Introduction

Effective monitoring and forecasting of water-related hazards are of paramount importance to ensure societal and environmental sustainability. Remote sensing has emerged as a potent tool for hazard monitoring1,2,3 and assessment4,5,6,7, particularly in data-poor or access-restricted regions8. Despite this technique’s unique ability to unveil insights into the integrity of water infrastructure9,10, detect operational anomalies, and identify early warning signs of structure compromise11, comprehensive case studies that demonstrate emergency response application on actual events remain scant.

On June 6, 2023, the Kakhovka Dam in Ukraine experienced a catastrophic breach. The environmental loss was approximated at 1.2 billion euros12, with at least 62 human lives lost, hundreds injured, and up to 100,000 individuals displaced13. These consequences were compounded by a near-total loss of irrigation systems in the agricultural sector following the disruption of four major canals fed by the Kakhovka reservoir, which in 2021 had enabled the harvest of approximately four million tons of crop products worth about $1.5 billion14. Rebuilding the dam is projected to require at least five years and ~$1 billion15, a stark illustration of the aftermath of such disasters. Various reports and studies have documented damage inflicted by missiles, shelling, or mines on bridges, piers, and even sluice gates at the dam site16. These war-related impacts, compounded by anomalous reservoir regulation and spring flood conditions, led to the lowest and highest water levels ever recorded17,18. Before the breach, the supporting structures downstream of the activating spillway started to show signs of being compromised19. These reports warrant an in-depth analysis to elucidate the factors contributing to the dam’s pre-failure condition, which may have exacerbated downstream flood damage post-breach, with the ultimate aim of identifying preventive measures to avoid such dangerous states.

In this study, we synergized multi-sourced remote sensing data, meteorological reanalysis products, and the dam design parameters to detect anomalous reservoir operation, estimate overflow conditions, and identify signs of potential structure compromise. We outlined possible safety checkpoints and proposed potential risk-mitigation strategies. Our objective is to underscore the utility of remote sensing data for water infrastructure monitoring by proposing new technical standards to effectively identify operational anomalies and irregularities. This approach bolsters the early warning capacity for water hazards, particularly in regions with limited access to traditional monitoring systems.

Results and discussion

Identification of irregular operation and structure compromise

Upon meticulous examination of multi-sourced optical images during the period from 2017 to 2023, we pinpointed the inception of the Kakhovka Dam’s anomalous operations starting on November 11, 2022.

As demonstrated in Fig. 1a, the conventional operation of the spillway sluice gates predominantly occurred in early summer, with activation varying from around 30–80% and typically lasting for a duration of less than 2 months, a pattern characteristic of flood mitigation. However, from November 2022 onwards for over 6 months, the spillway activation percentage remained approximately constant, 15% of the total capacity, and the activated gates of the spillway remained the same (Fig. 2c–f). Consequently, in 2023, the reservoir dropped to the lowest in late January and escalated to the highest water level in April, respectively, compared with the same period of the past three decades (Fig. 1c), and while the water levels of the other reservoirs in the Dnipro cascade remained within reasonable ranges (Supplementary Fig. 1). Figure 2a, b shows a damaged segment of the bridge to the river’s right bank on November 11, 2022. Around 3 days later, a limited number of gates near the left bank were activated (Fig. 2c). Furthermore, the outflow rate of the dam can be approximated as a function of the water level from November 2022 to June 2023, whereas such an approximation did not hold over the 2017–2022 period (Fig. 3a). This may indicate that effective regulation has been absent since November 2022, which can be confirmed in the tests ending in as early as the first week of January.

Fig. 1: The identification of operational anomalies of Kakhovka Dam using remote sensing data.
figure 1

a Daily activation percentage of spillways from 2017 to 2023: darker shades represent a higher percentage of activated spillways. b The cubic spline interpolated, and 7-day average smoothed reservoir water-level series over 1992 to 2023, in EGM2008 datum. c Water-level time series zoom-in during the pre-failure period, with sequential event time marks observed from PlanetScope images: right bank bridges missing (11–11), spillway activation fixed at 15% (11–23), right bank gates overtop (04–23), left bank bridges compromised (06–02), additional left bank parts missing (06–04), and dam breach (06–06).

Full size image
Fig. 2: Monitoring of spillway activation on key dates illustrated through PlanetScope images.
figure 2

a, b Before and after bridge damage near the right bank of the spillway on November 11, 2022. c Small portion of the spillway activated on November 14, 2022. d The spillway structure started overtopping on April 23, 2023. e The bridge near the left bank began to show signs of compromise on June 2, 2023. f Additional parts were missing by June 4, 2023. The white polygons represent the focus areas of anomaly events. The white arrow in (a) indicates the extent of the spillway structure. Image © 2022–2023 Planet Labs PBC.

Full size image
Fig. 3: Reservoir unregulated outflow hypothesis test and storage change simulation.
figure 3

a Non-regulated outflow hypothesis test metrics, measured by the coefficient of determination. The upper subplot presents the boxplot of metrics for 2017–2022, while the lower subplot depicts the same for 2022–2023. The start date for all tests is November 11, with end dates indicated on the x axis. b Storage change and outflow simulation for the period from November 2022 to June 2023. \(\bigtriangleup {V}_{{balance}}\) and \(\bigtriangleup {V}_{{HA}}\) represent the simulated and remotely sensed storage change of the Dnipro cascade, respectively, while \(\bigtriangleup {V}_{{upstream}}\) corresponds to the remotely sensed storage change for five upstream reservoirs. On 2023-04-22, the estimated outflow exceeded the design capacity of the spillway that was activated. Refer to “Methods” for further details.

Full size image

The spillway banks began overtopping around April 23, 2023 (Fig. 2d), and the estimated outflow through the activated spillway surpassed the designed capacity by April 22, 2023 (Fig. 3b). Subjected to a month-long overflow condition, the spillway ancillary structure exhibited visible damage to the pier on the downstream side of the opening gates by May 28, 2023, as revealed by ultra-high-resolution Maxar images19. The curved bridge and its supporting structures on the same location of the spillway began to show signs of compromise on June 2, 2023 (Fig. 2e), with additional parts missing by June 4 (Fig. 2f). This sequence of events highlights that the pattern of damage to the spillway during the pre-failure period was likely attributable to erosion at the spillway foundation, resulting from the extended period of overflow 20,21,22.

Safety checkpoints and mitigation measurements

As listed in Table 1, the initial safety checkpoint would be November 11, 2022, when the damage inflicted upon the sluice gates likely undermined the dam’s performance for flood protection. The subsequent checkpoint emerged on January 31, 2023, when it became clear that the dam’s outflow is unregulated, potentially resulting in a low water level that poses a threat to water resource safety. The third checkpoint likely surfaced on February 28, 2023, due to the sustained usage of the same portion of spillway which could bring about or accelerate the loss of structural integrity. The fourth checkpoint was on March 31, 2023, when the water level commenced its upward trend, and the estimated probability of overtopping, under the conditions of observed spillway activation, reached 23%. Given the dam was already overtopping based on PlanetScope image (Fig. 2d), the fifth checkpoint on April 23, 2023, signals an emergency state, with the risk of persistent hazardous conditions pegged at 79% in the absence of corrective measures. The final checkpoint presented itself on May 28, 2023, when visible structural compromises were confirmed, with a predicted likelihood of 45% of the dam remaining in an overtopping state in June in the absence of action.

Table 1 Safety checkpoints with the corresponding risk conditions and mitigation measurements
Full size table

The proposed mitigation strategy for the early stages (from November 2022 to February 2023) is of minor to moderate intensity. Restorative actions such as repair of the damaged sluice gates or maintenance of others could have ensured the functionality of flood reduction. Even without comprehensive repairs, implementing control mechanisms such as the deployment of management personnel or preemptive water release by partially activating more spillways in March 2023 could have curtailed the likelihood of overtopping. Opportunities for mitigating overtopped conditions also remained, contingent on the operational state of the spillway during the later stages (from April to May 2023). Furthermore, preventative actions should be implemented in the potentially affected regions, commencing with flood warnings and subsequent evacuations, which could have mitigated impacts.

Dam management in war conditions

A series of conspicuous anomalies and signs of neglect, observed since November 11, 2022, likely contributed to the overtopping condition of the Kakhovka Dam, which might have intensified the flood damage downstream after the breach. Despite the visible damage, when disregarded, the spillway outflow should still uphold more than 80% of its designed capacity 23. Nonetheless, our analysis might indicate that the reservoir operations utilized a scant 10% of the total design capacity. It should be noted that the issue of overtopping was unlikely to be rectified by intensifying cascade regulation, given that from January to May 2023, the upstream reservoirs’ storage capacities were occupied by over 8.5 cubic kilometers of water (Fig. 3b), constituting over 80% of the current usable volume. The peculiar reservoir operation stirred concerns by local media before and during the spring flood season18,24, yet no effective countermeasures were implemented to mitigate the risk. Furthermore, as the dam began to show signs of impending structure damage, there was a conspicuous absence of a hazard warning, implying a possible violation of safety procedures and disaster prevention protocols.

In a war situation, ensuring the safety of the reservoir infrastructure is of paramount importance. This necessitates a conservative reservoir regulation approach, which includes operating reservoirs at lower than normal water levels25,26, implementing pre-release of volume prior to the flood season27,28, and assuring the maintenance of flood protection facilities. Although these measures might reduce the water supply of irrigation and civil usage as well as the power generation, the trade-offs are indispensable in avoiding disasters. Such a conservative strategy not only enhances the system’s resilience to hydrological extremes, but also markedly reduces the risk of overtopping when operational norms may be impacted by external forces.

The destruction of Kakhovka dam has caused a substantial decrease in usable volume in the Dnipro cascade, which underscored the urgent need for a reevaluation of the cascade regulation rules29. Special attention should be paid on Kremenchuk, the reservoir with the largest usable volume and the only remaining facility with annual regulation ability 30 along the Dnipro cascade. Given its position as the third among five existing reservoirs, it is particularly vulnerable and cannot afford any mishap. The Zaporizhzhya Nuclear Power Plant (ZNPP), which was dependent on the Kakhovka reservoir for coolant water, is another critical concern. It has been disconnected due to low water levels, and although the coolant water pool of ZNPP remains relatively stable in the shutdown state31, it would be safer to compensate for water loss via the original Dnipro river channel using temporary pipelines. As an ancillary measure, the enhancement of hydrological monitoring and forecasting capabilities32,33 might help facilitate optimal utilization of the remaining volumes within the ideal framework of conservative regulation.

Conclusion

Our study highlights the value of remote sensing in identifying operational anomalies and predicting early warning signs in water resource infrastructure, which can enable the development of effective mitigation strategies.

As satellite data accessibility improves, we are advancing toward near real-time, high-resolution remote monitoring of both water quantities and qualities in rivers34,35, lakes36,37,38, and reservoirs39,40,41,42. Data assimilation in conjunction with meteorological and hydrological models43,44,45,46 can further enhance the dynamic simulation of water resources and infrastructure operations, paving the way for a more comprehensive and nuanced understanding of water system behaviors and vulnerabilities.

Beyond our technological advancements lies an urgent societal call: the imperative of transparent data sharing, especially concerning public water infrastructure. As our study on the Kakhovka Dam underscores, this transparency is pivotal at the watershed scale, where potential cascading impacts arise from interconnected systems. Open access to regional data, encompassing design parameters, and both remote and in-situ monitoring, not only accelerates the development and refinement of scientific methodologies but also ensures more accurate predictions of water system behavior. In turn, this fosters the establishment of proactive water hazard mitigation strategies, safeguarding the longevity and safety of our water systems.

Methods

Spillway activation monitoring

The activation of a spillway is a key measure for flood mitigation in reservoirs. Specifically, at the Kakhovka Dam, the activation of its top flow-type spillway is effectively monitorable using remote sensing data. In this study, we tracked the percentage of spillway activation by analyzing multisource optical images from 2017 to 2023.

Optical images from PlanetScope, Sentinel-2, and Landsat-8 were used to analyze the spillway activation states. Table 2 summarizes each sensor’s configuration and the number of clear images used in the study, after excluding those with cloud cover. Data availability covering the dam location is detailed in Supplementary Fig. 2.

Table 2 Optical images used for spillway activation monitoring
Full size table

We measured the percentage of spillway activation by relating the water splash width to the total length of the spillway on a daily scale from optical remote sensing images, as depicted in Fig. 1a. For dates lacking valid observations due to cloud coverage or data absence, we assumed that the activation state was identical to that of the previous date of valid record. Consequently, the spillway activation percentage remained at around 15% from late November 2022 until the dam’s breach. Supplementary Fig. 3a–c demonstrates comparative visualization of PlanetScope, Sentinel-2, and Landsat dated 2023-03-22. There are also some snapshots of the spillway activation during normal flood mitigation periods, as depicted in Supplementary Fig. 3d–f.

The measurement’s uncertainty primarily stems from two factors: sensor resolution and the dynamic nature of the water splash. With the spillway dam’s total length being 435 m23, the per-pixel measurement bias is ~0.7% for PlanetScope, 2.3% for Sentinel-2, and 6.9% for Landsat-8. High-resolution images from PlanetScope are most effective for assessing spillway activation, even under haze conditions (Supplementary Fig. 3a). While Landsat’s medium resolution introduces a higher degree of uncertainty (Supplementary Fig. 3c), it is still valuable for confirming when gates are fully closed. The dynamic nature of water splash poses challenges for precise measurement, and its width may also correlate with the outflow rate, influenced by the reservoir’s water level. However, these uncertainties do not impede the identification of operation anomalies in this case, as the same spillway activation condition remains consistent for over 6 months.

Outflow hypothesis test

Supplementary verification was required to substantiate the hypothesis of unregulated outflow given the remote sensing images’ incapability to offer information regarding the vertical opening degree of the sluice gates and the operation state of the hydroelectric power station. We approximated the outflow through spillway using the equation of free flow through rectangular weir47 as:

$$Q=N\frac{2}{3}{c}_{d}b2{g}^{1/2}{H}^{3/2}$$
(1)

where \(N\) is the number of functioning gates, \({c}_{d}\) is the contraction coefficient, \(b\) is the width of the gate, g is gravity acceleration constant, and H is the head difference between the upstream and downstream water levels.

Assuming in the period of unregulated outflow, all parameters other than \(H\) are constant, by incorporating the design discharge capacity of spillway outflow 23 \({Q}_{{design}}\) = 15,438 m3 s−1 and percentage of spillway activation \({P}_{{activate}}\), we deduced:

$$Q=15438\times {P}_{{activate}}{\left(\frac{H}{{H}_{{design}}}\right)}^{3/2}$$
(2)

Where \({H}_{{design}}\) is the design water head.

To simplify, we assumed that the water head difference is proportional to water level of the reservoir, thus utilized the water level from altimetry product to approximate \(H\), meaning \({H}_{{design}}\) is a parameter in need of calibration.

After fixing the \({H}_{{design}}\), we could pinpoint the estimated start date of spillway overflow from reservoir water level time series.

We validated the hypothesis based on water balance of the Dnipro cascades, represented by the following equation:

$${V}_{{balance}}^{t}={V}_{{balance}}^{t-1}+({Q}_{{inflow}}^{t}-{Q}^{t}-{Q}_{{usage}}^{t})\Delta t$$
(3)

where \({V}_{{balance}}^{t}\) is the estimated cascade volume in time t. \(Q\) is the outflow through spillway, \({Q}_{{inflow}}\) is the streamflow, and \({Q}_{{usage}}\) represents water usage, including irrigation and civil water usage, taken from the cascade storage, assumed as a constant value, which is also a parameter in need of calibration.

Given the absence of streamflow observation data for the Dnipro River, we sourced daily discharge simulation data for the site of the Kakhovka Dam from GloFAS-ERA545, which have accounted the reservoir evaporation in the reservoir module, spanning the period from 1979 to 2023. We then refined the annual flow volume by implementing quantile mapping, based on the specific annual streamflow volume and corresponding reliability values stipulated in an evaluation report29.

We further calculated the remotely sensed cascade storage time series as:

$${V}_{{HA}}^{t}={V}_{{HA}}^{t-1} + {\sum }_{i=1}^{6}({H}_{i}^{t}-{H}_{i}^{t-1})\cdot ({A}_{i}^{t}+{A}_{i}^{t-1})/2$$
(4)

Where \({H}_{i}^{t}\) and \({A}_{i}^{t}\) is the water level and surface extent area of the ith reservoir of the Dnipro cascade at time \(t\).

We set the initial storage values to zero for both the estimated and remotely sensed storage time series at the starting step. Then, we test the unregulated outflow hypothesis by computing the coefficient of determination between the two time series.

For consistency, we leverage the Area-Elevation (A–E) relationship to determine the reservoir surface area based on the water level. The water levels were collected from the Global Reservoir and Lake Monitor (G-REALM) dataset48. We bridged the 7–10 days temporal resolution gap in G-REALM’s point data by applying cubic spline interpolation, followed by a 7-day moving average to create a smooth, continuous water level time series. The reservoir water surface area was determined using multisource remote sensing data, with optical data primarily delineating surface water boundaries and Synthetic Aperture Radar (SAR) imagery compensating for areas obscured by clouds. Water surface extent was extracted from optical images using a supervised classification method based on water indices49, supplemented by manual refinement, while SAR images were processed with an automated algorithm named RAPID2,3. The SAR images used in this study are the Radiometrically Terrain Corrected Sentinel-1 products at, C-band, dual-polarization, and 10-m resolution from Alaska Satellite Facility. Supplementary Table 1 presents the water levels and corresponding surface areas for the Dnipro cascade, derived from remote sensing data.

The test period commences on November 11, 2022, and concludes on the first day of each subsequent month through June. We applied the hypothesis test for each time period from 2017 to 2023 for comparison, constraining \({H}_{{design}}\) and \({Q}_{{usage}}\) as [15.8, 17.4] m and [20, 100] m3 s−1, respectively, and calibrated them by minimizing the Root Mean Squared Error between \({V}_{{balance}}^{t}\) and \({V}_{{HA}}^{t}\). For the percentage of spillway activation, we use the measured value 15% from “Spillway activation monitoring” to implement our primary test (Fig. 3). To ensure the reliability of our conclusions, we also incorporated a sensitivity analysis by varying the spillway activation percentage from 10% to 20% and evaluating the implications for both the storage time series and the overflow start date.

Supplementary Fig. 4a reveals that unregulated outflow is plausible when spillway activation ranges from 13.8 to 17.5%, which corresponds to about 4–5 sluice gates remaining open. The inferred onset of overflow falls between April 19, 2023, and May 9, 2023 (Supplementary Fig. 4b), is consistent with the initial overtopping observed from remote sensing image (Fig. 2d).

Key uncertainties in this section, including the spillway activation percentage, inflow simulation from GloFAS-ERA5, and the A–E curve derived from remote sensing data, were thoroughly examined. The first uncertainty was directly assessed through a comprehensive sensitivity analysis. The simulation streamflow has demonstrated commendable performance against observation at locations near the site of Kakhovka45, and the values were further adjusted using quantile mapping based on localized statistics, bolstering their reliability as proxy of the inflow. G-REALM data is known to exhibit minimal bias for large reservoirs (<10 cm)50, indicating that the principal challenge stems from its relatively low temporal resolution, which could influence the accuracy of the reservoir storage retrieval. Given the study area’s stable terrain, reservoir storage is more influenced by water levels than by surface area, reducing the effect of area measurement errors. Although G-REALM captured water level for every 7–10 days, and the trapezoidal cross-section assumption introduced some level of uncertainty, the extended test period and high monthly R-squared values lend credence to our results. These findings demonstrate that, despite the inherent uncertainties from multisource input data, the assumption of unregulated outflow remains a valid and well-supported hypothesis.

Estimation of overtopping probability

Inspired by Stefanyshyn and Benatov 51, we define the occurrence of overtopping in the Kakhovka reservoir as a sign of danger. We estimate the probability of overtopping52 in the subsequent time step by taking into account the stochastic nature of the inflow and the reservoir’s current state:

$${P}_{f}=P({H}^{t+1} \ge {H}_{o})=P({Q}_{{inflow}}^{t+1} \ge Q+({H}_{o}-{H}^{t})* ({A}_{o}+{A}^{t})/2/ \Delta t)$$
(5)

Where \({H}_{o}\) is the water level that triggers overtopping, \({A}_{o}\) is the corresponding water extent, and Q is the current maximum available outflow capacity.

We investigate the probability of overtopping in the subsequent time step, considering two scenarios of outflow capacity: the actual capacity without any mitigation, as estimated from remote sensing imagery, referred to as the “probability of overtopping” (PO), and the capacity after applying proposed minimal mitigation measures, termed “PO after MM”, as detailed in Table 1.

We designate the 99th percentile of the historical water level, excluding the years 2022 and 2023 as the \({H}_{o}\)(equal to 16.56 m), and we retrieve \({A}_{o}\) using the A–E relationship from “Outflow hypothesis test”. Consequently, Eq. (5) simplifies to:

$${P}_{f}=P\left({Q}_{{inflow}}^{t+1}\ge f(\Delta t,Q)\right)$$
(6)

We then evaluate the possibility of overtopping at two scales, as

$${P}_{f}=P({Q}_{{mean}}^{t+1}\ge f(30,Q))+P({Q\le Q}_{{mean}}^{t+1} < f(30,Q)\,{{{{{\rm{\& }}}}}}{\,Q}_{\max 7}^{t+1}\ge f(7,Q))$$
(7)

Here, \({Q}_{{mean}}^{t+1}\) and \({{Q}}_{\max 7}^{t+1}\) denote the mean streamflow and maximum average streamflow within a 7-day window for the subsequent month, respectively.

We employ a Gaussian copula function53 to determine the correlation of streamflow between successive months:

$$C(u,v,w)={\varPhi }_{\varSigma }({\varPhi }^{-1}(u),{\varPhi }^{-1}(v),{\varPhi }^{-1}(w))$$
(8)

Where \(C(u,v,w)\) represents the three-dimensional copula function, with variables \(u,v,\) and \(w\) denoting the transformed values obtained by applying the inverse of the cumulative distribution function to the observed data series of \({Q}_{{mean}}^{t}\), \({Q}_{{mean}}^{t+1}\) and \({{Q}}_{\max 7}^{t+1}\), respectively.

We fit the copula function with refined discharge simulation described in “Outflow hypothesis test”, excluding the data from 2022 and 2023. We employ a log-normal distribution as the marginal distribution function. Subsequently, we implement Markov Chain Monte Carlo sampling54 of the conditional distribution presented in Eq. (7) to estimate the overtopping probability.

Resolving analytical uncertainties with next-gen remote sensing

The current landscape of widely accessible remote sensing data effectively facilitates the identification and analysis of notable operational anomalies within large water infrastructure projects. As remote sensing technology continues to evolve, it promises substantial enhancements to this analytical process. High-resolution imagery, finer than 1 m, can swiftly pinpoint issues such as improper sluice gate operations or structural integrity losses, with the capability for optical detection or revealing internal damages via InSAR techniques10. In addition, the precision in monitoring reservoir storage is poised for improvement with the advent of new satellite altimetry missions like SWOT50, or by integrating high-quality bathymetric data and increasing the frequency of shoreline measurements55. Moreover, incorporating data assimilation techniques with remotely sensed streamflow retrieval at lake37 or river-reach56 scales is expected to greatly enhance the precision of inflow estimates, thereby strengthening water resource management and hazard mitigation strategies. By harnessing these advanced satellite products and synchronizing them with engineering design parameters, the establishment of a globally comprehensive, remote sensing-based monitoring system becomes a tangible prospect.