River interlinking alters land-atmosphere feedback and changes the Indian summer monsoon

Massive river interlinking projects are proposed to offset observed increasing droughts and floods in India, the most populated country in the world. These projects involve water transfer from surplus to deficit river basins through reservoirs and canals without an in-depth understanding of the hydro-meteorological consequences. Here, we use causal delineation techniques, a coupled regional climate model, and multiple reanalysis datasets, and show that land-atmosphere feedbacks generate causal pathways between river basins in India. We further find that increased irrigation from the transferred water reduces mean rainfall in September by up to 12% in already water-stressed regions of India. We observe more drying in La Niña years compared to El Niño years. Reduced September precipitation can dry rivers post-monsoon, augmenting water stress across the country and rendering interlinking dysfunctional. Our findings highlight the need for model-guided impact assessment studies of large-scale hydrological projects across the globe.

algorithm called Peter Clark's Momentary Conditional Independence (PCMCI) 14 (See methods for details) to generate causal networks between different hydrometeorological variables. The TE is an information-theoretic metric that finds the dependence between two variables by excluding the effects from the history of the target variable. It can capture the nonlinear and lagged causal connections between variables, and can be considered as a nonlinear extension of Granger Causality 46 . The TE becomes difficult to estimate for a large number of variables because of the complexity due to dimensionality. The PCMCI adopts a two-step method to handle the curse of dimensionality while delineating causal structure from time series (methods).
We use both the approaches (TE and PCMCI) on the variables (Extended Table 1 Weather Forecast (ECMWF, ERA-5) 20 . The extended Figure 1 presents the climatology of soil moisture, precipitation, and Evapotranspiration (ET, generated from latent heat flux) for different basins generated using ERA-5 variables. All the basins receive maximum precipitation during the Indian summer monsoon (also called southwest monsoon) from June to September 47 . The Cauvery basin also receives significant amount of rainfall during October-December during the northeast monsoon 48 and has two peaks in annual precipitation. The soil moisture in Ganga, Godavari, Krishna Mahanadi, and Narmada-Tapi basin peaks during August and starts declining by the end of the summer monsoon season. The soil moisture in Cauvery peaks during late October, showing cumulative effects of rainfall from Indian summer monsoon and northeast monsoon. Evapotranspiration in all basins increases during the start of the Indian summer monsoon and is highest during post-monsoon because of moisture accumulation during monsoon getting exposed to solar radiation.
After performing causal analysis, we represent the association between variables across different basins as networks. We demonstrate the causal relationships between land variables across basins through land-atmosphere, atmosphere-atmosphere, and atmosphere-land interactions showing that the basins are not hydrologically independent. A perturbation in a river basin due to the proposed interlinking can travel to the neighboring basins by atmospheric pathways. Further, we used a modified regional climate model -Weather Research and Forecast coupled with Community Land Model 4 (WRF-CLM4, details in Methods) 18 -to test the hypothesis that by land-atmosphere feedback, the additional irrigation from river interlinking can lead to changes in the Indian summer monsoon spatial patterns and the hydrology of the neighboring basins. To the best of our knowledge, such feedbacks have not been considered in the literature for any globally existing or planned interlinking projects.
This study shows that river basins are linked to each other by land-atmosphere feedback and any perturbation in one basin can travel to neighbouring basins. This result is at odds with the general assumption of independence of river basins while planning hydrological projects. We also show that by land-atmosphere feedback, river interlinking projects in India will affect Indian summer monsoon leading to a reduction in September rainfall in dry regions of the country which can further aggravate the water stress. The methodology and results presented here pave way for similar scientific assessment of the impacts of interlinking for river basins and other large scale hydrological projects across the globe.

Information Links between the Basins
We computed the transfer entropy (TE, Details in Methods) between the land variables across the basins using 40 years of daily reanalysis data from ERA-5 (1981ERA-5 ( -2020. We used the entire period for deriving the TE with a maximum lag of 10 days. Figure 1(c) shows the directional bivariate TE links between land variables across basins showing a link only if it is found statistically significant at 95% confidence (see methods) . Each arc is a variable, and the arrow represents the link's direction to the other variable it projects onto. With the increase in the number of variables, estimating multivariate TE becomes computationally expensive; hence, we stick to bivariate analysis. We get a large number of links between land variables using TE across the basins. We found that the Ganga basin land variables produce a high number of causal links affecting land variables of other river basins with a low number of incoming links from other basins. For example, the latent heat flux and soil moisture of the Ganga basin shows the existence of causal links to at least one of the land variables of all other basins and there is no incoming link to soil moisture of Ganga basin from other basins. High recycled precipitation due to land-atmosphere feedback is well established for the Ganga basin 43,44,49 . Cauvery basin on the other hand has a large number of incoming links from all other basins, as evident from Figure 1 (c). Literature shows that the Cauvery basin receives recycled precipitation generated by evapotranspiration from the neighboring regions 50 . Stronger causal connections exist between the land variables of other basins as well.
The above bivariate TE analysis shows causal relationships across land variables; however, some of the links we see in the Figure 1 are likely due to common drivers such as the El Niño-Southern Oscillation (ENSO), and/or indirect links. A co-variability between the soil moisture in the Mahanadi basin with that of Ganga, for example, may be associated with monsoonal variations or the interannual impacts of ENSO. It is also possible that the local changes in soil moisture in a river basin may indirectly affect that in another river basin through local landatmosphere interaction, which in turn would affect the large scale flow. Thus, the most likely reason for these links would be indirect links or common drivers. To address these challenges, we used an advanced causality approach, PCMCI (see Methods for details), which tries to account for common drivers while controlling the high-dimensionality. The monsoon drives the climate and the water cycle in India. To understand the land-atmosphere processes and its impacts on water cycle, we have performed analysis using PCMCI for the summer monsoon season. We have applied PCMCI to all the land and atmospheric variables from reanalysis data separately for each year's monsoon seasons with 122 days each, considering a maximum of 10 days' lag. Figure 2 presents the links appearing for more than 10 years out of 40 years (1981-2020). We hypothesized that the causal connections between land variables of two different basins A and B exist through a series of indirect links: land variable (river basin A)  atmospheric variable (river basin A)  atmospheric variable (river basin B)  land variable (river basin B).
We present the links in the same way in Figure 2 Interestingly, we also see a similar pathway from the Mahanadi variable to that of Ganga, indicating a feedback.
The intra-basin land-to-atmosphere connection happens by evapotranspiration contributing to the moisture content of the air while causing surface cooling, whereas atmosphere-to-land connections occur by precipitation and temperature changing soil moisture and surface energy balance 45,51 . The atmosphere-to-atmosphere interactions between different basins occur through moisture and heat transported by winds across basin boundaries. To make sure that the links are not specific to a single reanalysis product, we applied PCMCI to another reanalysis data: Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2) 21 . The results are shown in Extended Figure 2. We found similar links existing in the network derived from MERRA-2. There is a clear resemblance in the nodes present in the 4 layers across the reanalysis, though they are not exactly the same. This proves that, due to the landatmosphere feedback, the assumption of hydrological independence of neighboring basins does not hold true.

Feedback from the proposed interlinking
Based on the above causal analysis, we hypothesize that a perturbation in the land variables of a receiver basin due to the proposed interlinking can also affect its neighboring basins (for example, feedback between Ganga and Mahanadi explained above), including the donor basin previously been demonstrated to posess a reasonable skill in simulating Indian summer monsoon 9 . We performed another simulation (the irrigation run, hereafter IRR) by increasing the percentage irrigated area to 80% in regions where interlinking projects target an increase in the culturable command area as shown in Figure 1 (b). Figure 1(b) shows the increase in the percentage irrigated area in each grid cell done to achieve 80% irrigated aria in IRR run. The IRR run provides irrigation (in addition to irrigation in CTL run) of around 600mm/day (1400mm/day for paddy 9 ) to the area of around 30 million hectares across the country ( Figure   1(b)). The simulations consider the India-specific crop and irrigation practices 9,52,53 (details in methods). The differences in results between the two simulations highlight the feedback from the interlinking through land-atmosphere interactions.
Extended Figure 3 shows difference in mean daily precipitation between IRR and CTL runs (IRR-CTL) for the monsoon season (b); June to September, JJAS), June (c), July (d), August (e), September (f). Hatched lines in plots indicate the regions where the difference is statistically significant at 90% confidence level. Spatially, there is not much statistically significant changes in precipitation during June, July, and August, though during July, we see some rainfall deficit in the western India and a slight increase in precipitation for the rest of the country. Importantly, September sees a widespread and maximum statistically significant reduction in precipitation.
The simulated changes in September rainfall can be attributed to land-atmosphere feedback.
Contribution of land-atmosphere feedback to Indian summer monsoon is known to peak during September when the soil contains high moisture leading to high evapotranspiration 30   What is more, the interlinking will also result in a changing spatial pattern of temperature over India (Extended Figure 4, a). Extended Figure 4 shows the difference in mean daily values for daily maximum temperature, surface latent heat flux(b), and root zone soil moisture (c) between IRR and CTL runs for September. The changing meteorological patterns result in statistically significant changes in mean monthly soil moisture and latent heat flux. However, there is a lack of one-to-one consistency everywhere due to complex hydrometeorological processes. The regions with less precipitation are accompanied by an increase in daily maximum temperatures of up to 1°C. and a decrease in soil moisture of around 15 mm (Extended Figure 4; a,c, respectively). The changes in soil moisture in the grids receiving daily irrigation cannot be used to quantify impacts of land-atmospheric feedback, and hence, masked with grey color. The irrigated grids are also visible to some extent as having high latent heat flux in Extended Figure   4 Figure 6). The dry western region shows a decline in rainfall and soil moisture even for the El Niño years with an increase in temperature (Extended Figure 5). The further drying patterns of the arid region due to interlinking could be alarming and hence, needs to be addressed in the planning for interlinking. Central Indian regions show an improvement in the rain due to interlinking in El Niño years, which is good for the dry years. Overall, we found that the perturbed water management from the proposed interlinking can lead to changes in the spatial distribution of the Indian Summer monsoon and a systematic reduction of precipitation in many regions, including the dry arid regions.
Further, to see whether the reduction in precipitation in IRR experiment is related to landatmosphere feedback from extra irrigation provided in IRR experiment, we apply TE to find causal connections from the LH of extra irrigated regions to the precipitation over the drying regions. We considered the differences between IRR and CTL Simulations for each year separately for the same. The results are shown in Figure 4. We use TE here, as we want to capture both direct and indirect connections from latent heat flux to precipitation in model simulations. Figure 4

(a) shows 3 chosen regions (southern peninsula -region A, western India
-region B, and a part of ganga basin -region C) out of areas where irrigation was applied and

Implications of Interlinking Projects
India has a rapidly growing problem of water stress due to global warming, population growth, pollution, and change in land use. As per the Central Water Commission, Government of India, the current per capita availability of water in India is around 1400 cubic meters which is slated to reduce to around 1200 cubic meters by 2050 and a large portion of country is already classified as water stressed 57,58 . A large fraction of India's water resources is used for agriculture and the irrigation is practiced across the country. The demand for water will further increase with rapid intensification of agriculture. As water demand is rising rapidly, within the next 20 years, India might need most of its runoff to meet its urban and agricultural needs 59 . As a solution to this problem, India has planned river-interlinking projects to transfer water from surplus to deficit basins to cater to the water demand of growing population. The goal is to keep maximum possible water on land for utilization which originally used to reach oceans from river basins. The assumption behind such planning is that 1) the hydrology of the river basins is independent; 2) the feedback from interlinking will not affect the rainfall patterns.
Here, we find that both the assumptions made for the interlinking are not valid. The perturbed hydrological processes of the receiving river basins send feedback to the Indian monsoon and norther (states of Punjab, Haryana, and Uttarakhand) parts of India, which, based on our experimental set up, can be attributed to land-atmosphere feedback from interlinking. The reduction in September precipitation will dry up the rivers in the subsequent months amplifying water stress manifolds in various parts of the country, which is an unexpected and unintended result from interlinking. The majority of the population in these areas is dependent on agriculture.
A reduction in monsoon rainfall would cause damages to socio-economy of these regions increasing climate vulnerability and risk. It is noteworthy that we have not considered the feedback on the monsoon rainfall in response to the reduced runoff to ocean due to interlinking.
Recent studies show land to ocean runoff can perturb the monsoon rain 60 and may intensify the feedback quantified by us. Hence, proper quantification of the feedback from proposed interlinking policy needs careful scientific investigation. Our study is the first attempt to quantify the impacts of any large scale hydrological project like river interlinking on Indian Summer Monsoon, which was not considered in planning these projects. Our results highlight the importance of regional land-atmosphere model-driven hypothesis testing and impacts assessment while planning for large-scale hydrological projects.

Data
We use 40 years (1980-2019) of daily data from two sources (Extended Table 1 December ONI values of any year being above 1 and below -1 respectively (Extended Table   2). Information exchange takes place between two variables ( and  with time 17,62,67,68 . In this study, number of bins taken is 11 which has been argued to be appropriate for measuring TE given sufficient data length 17  October). Control experiment (CTL) prescribes current irrigation water application over India using estimates from the agricultural census and a gridded reconstructed data 52 . The Irrigation experiment (IRR) adds additional water as irrigation by maximizing the irrigated area fractions on the grids which are going to be benefitted from interlinking. Figure 1