from 97,000 kg s^–1 to 78,000 kg s^–1; in the Mahanadi from 1,700 kg s^–1 to 550 kg s^–1, in the Godavari from 8,600 kg s^–1 to 1200 kg s^–1; and in the Krishna from 50 kg s^–1 to 5 kg s^–1 despite that river experiencing no net change in mean annual water discharge. In the Kaveri, sediment discharge will increase year-round, but by a negligible amount.

Total mean annual suspended sediment load for each delta is available in Table 6. Current conditions (Q_s) are calculated from available data, and NRLP conditions are simulated. Given full NRLP implementation, all rivers but the Kaveri are projected to experience a decrease in mean annual sediment load. With large reservoirs planned near or at their deltas, the Mahanadi, Godavari, and Krishna Rivers are projected to experience reductions of 40–85%, 71–99% and 60–97% of their annual suspended loads, respectively. The Ganga is projected to experience a 39–75% reduction in annual suspended load, and the Brahmaputra a 9–25% reduction despite losing only 6% of its annual water flow. For the Ganga, Brahmaputra, and Mahanadi Rivers, approximately two-thirds of the decrease in mean annual sediment load is due to the redistribution of flows from the monsoon to the dry season, with the remaining one-third due to trapping in reservoirs. For the Godavari and Krishna rivers, approximately half of the reduction is due to flow redistribution and half to trapping. The Kaveri is projected to experience a 7–43% increase over current conditions; however, the current suspended load is near zero, so this will have little, if any, impact on the delta.

R² values for fitted Q-Q_s rating curves range from 0.69–0.95 and are lowest for the Ganga and Brahmaputra Deltas, reflecting current uncertain knowledge of the sediment loads of those rivers as presented in the literature. Darby et al. (2015) notes that sediment load estimates for the Ganga range from 390–548 Mt/y, while Fischer et al. (2016) finds a suspended load of 260–720 Mt/y. For the Brahmaputra, Darby et al. (2015) notes a best estimate of approximately 475–523 Mt/y but acknowledges published values as high as 1100 Mt/y. The large uncertainties for suspended loads highlight the need for ongoing data collection on these vital international rivers.

At deltas, reduced suspended loads will translate to reduced aggradation rates, putting them at greater risk of coastal erosion and flooding (Table 6). For the Ganga-Brahmaputra delta, the aggradation rate is projected to drop from 3.6 to 2.5 mm y^–1 given full implementation of the NRLP, a 30% reduction in elevation-building capacity. Aggradation on the Mahanadi delta would fall to half of its current rate, and on the Godavari to just 13%. These values do not include bed load, which is typically ~10% of the suspended load and contributes additional aggradation capacity (Syvitski et al., 2003). However, little data is available to constrain bed load estimates. Both the Krishna and the Kaveri deltas, which currently receive negligible sediment loads, would continue to experience negligible aggradation under the NRLP.

Figure 9 shows the changes in watersheds for six major basins and the proportion of water that will be sourced from new areas. Given full NRLP implementation, 12% of the water discharge at Hardinge Bridge on the Ganga will be sourced from a new watershed, along with 3% at the mouth of the Hooghly. It is worth noting that more than half of the water in the Hooghly is already transferred from the Ganga via Farakka Barrage. While the NRLP will result in only 3% of the Hooghly discharge coming from the new watersheds of the Brahmaputra tributaries, it will represent an increase from 56% to 64% of discharge coming from a foreign basin.

Under the NRLP, 22% of the water discharge at the mouth of the Godavari will be sourced from a new watershed, along with 38% at the mouth of the Krishna, 95% at the mouth of the Penna, and 59% at the mouth of the Kaveri. Of all the rivers affected by the NRLP, the Penna is expected to experience the largest change in runoff source area. As a “middle-man” in the great exchange, the Penna will experience through-moving water transfers totaling more than ten times its current discharge. The effect will be to almost entirely replace the water of the Penna with water from other basins. Similarly, the Krishna is projected to experience no change in the total volume of water exiting the river mouth, but it will nevertheless experience a large change in the proportion of water that comes from its natural watershed. This is due to the compensating transfers built into the NRLP, in which rivers receive water from foreign basins to compensate for sending water further down the system.

From the data compiled herein, it is possible to calculate the full annual transfer volume of the NRLP canals. Disregarding water that discharges into one canal and continues directly into a second, but including water that mixes with an existing river before being transferred again, the total annual transfer volume proposed for the NRLP is 245 km³ y^–1. This value is larger than the 174 km³ y^–1 reported by Bagla (2014) and Joshi (2013) because we have included utilization and intra-Ganga basin transfers (transfers between Ganga tributaries).

Among interbasin transfer systems, the NRLP is vastly larger than any existing or proposed system worldwide. The project is larger even than China’s South–North Water Transfer Project, which aims to transfer 44.8 km³ y^–1 from the Yangtze river to the Yellow river via three mega-canals (Zhang, 2009). The Central Route of the South–North Water Transfer Project was completed in 2014 and currently moves 9.5 km³ y^–1 through 1,264 km of canal. The largest canals of the NRLP will be larger than any currently existing canal, including the Central Route. If fully realized, the NRLP will contain nine of the ten highest-volume canals in the world. Even if Ganga-Brahmaputra intra-basin transfers were removed from the project, The NRLP would remain the world’s largest interbasin water transfer system at 129 km³ y^–1. By comparison, the largest interbasin water transfer system in the United States, the California State Water Project, transfers water from Northern California to Southern California in the amount of 3 km³ y^–1 (DWR, 2011). Additional large canals, such as the All-American Canal and the Colorado River Aqueduct, transport 3.2 and 1.5 km³ y^–1, respectively (Ghassemi and White, 2007), and structures in coastal Louisiana move 1–3 km³ y^–1 within the Mississippi delta.

The NRLP was planned with the expectation that it will bring many benefits to India. This study found that, if implemented in full, NRLP irrigation volumes would indeed exceed 100 km³ y^–1 and provide 11 km³ y^–1 of water for domestic and industrial needs. Increased domestic water supplies and more continuous availability throughout the dry season would help mitigate the risk of dehydration during heat waves, which killed 2,500 people in 2015. Hydropower will increase the standard of living. Flooding in northern India may be reduced; however, seismic hazard in the region is considerable, and the feasibility of large dams under these conditions is a matter of ongoing research.

However, the project will not benefit all people equally. We show here that dams proposed in the NRLP will submerge 3,400 km² of land and displace approximately 700,000 people (Table 3). Eleven dams will each displace 10,000 people or more, with the two highest displacements associated with dams on the Godavari: 172,000 at the Inchampalli Dam and 203,000 at the Polavaram Dam. Additional displacement may occur along canal routes, displacement which is unaccounted for here. Some displacement may be mitigated by voluntary migration in advance of construction.

Though the NRLP will increase water availability within basins, we calculate mean annual water discharge at river mouths/basin outlets will decrease for 24 out of the 29 rivers studied (Figure 6). This will likely drive salinity incursion that may damage wetlands and contribute to the deterioration of freshwater and estuarine ecosystems (Werner et al., 2013). The decrease in river discharge would come on top of an already decreasing trend in basin outlet discharge across the subcontinent caused by increased water usage (Gupta et al., 2012). For the Krishna, mean annual discharge in the 1950s totaled 62 km³ y^–1, but the modern river basin is nearly closed with a discharge of only 12 km³ y^–1 and no flow for several months per year (Rao et al., 2010). Despite increasing water utilization within the basin, the NRLP will not change river mouth conditions for the Krishna. Increased usage and transfers upstream will offset water flows designated for the Krishna delta. Of the five rivers that will experience increased discharge, two (the Yamuna and Betwa) are Ganga tributaries that do not flow into the sea. The remaining three – the Hooghly, Penna and Kaveri – may see an overall decrease in salinity intrusion as freshwater is added to these systems. The NRLP will also modify the hydrographs of the major rivers by impounding flows during the monsoon season and releasing them in the dry season. For monthly flows, mean changes will be within natural variability; however, average-flow years would drop to current dry-year levels, and dry years would experience unprecedented low flows (Figure 7).

Under the NRLP, suspended sediment load would decrease for five of the six rivers, with a scant increase for the Kaveri compared with current conditions. Approximately one-third to one-half of the reduction will come from changes to monthly water discharge, which will convey new suspended loads as approximated by the unique sediment rating curve for each river. The remaining reduction to transport is driven by sediment trapping in new reservoirs, which together have a gross storage of 98.7 km³. We note that these values are estimates, as canals may change the sediment rating curves once emplaced; sedimentation may occur within the canals and at their offtake locations, and trapping may be partially mitigated by design features such as gates for sediment sluicing. However, given the set of assumptions herein, the NRLP will reduce sediment loads dramatically for all but the Kaveri river. In the Kaveri, the increased sediment load represents only a negligible increase over a load that is currently near zero.

Deltas are sustained by sediment deliveries, and reduced sediment loads will translate directly to decreased aggradation. For the Ganga-Brahmaputra delta, a projected aggradation decrease from 3.6 to 2.6 mm y^–1 is enough to compromise the delta’s elevation with respect to sea level. Global mean sea level is rising at a rate of 3.3 ± 0.4 mm y^–1 and is expected to accelerate in the coming century (Church et al.,. 2013). Relative sea-level rise on the Ganga-Brahmaputra delta is currently even higher, with a mean value of 5.6 mm y^–1 across the delta (Brown and Nicholls, 2015). Relative sea-level rise encompasses eustatic sea-level increases along with the added influence of tectonics, sediment compaction, and accelerated compaction due to human activities such as hydrocarbon extraction and groundwater abstraction. Aggradation is the only process currently countering relative sea-level rise in the near-shore portions of the Ganga-Brahmapura delta; reduced aggradation will further compromise the population of 40–50 million people living near the coast and endanger the world’s largest mangrove forest, the Sundarbans.

As in all deltas, relative sea level rise rates in the Ganga-Brahmaputra delta are spatially variable, ranging from –1 mm y^–1 from tectonic uplift, to 43 mm y^–1 with a standard deviation of 7.3 mm y^–1 (Brown and Nicholls, 2015). The high standard deviation reflects both differences in methodology and true differences in aggradation rates across the delta. In the Ganga-Brahmaputra delta, aggradation rates vary between active and inactive portions of the delta, and on a smaller scale rates with distance from the nearest active channel (Rogers et al., 2013). Delta aggradation rates are also affected by river floods, tides and wave conditions, as well as human activities such as embanking (Auerbach et al., 2105). In this work, estimates of aggradation changes assume uniform deposition over the entire delta, which is a simplification relative to the complex depositional environments of active delta systems. Nevertheless, these are useful calculations for giving first-order estimates of the impacts of reduced sediment loads on delta surface elevation. In reality, spatial variability of sediment deposition would concentrate the impacts into a smaller area, resulting in more dramatic reductions than reported here for the active portions of the deltas.

Table 6 shows current and NRLP aggradation rates compared to historical aggradation rates for each river. Current aggradation rates have fallen by 80–100% compared to historical rates for all but the Ganga and Brahmaputra rivers, primarily due to reservoir trapping (Sarker, 2004; Gupta et al., 2012). This has likely driven the significant shoreline erosion that has been observed at the Godavari and Krishna deltas, where 76 km² have been lost over the past 43 years (Rao et al., 2010). At the Kaveri delta, which hosts a population of 4.8 million people, erosion has been observed along most of the coastline (ADB, 2011). The NRLP would drive aggradation rates further downward, exacerbating existing conditions in the Godavari and Mahanadi deltas while doing nothing to increase sediment flow to the Kaveri and Krishna deltas despite nominally directing water to these areas.

Tessler et al. (2015) quantified the risk of fluvial or coastal flooding and its expected economic impacts and losses of life for 48 deltas and found the Ganga-Brahmaputra and the Krishna to be the highest-risk deltas in the world, with the Ganga-Brahmaputra, Krishna, Godavari and Mahanadi all among the ten highest-risk. The Ganga-Brahmaputra, Mahanadi, and Godavari are the 1st, 10th, and 15th most populous deltas in the world, respectively (Higgins, 2015). Under the NRLP, the risk to these populations would increase. Rare ecosystems and vital agricultural areas would become more vulnerable to storm surges, river flooding, and heightened salinity. The NRLP system has no benefit to delta surface elevation or delta aggradation rates. With respect to sediment deposition and surface elevation, the system will push the deltas further in the wrong direction.

If completed, the NRLP will transform watershed boundaries, with more than half of the land in India contributing a portion of its runoff to a new mouth. Public health impacts from changed runoff regions may be severe. Currently, 52% of rural residents use open defecation without designated toilet areas (NSSO, 2016). Fecal contamination of water sources has made diarrheal disease the third leading cause of child mortality, at more than 300,000 child deaths per year (Lakshminarayanan and Jayalakshmy, 2015). Although fecal coliforms are short-lived, increased connectivity within the Ganga basin will expose new waterways to nearby contaminants. The Godavari would be exposed to runoff from the heavily polluted Ganga basin.

Exposure risk becomes clear when the NRLP is modeled as a full, connected system. Reservoirs on the Krishna, Godavari and Penna will host mass transfers of water moving from north to south. Mixing between river and canal waters will occur at many canal-river crossings, which in some cases will replace more than 50% of river water with water from other basins. Exact mixing proportions within reservoirs will depend upon residence time and on density differences between reservoir and canal water and may exceed the estimates in this study. They will not be smaller than these estimates unless the system is redesigned, as we have assumed zero mixing in reservoirs except when the lag between discharge and withdrawal is on the order of months.

Changes to habitat network connectivity will also have implications for invasive species management and biodiversity. As noted by Grant et al. (2012), canals 24 and 25 are expected to be particularly destructive to freshwater fish biodiversity, as they will link highly biodiverse western rivers to the much larger Krishna and Cauvery basins, allowing the spread of common species into previously isolated basins.

Future climate change may alter the feasibility of the NRLP. Compared to the late 20th century, climate model projections for India suggest a 2–5°C increase in the mean annual temperature by the end of the 21st century (Kumar et al., 2006). This warming will accelerate evaporation and evapotranspiration and increase water demand in urban and agricultural regions (Hasson et al., 2013). Along the link canals, the NWDA estimates that approximately 9.8 km³ y^–1 will be lost to seepage and evaporation, which may increase in a warmer climate. The NWDA estimate does not include evaporation losses from reservoirs, which may also be substantial.

At the same time, global climate models (GCMs) suggest an increase in accumulated rainfall from the summer monsoon, particularly in western India (Kumar et al., 2006; Immerzeel et al., 2010), with some geographic and intra-annual variability evident in downscaled data (Pervez and Henebry, 2014). In the Ganga and Brahmaputra basins, basin-integrated precipitation minus evaporation (P – E) is expected to increase (Hasson et al., 2013). The combination of increased P – E and accelerated melting of snowfields and glaciers in the Himalaya will temporarily increase flows on meltwater-fed major rivers and their tributaries (Lutz et al., 2014). However, once glacier water reserves are depleted, likely by mid-century, decreased streamflows on the Ganga (–18%) and Brahmaputra (–20%) are expected, even when the projected increase in annual rainfall is taken into account (Immerzeel et al., 2010).

The NRLP is expected to further decrease streamflows, which will likely affect delta salinity conditions. Under a future rising sea level scenario, the salinity of groundwater and river channels is expected to increase (Werner et al., 2013). Climate-related salinity incursion in rivers and deltas will be exacerbated by the decrease in river mouth discharge brought about by the NRLP.

This study has resulted in the creation of a database and graph database model that allows calculations of changes to river water discharge for 29 rivers in India given realization of the proposed National River Linking Project. We have further created a prototype method for evaluating the impacts on suspended sediment transport given the changes in water discharge estimated from the most up-to-date empirical data. Considering the complexity and disputed nature of the project, is likely that the project will follow an evolving path to either partial or full completion. To illustrate the potential large-scale magnitude and impact of the project, this study operates the graph database model in the worst-case scenario (“all infrastructure will be realized,”) but many other scenarios can be simulated and explored with this flexible model. If the NRLP is to be fully implemented as is proposed, it will decrease mean annual basin outlet discharge for 24 of the 29 affected rivers (as much as 73%), reducing freshwater deliveries to wetlands and estuaries. Additionally, more than 700,000 people will be displaced in the construction of the dams, and waterways will be exposed to new contaminants, invasive species, and disease-causing agents. Finally, the NRLP will compromise the already vulnerable deltas of the Indian subcontinent by decreasing aggradation rates as much as 87%. Additional water and sediment discharge data, particularly for Himalayan rivers, would be invaluable for refining estimates of the impacts of this colossal system.

Watershed change and river/canal discharge data used in this study are included as supplemental material. They can also be accessed at https://csdms.colorado.edu/wiki/Data:NRLP_India. These data are available in csv, shapefile, and Neo4J (graph database) formats. The code used to process the data are available at the above website and on GitHub, https://github.com/sahiggin/NRLP.

Krishna and Godavari river discharge data were kindly provided by the Office of the Chief Engineer, Krishna and Godavari Basin Organisation, Central Water Commission, Government of India. Ganges and Brahmaputra river discharge data were kindly provided by the Regional Integrated Multi-Hazard Early Warning System for Africa and Asia (RIMES Project), via the Government of Bangladesh. We thank Rochelle Worsnop for providing helpful assistance with figure design and Gabriel Ycas for productive feedback regarding study conception, methodology and philosophy. We are grateful for the thoughtful comments from Alex Kolker and an anonymous reviewer, whose feedback improved the paper.

S.H. was supported by NASA award 10-LCLUC10-2-0038; S.H. and I.O. were supported by NSF/Belmont Forum grant G8MUREFU3FP-2201-037. I.O. received additional funding from NSF award OCE 1600287. K.R. was supported by NSF award SES 1415431. E.K. held a National Research Council (NRC) Research Associateship award at the Atlantic Oceanographic and Meteorological Laboratory (AOML) and the Earth System Research Laboratory (ESRL).

The authors have no competing interests to declare.

• Contributed to conception and design: SH, IO, KR

• Contributed to acquisition of data: SH, IO

• Contributed to analysis and interpretation of data: SH, IO, KR Drafted and/or revised the article: SH, IO, EK

• Approved the submitted version for publication: SH, IO, KR, EK