Introduction

The increasing frequency and severity of extreme weather events, particularly as global average temperatures rise, is a critical issue today1,2,3,4,5,6,7. This trend, primarily driven by human-induced climate change8,9,10,11 and increased carbon emissions12, disproportionately affects low-income and developing countries, especially in South Asia10,13,14. Despite its low contribution to global carbon emissions, Pakistan is highly vulnerable to climate-related disasters such as severe droughts and intensified monsoon rains14,15,16. These climate-related changes have led to a notable increase in flooding17, as evidenced by the devastating floods in Pakistan during the summers of 2010, 2011, and 201416,18,19. Nearly 1.47 billion people, about 19% of the global population, are directly exposed to substantial flood risks11,20. Developing nations are particularly vulnerable due to rapid population growth, inadequate infrastructure, and limited adaptation capacity12,13,21. These countries account for approximately 95% of all disaster-related deaths22. These events highlight the severe consequences of climate change, resulting in substantial economic losses and underscoring the global challenge of managing flood risks23.

In Pakistan, the monsoon season, driven by storms from the Bay of Bengal, brings considerable flooding challenges12,22,24, especially in regions like Punjab, Sindh, Khyber Pakhtunkhwa, Baluchistan, and Gilgit Baltistan. The unprecedented flood disaster of 2022 exemplifies the increasing severity of such events, causing extensive destruction and prompting a national emergency25,26. The 2022 flood resulted in over 1718 fatalities, 12,800 injuries, and approximately 33 million people affected, amounting to 15% of the country’s population (Fig. 1a) (NDMA, 2022). This devastating event led to widespread displacement, considerable infrastructure damage, and enormous economic losses, underscoring the urgent need to understand the driving factors behind such extreme weather events19. Emerging research increasingly links these events to larger climate change patterns, which influence the frequency and severity of extreme weather. According to the IPCC AR6, each 1 °C increase in global temperature could amplify the frequency of extreme precipitation events by 1.3 to 1.7 times and intensify them by approximately 7%. This increase in both frequency and intensity of extreme precipitation points to a future with natural disasters occurring at larger scales, with greater severity, and potentially reaching peak impacts earlier in the season.

Fig. 1: Overview of Pakistan’s 2022 flood and its inundation extent and growth.
figure 1

Panel (a) displays a map depicting the 2022 flooding in Pakistan, with dark blue indicating severe inundation and lighter blue representing areas with less severe flooding. Concentric rings illustrate statistics on affected populations, calamity-declared districts, regions, and displaced populations. Arrows indicate monsoon spells, with their occurrence dates shown. Panel (b) illustrates the flood’s progression from May to September. Panels (c–f) focus on the most affected Larkana Plain, where a 100 km² inland lake formed. (The base map in panels (a) is from the USGS, and the image in panels (c–f) is from a satellite image).

Full size image

One key innovative aspect of this study is integrating high-resolution remote sensing data (e.g., SAR imagery) and advanced hydrodynamic modeling coupled with historical flood records (1950–2022) to provide unprecedented insights into floodwater dynamics and spatial distribution. This study introduces an automated flood risk prediction tool capable of forecasting flood risks under future climate scenarios by leveraging machine learning techniques for a rapid application and analyzing historical meteorological data and Global Climate Model (GCM) outputs. Specifically, the study employs advanced machine learning methods, such as the Geometry-Adaptive Physics-Informed Neural Solver (GeoPINS), to enhance flood risk assessment and prediction in data-deficient areas. This novel approach addresses the gap in understanding flood behavior and sources of floodwater contributing to the Indus River and its tributaries.

It is essential to delve into the root causes of these floods, particularly the changing atmospheric conditions that contribute to their increasing frequency and severity. For instance, the 2022 flood was characterized by a unique set of atmospheric conditions that deviated from previous patterns, such as those observed during the 2010 floods (Fig. S1)27. Understanding these atmospheric drivers, including the role of intensified monsoon systems, is critical for developing effective flood management strategies. However, a notable knowledge gap exists in understanding the historical behavior of floods in Pakistan, including the types, frequency, and sources of floodwater that contribute to the Indus River and its tributaries. The country experiences various flood types, each with distinct impacts on homes, the economy, and infrastructure. For instance, the northern parts of Pakistan are prone to glacial lake outburst floods, debris flow, mudflow, and avalanches, while the central and southern regions experience flash floods and extreme rain events, each with unique impacts and damage patterns.

There are undoubtedly several published papers; however, the current studies have limitations, as only a few cases have been explored. For instance, Nanditha, et al.28 mainly focused on the flood-hit region of Larkana Plain. At the same time, Khan et al.24 evaluated livelihood vulnerability, Wang et al.25 focused on flood mapping, and Wang, et al.29 atmospheric circulations. None of these studies, however, describe the leading causes, historical behavior of the sources of the flood, and future challenges, including the main vulnerability in high-risk regions across the country.

This research aims to comprehensively examine the 2022 flood, evaluating its social and economic implications, causes, and future challenges. It will utilize historical flood data (1950–2022), remote sensing (SAR), meteorological data (1960–2021), and Global Climate Model (GCM) analysis, along with field validation in 58 districts, to compare the 2022 flood with previous major events, such as the 2010 flood. The study will also assess Pakistan’s future flood risk under the SSP5-8.5 climate change scenario and propose strategies for reducing the vulnerability of both infrastructure and communities to flood impacts. The SSP5-8.5 scenario represents a fossil-fueled development pathway where the world experiences high emissions, rapid economic growth, and minimal focus on environmental sustainability, leading to severe climate change impacts. This scenario provides a worst-case emissions benchmark for assessing extreme climate impacts. Additionally, other SSP scenarios, such as SSP1-2.6, SSP2-4.5, and SSP3-7.0, are considered to compare various emissions pathways and their implications for flood risks. Understanding the dynamics, causes, effects, and challenges of the 2022 flood is essential for developing effective mitigation strategies, enhancing resilience against climate change impacts, and informing policymakers and stakeholders. This approach aims to minimize the adverse effects of future floods in Pakistan, contributing to the nation’s sustainable development and resilience.

Results

Progression, impact, and advanced modeling of the 2022 pakistan floods

Between June and September 2022, Pakistan experienced a dramatic increase in floodwater coverage. Initially covering 3458 km² in late June, the inundated area expanded to approximately 16,000 km² by mid-August, peaking at 29,348 km² by the end of August (Fig. 1b, Table S2). Sindh Province was particularly affected, with 16,734.61 km² submerged, representing about 12% of its total area. The flood subsided in September, halving the inundated size compared to late August (Fig. 1c–f). The southern plains along the Indus River and the western mountainous regions were heavily impacted, affecting nearly a third of the country, with Baluchistan and Sindh provinces suffering the most.

Using spatial analysis, flood depths were computed in different periods (floodplains’ inundation depth, termed flood depths) (Fig. S2). Initially averaging 1.93 m in late June, the depth increased to around 3.37 meters by late August, with an average of 2.20 m across Pakistan. Notably, 80% of affected areas had depths below 1 m, but 6.51% of the area exceeded 5 m or more. In some regions, river channels caused depths to surpass 20 meters. The increased flood depths were also influenced by mountain runoff, sediment deposition, and local hydrodynamics, which caused river channels to change course and exacerbate inundation. The flood’s impact on infrastructure was profound: over 1.21 million houses were affected, 248.28 km² of infrastructure was inundated, and Baluchistan and Sindh had the highest damage. Around 1,784,126 ha of agricultural land were impacted, and over 800,000 animals were lost, which led to food insecurity for 14.6 million people. The reconstruction cost estimates were hampered by limited baseline data (Fig. S7e). A comprehensive recovery strategy emphasizing climate resilience and adaptation is crucial.

Furthermore, we employ an advanced physics-aware machine learning framework to achieve fast, accurate, and downscaled large-scale flood forecasting in Pakistan to simulate and analyze severe flood events. Specifically, we use the machine learning model from FloodCast30. To efficiently simulate the evolution of flood depths across Pakistan, we employ coarse spatiotemporal resolutions (480 m×480 m spatial resolution and 300 s temporal resolution) over 14 days from August 18th to August 31st. The flood inundation extents and depths generated by the machine learning model are validated using SAR-based flood data and traditional hydrodynamic-based results. Both traditional hydrodynamic methods and the advanced machine learning model demonstrate strong agreement during high water levels. However, comparative assessments with SAR-based flood depth data reveal that the advanced machine learning model outperforms traditional hydrodynamics, exhibiting fewer simulation errors. The model also better accounts for complex terrain variations and localized hydrodynamic interactions, which are critical in flood generation. Therefore, these advanced methods represent a promising tool for high-precision and real-time flood forecasts under climate change.

Main causes of floods

Geographical and climatic factors contributing to the 2022 pakistan floods

Pakistan’s geographical environment, situated at the convergence of the Eurasian and Indian tectonic plates and influenced by the South Asian monsoon, makes it highly vulnerable to intense flooding. The varied topography-from high mountains and riverine plains to arid deserts, is shaped by both tectonic activity and seasonal monsoon systems, which play a central role in the country’s flood dynamics. The primary contributor to flooding in Pakistan is the low-pressure monsoon depression systems originating from the Bay of Bengal during the summer months (July-September). In 2022, eight monsoon systems reached Pakistan, four particularly intense (Fig. 1a), resulting in exceptional rainfall and subsequent flooding.

These monsoon systems are crucial as they direct large amounts of moisture from the Indian Ocean toward southern Pakistan, amplifying the region’s precipitation patterns. The unusually high sea surface temperatures in the Arabian Sea and Indian Ocean during 2022 markedly increased atmospheric moisture levels. This increase in moisture content, combined with the monsoon’s convergence with Pakistan’s orographic features, produced extreme rainfall intensities central to the flooding dynamics. The Bay of Bengal monsoon depression systems, when they form, create low-pressure areas that draw moisture-laden air masses from the Arabian Sea toward Pakistan. This process greatly increases the amount of atmospheric moisture available to fuel intense rainfall, especially in regions with favorable orographic conditions, such as Pakistan’s southwestern and southeastern parts. In August 2022, a notably robust South Asian monsoon transported Indian Ocean moisture towards southern Pakistan. This effect intensified when a closed cyclone system formed over Sindh Province, acting as a conduit for moisture from the Arabian Sea into the areas bordering Sindh, Balochistan, Khyber Pakhtunkhwa (KPK), and Punjab provinces. Persistent low-pressure systems interacted with these elevated sea surface temperatures, amplifying monsoon activity and leading to unprecedented rainfall. This interaction resulted in abnormally heavy rainfall across these regions.

As illustrated in Fig. 2, intense precipitation along the Koh-e-Suleman and Kirthar mountain ranges and the Indus Plain caused rapid runoff and extensive riverine flooding downstream. The large-scale orographic lifting, in which moist air from the Bay of Bengal is forced upward by the mountains, leads to condensation and heavy rainfall, intensifying runoff and river flooding. This flooding was further intensified by the high-magnitude flows from numerous mountain streams in these regions, contributing to large-scale inundation and the catastrophic flooding observed31. In particular, steep slopes in these regions contributed to rapid water flow, which overwhelmed flood defenses and natural drainage channels when combined with excessive rainfall. This rapid runoff, driven by intense rainfall and the steep terrain, led to flash floods in mountainous and riverine areas, further contributing to widespread devastation. The combination of Pakistan’s unique geography and the powerful monsoon system of 2022 underscores the role of climate-driven precipitation patterns and topographical influences in exacerbating flood risks.

Fig. 2: The July and August 2022 spatial distribution of precipitation and its historical anomaly.
figure 2

Spatial distribution of precipitation (mm) in July and August 2022, compared with the corresponding historical period in Pakistan: (a) July 2022 precipitation; (b) August 2022 precipitation; c and d) Precipitation anomalies between July and August 2022 and the historical mean (mean of 1990–2021).

Full size image

Factors behind the unprecedented July–August 2022 rainfall

The extremely heavy precipitation during July and August 2022 was a critical factor in the floods, particularly in Pakistan’s southern regions (Baluchistan and Sindh). National average rainfall in August 2022 was 243% above normal, with Balochistan and Sindh experiencing unprecedented increases of 590% and 784%, respectively. This marked the wettest August on record, with rainfall eight to nine times higher than usual in these areas. Additionally, the extent and duration of monsoon activity increased, particularly over the typically arid Balochistan province, where rainfall usually ranges between 0-5 mm during the monsoon season. In 2022, however, southern parts of Balochistan received their highest monsoon rainfall in the past 60 years.

In July, concentrated heavy precipitation occurred across Pakistan, especially in Sindh and Balochistan (Fig. 1a), with specific areas experiencing over 300 mm of rainfall, significantly exceeding the 1990–2021 average (Fig. 2a–c). This pattern persisted into August, with Sindh receiving 180–200 mm, nearly double its historical average (Fig. 2b–d). Similarly, the central region of Balochistan received double the typical rainfall for this period (Fig. 2c, d). Additional analysis suggests that unusually high sea surface temperatures in the Arabian Sea and Indian Ocean during 2022 increased atmospheric moisture levels, creating ideal conditions for intense monsoon activity in Pakistan. This moisture excess, compounded by atmospheric instability, intensified rainfall events, leading to the observed deluge. The high sea surface temperatures facilitated increased evaporation rates, enriching the moisture content in the atmosphere, which then contributed to heavy rainfall once these moist air masses encountered Pakistan’s mountainous regions.

Further contributing to this extreme rainfall were persistent low-pressure systems that formed over the Indian Ocean and moved toward Pakistan, drawing in substantial moisture from the Arabian Sea. These systems intensified in August, creating a closed cyclone over Sindh, which channeled even more moisture toward the southern regions, particularly Balochistan and Sindh. The cyclone system’s structure and slow movement allowed for continuous moisture convergence, resulting in prolonged heavy rainfall over southern Pakistan. This combination of elevated sea surface temperatures, persistent low-pressure systems, and monsoon currents explains the extreme rainfall observed during the 2022 flooding season. The coupling of atmospheric conditions with increased surface moisture loading highlighted the compound nature of the flood-producing mechanisms. This cyclone-like system and low-pressure activity over the Arabian Sea helped funnel moisture-laden air masses, amplifying rainfall and exacerbating flood risks in typically drier areas.

Topographical influence on flood water sources to the Indus River

The Indus River experienced a series of floods in 2022, with hydrological data revealing distinct patterns across various monitoring stations. At upstream stations, such as Tarbela and Kalabagh, flood peaks occurred multiple times in July, with Tarbela’s peak discharge reaching 9000 m³/s and Kalabagh’s 12,000 m³/s (Fig. 3a). The Tarbela peak was primarily driven by glacial melt, though this contribution was within average levels, suggesting limited glacier influence during the peak flood period. By late August, Kalabagh’s discharge contributed 50% to the total flow in the mid and lower Indus reaches.

Fig. 3: The central flooded region, the flood level, and its topographic drainage flow.
figure 3

Panels (a) and (b) depict observed discharge data from hydrological stations along the Indus River, with the red dashed line indicating the maximum flood level at the barrages. The barrage sequence is shown in panel (c). The panel (e) illustrates topographic changes and hanging river landforms in the alluvial diluvial flooding areas of southern Pakistan’s lower reach of the Indus River. (The base map in panels (c) is from the USGS, and the image in panel (d) is from a satellite image).

Full size image

In the mid-Indus region, Cheshma and Taunsa stations recorded peak flows between 12,000–14,000 m³/s by late July and even higher rates, 16,000–18,000 m³/s, by late August (Fig. 3b). Downstream barrages like Guddu, Sukkur, and Kotri exhibited similar trends, with Guddu and Sukkur experiencing consistent flows of 16,000-18,000 m³/s. These barrages, located in low-lying regions, were heavily impacted by localized rainfall and inflows from surrounding topographies, particularly the Kerther mountain ranges. This inflow and extensive monsoon rainfall caused severe flooding in populated areas near the Guddu and Sukkur barrages.

The flood’s peak occurred on September 4th, with the most intense flooding observed between Chashma and Kotri barrages. At Kotri downstream, peak discharge was recorded on September 4th, subsiding by September 12th. The unique topographical structure of the Indus River basin, including a low gradient of about 0.2‰, promotes sediment accumulation and impedes drainage, contributing to significant water retention and flood conditions in the Larkana plains. The extensive flow from over 2100–2150 streams from central mountain ranges resulted in substantial floodwater volumes among 1250, generating high discharging rates between 3000–4000 m³/s, which account for 58% of total streams in the region. This unique confluence of heavy precipitation and complex topography led to a 100 km inland lake in the Larkana plains (Fig. 1a, 1c–f). The low-lying western floodplain, with a mean water level of 5.5 feet in cross-sections AA’ to CC’, was particularly vulnerable to floodwater overflow. In contrast, the eastern floodplain faced less severe inundation (Fig. 3d, e). Due to this lower elevation and poor drainage in the west, excess runoff formed the inland lake observed at the Larkana plains.

Contribution and devastating impacts of breaching small check dams

Pakistan, reliant on the Indus River, prioritizes floodwater conservation, especially in Balochistan, its driest province. To enhance groundwater recharge and retain agricultural water, 1012 check dams were constructed. However, the 2022 floods severely damaged 177 of these dams, with 69 completely breached (Fig. 4a, red points) and others partially breached (orange points). Most breaches occurred in the Quetta division, triggered by overtopping (62%), seepage (18%), and foundation erosion (20%) (Fig. 4b–d; Table S3). For example, the breach of Machka Dam 1 released 33.0 × 104 m3 of water from its pre-flood capacity of 144.96 × 104 m3, exacerbating downstream flooding.

Fig. 4: The number of beaching and breaching models of check dams.
figure 4

(a) The figure illustrates the number of small check dams that were breached or partially breached in each district of Balochistan province due to extreme monsoon rainfall in 2022. High dam breaches were identified in the most populous districts, and their impacts are also illustrated per district. In panels (b–d), the three breaching models are presented in response to extreme rainfall, with blue arrows indicating the flow direction and red areas representing the breaching dimensions. Photograph taken by Dr. N.A. Bazai.

Full size image

Overtopping breaches caused immediate and severe downstream impacts by releasing large volumes of water over dam crests. Seepage and foundation erosion breaches led to delayed failures by undermining dam bases through prolonged infiltration. These breaches destroyed 590,440 acres of agricultural land (categorized as 1–20,000, 20,000–50,000, and >50,000 acres), damaged 346,303 homes (categorized as low, moderate, and high damage), and killed 292,526 livestock. Total damages amounted to PKR 349 billion (US$ 1.625 billion). Additionally, 238 people died, 106 were injured, and telecommunication and electricity systems collapsed for nearly a week.

Floodwaters merged with major rivers, including the Zhob, Mula, and Hingol, which were already swollen by 2022 monsoon rainfall that exceeded 450% of Balochistan’s annual average. This overwhelmed dam capacities, causing widespread inundation. 38 districts declared calamities, with 35 severely affected by dam breaches. Affected populations (Fig. 4) were categorized as 0–0.7 million, 0.7–1.5 million, and 1.5–2.5 million.

Field observations identified overtopping as the primary breach mechanism, rapidly increasing river flows such as the Hub, Zhob, Bolan, Mula, Porali, and Hingol. Seepage and foundation erosion compounded floodwaters, intensifying disaster impacts. Approximately 30% of the dams were constructed on unsuitable sites, amplifying flood risks. Sudden breaches of small check dams caused extensive damage within a 1–2 km buffer zone, demolishing infrastructure and intensifying downstream flooding. The high-intensity rainfall overwhelmed dam capacities, with breaches contributing to 80–85% of the total economic losses. Dam breaches caused 75% of agricultural damage, while heavy rainfall accounted for 25%. This marked the first-time locals experienced such large-scale dam failures, as check dams are relatively new to the region, leaving communities unprepared for the scale of destruction.

Discussion

Pakistan has experienced 31 catastrophic floods since 195032, occurring roughly every two years and causing substantial economic and human losses, as illustrated in Figure 05a–e and Table S1132,33. On average, each flood has resulted in over 500 fatalities and $4000 million in financial damages. The 2022 floods, however, were historically devastating, causing approximately $40 billion in losses and affecting a third of the country’s land area, as detailed in Table S4. This disaster was primarily due to hefty monsoon rains in the southwest, exacerbated by local topography and inadequate flood management. The extreme rainfall in the central Hindukush mountains, including the Kerther and Koh-e-Suleman ranges, triggered flash floods and debris flows, which merged into tributary rivers and significantly increased water flow into the lower Indus region. This phenomenon contributed to localized surges in water levels and sediment transport, intensifying downstream flooding. The situation was compounded by the low-lying floodplain of the Indus River and the collapse of 177 check dams in Pakistan’s driest province, resulting in 1739 deaths.

In 2022, Pakistan faced catastrophic floods, marking the worst disaster in its 75-year history, with economic impacts exceeding $40 billion. A staggering 33 million people were affected, 8 million displaced, and the agriculture, food, livestock, and fisheries sectors suffered damages amounting to $3.7 billion (Fig. S7a-e and Table S4). These floods differed greatly from the 2010 floods, covering one-third of the country and causing unprecedented destruction Figure 05e. The 2010 floods were primarily riverine, linked to glacier melt and monsoon rains along the Indus River. In contrast, the 2022 floods were driven by urban floods and flash floods from Baluchistan (e.g., streams originating in the Koh-e-Suleman, Marri-Bugti hills, and Kirther ranges). Lower Hindukush Mountain contributed, with glacier melt playing a lesser role (Fig. 5e). Particularly, the Koh-e-Suleman, Marry-Bugti hills, and Kirther range saw a considerable number of streams, around 2200 to 2500, causing severe flooding and damage. With discharges of some, these streams reached 6000–7600 m3/s, but an average of 3000 to 4000 m3/s flows over a month, inundating western regions of the Indus River, resulting in significant casualties and destruction.

Fig. 5: Comparison of historical floods to 2022 and their behavior in 2010.
figure 5

Panel (a–d) shows the chronological sequence of Major Flood Events in Pakistan, including the Number of Deaths, Affected Areas, Affected People, and Economic Losses. Panel (e) shows the primary source of the two floodwaters of the mega-flood in 2010 and 2022. In 2010, the flood mainly originated in northern Pakistan (Gilgit-Baltistan Province) and Panjab Province. However, the flood in 2020 primarily derived from the Koh-e-Suleman Range and the Kerther Range. (The base map in panel (e) is from the USGS).

Full size image

Deforestation was critical, leading to high runoff and debris flow, particularly along the Swat River. For instance, a notable debris flow from the Bodai Kamar gully in Swat obstructed the river for approximately four hours. This event, caused by a temporary dam (29–30 m high), breached and released massive sediment loads, intensifying downstream damage. These debris flows severely impacted the agricultural regions along the Swat and as well as Indus rivers, known for their intricate irrigation systems and dense populations. Massive sedimentation and an average of 1.8 to 2 m were identified between DG Khan district and Larkana plain (refer to Fig. S8a–c), leading to unprecedented economic losses and posing a formidable challenge to future flood management and recovery efforts. The upper Indus River is characterized by narrow dimensions, steep gradients, and rapid flow, leading to substantial sediment migration and deposition at lower reaches and elevated riverbeds34. This unique topography contributes to frequent and severe flooding, notably the 2010 flood. Additionally, the trumpet-shaped terrain in the south (Fig. 3c–e) increases the susceptibility of areas like the Larkana floodplain to prolonged inundation during major floods.

Adequate drainage and flood mitigation require comprehensive water resource management across the entire basin, as seen in successful initiatives like wetland restoration in the upper Mississippi River35,36 and soil and water conservation in China’s Wuding River basin37,38. Despite having extensive reservoirs and barrages, the Lower Indus Plain needs help with maintenance issues and high sediment deposits from floods. Encroachments along waterways in densely populated areas, such as the Larkana floodplain, have led to inadequate drainage and the formation of a 100 km2 flash flood lake, the largest in Pakistan’s history (see Fig. 3c, d). Current hydraulic projects are insufficient against mega-floods (Table S5), as seen in the 2022 flood event, where runoff far exceeded the design capacity of sluice gates and other engineering designs failed, including flood conservation projects were unable to hold the flood water and breaching 177 dams (Fig. 4a). This study proposes new floodwater retention and drainage systems considering sediment migration patterns and the increased flow rates projected under future climate conditions. These innovations could improve Pakistan’s flood resilience and provide a blueprint for other regions facing similar challenges.

Global warming intensifies flood risks in the area due to extreme monsoons and increased snowmelt39,40. A key innovation in this study is the integration of hydrodynamic modeling with data from Global Climate Models (GCMs) to project future flood risk under various climate scenarios, such as SSP5-8.5. GCMs from the Coupled Model Intercomparison Project (CMIP6) indicate that all scenarios of SSP126, SSP245, SSP370, and SSP585 project rising summer precipitation, which is a primary source of floods in Pakistan (Fig. S9a-l and details in the Supplementary file). The most pronounced trend of heightened intensity is observed in SSP585, showing an approximate increase of 0.065 mm/10a24,41. Summer rainfall constitutes Pakistan’s primary source of floods39. Researchers combined precipitation data with topographic factors like elevation differences in river channels to assess flood risk under changing climate scenarios. The findings indicate that over the next 80 years, regions with high flood risk will be concentrated in southern Pakistan and at the Himalayan-Indian plains junction in Sindh and northern Punjab Province (Fig. 6). The gradual expansion trend of relatively high-risk and high-risk zones due to climate change is evident, with the area of these zones increasing from 6.11% between 2015 and 2020 to 6.62% (an area increases of 4975.04 km²) in 2061–2080 (Fig. 6 and Table S6). Flood risk prevention and management strategies in Sindh and Punjab should be prioritized. This comprehensive approach enables a more precise assessment by integrating elevation and precipitation data, marking an advancement in flood prediction technologies.

Fig. 6: Time series of future flood risk zones in Pakistan from 2021 to 2100.
figure 6

Panel (a) shows the extent of the 2022 flood, while panels (b–d) depict the predicted risk maps for the periods 2021–2040, 2041–2060, 2061–2080, and 2081–2100, respectively. (The base map in panels is from the USGS).

Full size image

The flat terrain of the lower Indus Plain, with its dense population and human activity, is particularly vulnerable to flooding42,43,44. Most housing structures in Pakistan are masonry-timber buildings lacking adequate flood protection (Table S7). Many buildings constructed within 1 km along the banks of the Indus River and streams show relatively high physical vulnerability and lack adequate flood protection45,46. Social resilience to flooding is low, with approximately 50% of households needing flood education47. Schools in Pakistan also need more adequate emergency preparedness. Only 52% have emergency plans (Table S8), 38% have emergency equipment, and 58% have conducted comprehensive emergency training (Tables S9 and S10). These statistics highlight the need for improved flood disaster management strategies and disaster knowledge education44,47.

In Pakistan, persistent flood risks under climate change demand enhanced strategies for risk management and mega-flood handling. Disaster resilience hinges on addressing climate impacts, bolstering scientific knowledge, and implementing flood mitigation. Challenges in post-disaster recovery include climate impacts on flood risk, disorganized urban planning, housing vulnerability, lifeline project design, water pollution, weak emergency management, and insufficient early warning systems. Strengthening climate change research and disaster preparedness is crucial for tackling extreme weather conditions. Establishing forecasting systems and enhancing response capacity is imperative for sustainable development in vulnerable regions40,48.

Against this backdrop, understanding and addressing Pakistan’s 2022 mega-flood is paramount, especially for nations with limited economic resources. The flood in the southern and central regions presented unprecedented levels, reaching 16,000 to 18,000 m3/s in the Indus River. This marked a significant departure from the 2010 event, primarily attributed to prolonged monsoons. The aftermath of the 2022 flood was staggering, impacting 33 million people, causing extensive damage to 1,784,126 hectares of agricultural land, resulting in over 1739 casualties and inflicting economic losses surpassing $40 billion.

The root causes of this catastrophic flood lie in the convergence of intense Indian monsoons, leading to extraordinary rainfall in Baluchistan and Sindh, surpassing norms by +243%. The impact was exacerbated by the destruction of the water conservation project and the failure of 177 check dams. Weaknesses in flood management and check dam design, coupled with deforestation in the Hindukush ranges, likely contributed to the increased vulnerability by intensifying runoff and debris flows. An analysis of seven decades of historical flood data reveals two dominant flood types: riverine floods from the upper region (as seen in 2010) and scattered floods from the central and southern areas (as witnessed in 2022). Projections under the SSP5-8.5 climate scenario underscore an increasing trend in precipitation, particularly in high-risk areas, with potential simultaneous occurrences of both flood types posing heightened risks7.

Addressing the consequences of extreme climate events demands immediate and strategic action. The imperative mandate includes implementing an early warning system, meticulous execution of evacuation plans, and engineering interventions for riverbanks and sediment migration management. In this study, we adopted advanced machine-learning methods, such as the Geometry-Adaptive Physics-Informed Neural Solver (GeoPINS) within the FloodCast framework, demonstrating notable improvements in simulating flood dynamics, especially in data-deficient regions. These advanced methodologies enable more precise flood predictions, enhancing disaster preparedness and real-time response capabilities49. Furthermore, developing innovative floodwater conservation systems tailored to the topographical and climatic specificities of the Indus River basin is essential for mitigating future risks. These measures are advisable and critical for building a resilient and sustainable future. By focusing on scientific understanding, efficient mitigation, and strategic development, we can significantly reduce the impacts of climate-induced disasters and strengthen resilience in vulnerable regions like Pakistan.

Future research should focus on integrating advanced hydrodynamic modeling with more localized climate projections to develop more accurate flood risk assessments at the regional and district levels. Additionally, studies could explore the effectiveness of different flood retention structures, sediment management techniques, and the role of reforestation in reducing runoff and flood intensity. Researchers should further evaluate the potential of state-of-the-art machine-learning approaches, such as GeoPINS, for improving real-time flood prediction and response systems, particularly in regions with limited data availability. Further interdisciplinary collaboration is required to address the social aspects of flood risk, particularly the role of community resilience, education, and preparedness in mitigating the impacts of future mega-floods.

Methods

Acquired data

This study utilized a variety of data sources to analyze the 2022 mega-flood in Pakistan comprehensively. A total of 372 Synthetic Aperture Radar (SAR) images from 31 frames were obtained from the Sentinel-1 satellite via NASA’s Earth Data platform (https://search.earthdata.nasa.gov). Visual assessments of flood impacts were supported by optical imagery from Gaofen-2 and Google Earth. Flood runoff data was sourced from Pakistan’s National and Provincial Disaster Management Authority (NDMA and PDMA), Irrigation and Water Resources Departments (IWRD), and Flood Forecasting Division (FFD), while additional data on dam breaches and infrastructure damage came from the Baluchistan Irrigation Department (BID). Field surveys across 58 flood-affected districts provided qualitative data on stream bank damages, flood levels, sedimentation, and flood extent, supplemented by approximately 900 GPS-sampled points. Meteorological data were provided by the Pakistan Meteorological Department (PMD) (https://nwfc.pmd.gov.pk/new/monthly-reports.php) and included precipitation patterns analyzed using the Multi-Source Weighted-Ensemble Precipitation (MSWEP) dataset (http://www.gloh2o.org/mswep/), combining gauge, satellite, and reanalysis data. Future flood risk under climate change scenarios was assessed using data from the Coupled Model Intercomparison Project (CMIP6) (https://esgf-node.llnl.gov/search/cmip6/).

Remote sensing data

The rapid organization and dissemination of primary data play a crucial role in monitoring and assessing flood disasters. This study utilized 372 Synthetic Aperture Radar (SAR) images from 31 frames obtained from https://search.earthdata.nasa.gov. In seven orbits of the Sentinel-1 satellite (refer to Table S1 for acquisition times). For the two SAR images, S1 and S2, acquired at different times after registration, the neighborhood ratio model can be defined as Eq. (1):

$$NR(r,c)=1-min\left(\frac{\sum {f}_{w}({S}_{1(r,c))}}{\sum {{f}_{w}}_{({S}_{1}(r,c))}\text{'}}\frac{\sum {f}_{w}({S}_{2}(r,c))}{\sum {f}_{w}({S}_{1}(r,c))}\right)$$
(1)

where NR refers to the log ratio chart; r and c refer to the line number and row number of SAR images, respectively; min refers to minimization; fw refers to smoothing filter functions (such as average filtering and Gaussian filter) of images at the position (r, c) with the filtering window w. The initial view of each frame during the monitoring period served as the primary image. Enhancing water area information in the SAR image was achieved using a multi-view ratio of 10:2 and the neighborhood ratio model (Fig. S6). A threshold of 1.8 was applied to extract changes in water areas during various periods relative to the beginning of May in Pakistan (see Fig. 1b and S2). Furthermore, visual interpretation and human-computer interaction were utilized to assess damages to roads, buildings, and crops in flood-affected areas based on optical images from Gaofen-2 and Google Earth. Flood runoff data were sourced from Pakistan’s Provincial Disaster Management Authority (PDMA).

Field data and discharge measurements

The field data in this study, primarily qualitative, covers the entire country’s post-flood event. Collected during a field trip across 58 affected districts nationwide, information was gathered from residents along major thoroughfares. Investigative interests included assessing stream bank damages, determining stream base flow, evaluating stream flood levels, analyzing sedimentation patterns pre- and post-flood, and mapping flood extent using GPS points (G639; Accuracy: Single: 1–3 m; SBAS: 0.6 m). Flood depth, the primary cause of check dam breaches, and the status of 177 dams (partially and completely breached) among 42 dames were investigated during the field visit in September 2022 (the rest of the dam data was obtained from the Balochistan irrigation department and local authorities) and during the visit also compiled the dam and lake hights flood level and breaching points dimensions. These data were applied to Eq. (2), adapted from Froehlich’s embankment-dam breach model, to estimate each dam’s peak outflow:

$${Q}_{p}=0.298\cdot {\left({V}_{w}\right)}^{0.312}{\left({H}_{w}\right)}^{1.417};{R}^{2}=0.917$$
(2)

Where Qp is the peak breach discharge (m³/s), Vw is the reservoir volume above the breach bottom (m³), and Hw is the water depth above the breach bottom at failure (m).

This equation aligns with Froehlich’s peak outflow formulas for embankment dams. The details are in Tables S12 and S13, widely cited in dam breach studies50,51. Additional data were obtained from Baluchistan’s irrigation department (see Fig. S2and Table S3). Approximately 900 sample points were surveyed, using GPS for latitude and longitude acquisition, and images were captured from various angles of the affected areas. Information on water levels, casualties, agricultural losses, and infrastructural damage was collected through local interactions or trial-and-error methods. A risk awareness and education survey covered the most affected districts in provinces such as Sawat, Nowshera, Charsadda, DI Khan, DG Khan, Rajanpur, Quetta, Killa Saifullah, Larkana, Jakababad, Naseerabad, and Jaferabad. Beyond field data, insights were gathered through interviews with locals, former residents, local news, government departments, and select NGO sectors. Meteorological data were provided by the Pakistan Meteorological Department (PMD), and additional hydro-meteorological and geo-hazard data were sourced from the Irrigation and Water Resources Departments and the National Disaster Management Authority (NDMA). All photographs used in the supplementary materials, including Fig. S8, were taken by the authors during the field visit. Individuals appearing in the photos (e.g., Dr. Bazai’s team members) are part of the research team and have given their written consent for the publication of these images.

Meteorology data and process

This study relies on the Multi-Source Weighted-Ensemble Precipitation (MSWEP) dataset to analyze precipitation patterns in Pakistan and drive a hydrological model for simulating flooding (Fig. S9a-l). The MSWEP dataset combines gauge, satellite, and reanalysis data to provide high-quality precipitation estimates52 at a 3-hourly resolution and 0.1° granularity, available from 1979 to nearly real-time. Access the data at http://www.gloh2o.org/mswep/. Future flood risk under various climate change scenarios is computed using precipitation data from the Coupled Model Intercomparison Project (CMIP6). Find more information on CMIP6 data at https://esgf-node.llnl.gov/search/cmip6/.

Flood risk assessment method

Spatial and spectral analysis identifies the shape, size, graphics, shadow, and texture of elements at risk in remote sensing images. Google Earth’s high-resolution data aids hazard-affected object interpretation due to its large-scale capabilities. Verification through field investigation extracts object types, quantities, and spatial distribution from the flooded area’s remote sensing images. These data facilitate flood impact analysis and vulnerability calculation. Initial interpretation data is improved using Arcgis to rectify topological errors, yielding final datasets that include affected buildings and roads. Using spatial analysis and DEM, the flooded area’s maximum height (Hmax) is derived as the water surface elevation. Flood depth (Hdepth) for each grid is computed as the difference between the grid elevation (H) and water surface elevation (Hmax) using Eq. 3.

$${H}_{{depth}}={H}_{\max }-H$$
(3)

Flood hazards (HFlood) are quantified using flood depth, while vulnerability (V) is determined by the ratio of affected building area/road length and total building area/road length. Flood disaster risk (RFlood) is calculated as HFlood × V using Eq. 4. Normalized RFlood is categorized into five levels: very high (the result is 0.2–1), high (0.1–0.2), medium (0.05–0.1), low (0.01–0.05), and very low (0–0.01).

$${R}_{{Flood}}={H}_{{Flood}}\times V$$
(4)

Note: Here, V is a dimensionless ratio (0–1) of affected area (or road length) to total area (or length). Thus, the product HFlood × V is a simple scalar multiplication of flood depth (in m) by a vulnerability factor, yielding a flood-risk index that retains the physical units of depth but is weighted by exposure.

Flood risk prediction

This process predicts future flood risk trends under climate change scenarios. The area threshold technique identified major rivers in Pakistan and divided them into 5 km segments with consistent slopes. Flood susceptibility was quantified by extracting the height difference of grid points within 20 km of each river’s section from its riverbed: more minor differences indicated higher flood potential. Summer precipitation acted as the climate driver, while the height difference from the riverbed was the topographic factor. By integrating the topographic factor and flow accumulation weighted by summer precipitation (climate driver), this analysis forecasts the trend of flood hazards under future climate changes. Subsequently, the current building area was employed to quantify vulnerability. The conjunction of future flood hazards and vulnerability enables the prediction of future flood risks. The details of the future prediction are displayed in the flowchart (Figs. S5 and S10).

Flood risk assessment and prediction using advanced machine learning methods

Flood risk assessment and prediction often rely on various hydraulic or hydrodynamic models. Accurate flood inundation predictions are conventionally achieved by implementing two-dimensional hydrodynamic models that solve the shallow water equations (SWEs). In this study, we employed a two-dimensional hydrodynamic model based on the finite difference method to solve the SWEs, incorporating a well-balanced weighted essentially non-oscillatory (WENO) scheme to enhance accuracy and stability in complex terrains. Despite their accuracy, traditional hydrodynamic techniques are heavily influenced by grid resolution and are computationally expensive. To address these challenges, we integrated physics-based principles with machine learning algorithms, developing a hybrid approach that optimizes computational efficiency while maintaining predictive accuracy. This integration allows for effective scalability in hydrodynamic modeling, enabling more efficient flood risk assessments53.

We employ the advanced physics-informed machine learning approach, Geometry-Adaptive Physics-Informed Neural Solver (GeoPINS), within the FloodCast framework30 to achieve a fast, stable, accurate, downscaled, and geometry-adaptative flood assessment and prediction, as illustrated in Fig. 7a. This machine learning method uses input channels composed of coordinates, time domain, and initial conditions of flood height, which are first lifted to a higher-dimensional representation by a multi-layer perceptron (MLP). Several Fourier layers then extract efficient spatiotemporal representations, and the outputs (flood depth) are obtained via MLP projection. GeoPINS is trained using a geometry-adaptive physics-constrained loss function, which includes physics loss for controlling the residuals of 2-D depth-averaged SWE and data loss for measuring the mismatch between the NN prediction and the initial/boundary conditions. This physics-informed machine learning method provides several advantages, including working with incomplete models and imperfect data, robust generalization with limited data, particularly in flood prediction for data-deficient areas, and an improved understanding of deep learning through the physics-constrained optimization process.

Fig. 7: Flood risk assessment and prediction using advanced machine learning methods.
figure 7

a Architecture of the advanced physics-informed machine learning method used in FloodCast. b Study area of the 2022 Pakistan flood. c SAR-based flood extent on 30 August 2022. d Flood extent is based on the traditional hydrodynamic method (Finite Difference) (flood depth ≥ 0.1 m). e Flood extent is based on an advanced machine learning method (flood depth ≥ 0.1 m). f Comparison between SAR-based flood extents and flood extents from the traditional hydrodynamic method (flood depth ≥ 0.1 m). g Comparison between SAR-based and machine learning-based flood extents (flood depth ≥ 0.1 m). h Flood extent is based on the traditional hydrodynamic method (flood depth ≥ 0.5 m). i Flood extent based on an advanced machine learning method (flood depth ≥ 0.5 m). j Comparison between flood extents from the traditional hydrodynamic method and the machine learning method (flood depth ≥ 0.5 m).

Full size image

We utilize advanced machine learning methods to predict the severe floods that occurred in Pakistan between August and September 2022, primarily caused by heavy rainfall. Fig. 7c–j presents the flood inundation extents on August 30, obtained by different methods (SAR, traditional hydrodynamic models, and machine learning-based approaches) and their comparative analysis. The visualization results reveal that deterministic hydrodynamics-based flood extents align better with SAR-based results in peripheral regions. In contrast, the advanced machine learning method-based flood extents are accurate within inundated areas, especially in detail. By leveraging the predicted flood depths generated through advanced machine learning techniques, high-precision and real-time flood risk assessment and prediction can be effectively achieved.