A reverse causality also operates growing urban settlements increasingly drive flood risk by encroaching on floodplains. Rapid development replaces natural buffer zones with impermeable surfaces, boosting runoff and exposure. In Malaysia, scholars note that uncontrolled urban growth “has destroyed pervious surfaces” and significantly heightened flash-flood vulnerability (Kumaresen et al., 2025). (How can Changes in the Human‐Flood Distance Mitigate Flood Fatalities and Displacements?, 2023) quantified a global pattern: in more than half of examined countries people are moving away from rivers to reduce casualties, yet where flood-protection infrastructure exists, communities stay closer and “human–flood distance” even decreases. In other words, while some retreat occurs, new construction often proceeds right into formerly risky areas under the illusion of safety. Case studies illustrate this trend: as Greater Kuala Lumpur expands, natural floodplains are “replaced with concrete structures,” drastically reducing water absorption and amplifying runoff during storms (Abid et al., 2024). The 2021–2022 Malaysian floods underscored this dynamic by devastating areas that had been heavily developed for housing and agriculture. Thus, in Padang Terap the spatial flood vulnerability identified by hotspot analysis is likely reinforced by settlement patterns: villages and farms have sprawled onto low-lying terrain and riparian zones. This matches the global finding of (How can Changes in the Human‐Flood Distance Mitigate Flood Fatalities and Displacements?, 2023) that urban sprawl in floodplains intensifies flood exposure. In short, increased urbanization in Malaysia not only suffers from floods but actively exacerbates them by encroaching on natural drainage corridors (Abid et al., 2024).

Padang Terap’s flooding can be contrasted with other regional hotspots to highlight distinctive pressures. In Kelantan (Peninsular Malaysia’s northeast coast), flood risk is driven by seasonal Northeast Monsoon rainfall that swells the Kelantan River basin. In fact, Malaysia “experiences seasonal floods, particularly during the Northeast Monsoon” that heavily impact Kelantan (among other states) (Abid et al., 2024). Like Padang Terap, Kelantan’s threat is mostly hydrometeorological: it stems from intense rainfall over floodplains. However, Kelantan is even more exposed due to its location on Malaysia’s flood-prone east coast, and studies have documented severe inundations and erosion in its river network. In contrast, Jakarta’s floods illustrate a heavily urbanized context. Continuing urbanization in Jakarta has dramatically increased flood vulnerability especially in river-adjacent slum areas (Nasution et al., 2022). (Nasution et al., 2022) report that as floods become routine in Jakarta, unabated urban sprawl (often on reclaimed or subsiding land) leads to routine inundation, with more severe floods occurring when large rivers overflow. The March 2025 Jakarta disaster (over 100,000 people displaced by 1–3 m of water) exemplifies how dense development and aging infrastructure amplify climate threats. By comparison, Padang Terap’s setting is rural: its flood hotspots arise from upland rainfall and river dynamics more akin to Kelantan’s monsoonal pattern, rather than Jakarta’s fast-flash, drainage-limited regime. This suggests that while all three regions face flood exposure and development pressures, the underlying mechanisms differ. Kelantan and Padang Terap share pronounced hydroclimatic drivers and contiguous floodplain inundation, whereas Jakarta’s hazard is worsened by impervious surfaces, land subsidence, and inadequate urban drainage ((Abid et al., 2024); (Nasution et al., 2022)). Recognizing these differences is critical for resilience: Padang Terap (and Kelantan) may benefit most from catchment-level land management and green infrastructure, whereas Jakarta-like contexts require aggressive urban planning, drainage upgrades, and subsidence control.

In 2010, the Padang Terap District in Kedah was significantly affected by flooding, with 11 sub-districts impacted (Table 1). Although specific evacuation figures for Padang Terap were not detailed in available reports, statewide data indicated that more than 28,000 people had been evacuated in Kedah as of November 4, 2010, with conditions in Padang Terap reportedly beginning to improve at that time (Star, 2010). Despite the lack of detailed statistics for that year, Padang Terap has continued to face severe flooding in subsequent years. For example, in November 2024, a total of 1,298 individuals from 386 families were reported to be affected by floods in the district (Hilmy, 2024). According to data from the Department of Social Welfare (JKM), a total of 1,427 families were registered as victims during the 2010 flood. For the purposes of this study, a sample size of 680 individuals representing 47.7% of the affected population was selected. It is important to note that not all those registered with JKM were considered respondents, as some were stranded due to the flood but did not experience direct flooding of their homes. Consequently, only those whose residences were inundated were included in the study sample.

According to the research findings, recorded floodwater levels were observed at 0.3 meters, 2.0 meters, 2.5 meters, and 3.7 meters. These measurements indicate the vertical height of floodwaters in relation to the highest structural point of the affected houses, typically the floor level or platform of the living area. A flood level of 0.3 meters suggests minor inundation, likely affecting only ground-level areas such as yards or steps. However, levels reaching 2.0 meters or more signify severe flooding, with water submerging significant portions of the house, including interior spaces, furniture, and electrical systems. The highest recorded level of 3.7 meters implies a catastrophic flood event, likely overwhelming entire homes and posing serious threats to safety, property, and livelihood. These data points underscore the varying degrees of flood impact and highlight the vulnerability of residential structures depending on their elevation and proximity to flood-prone areas (Said, 2017).

Padang Terap, located in the northern state of Kedah, Malaysia, represents a quintessential example of a district characterized by a mosaic of agricultural landscapes interspersed with rural settlements (Figure 1). The district’s socio-economic fabric is largely shaped by its agrarian economy, yet it remains highly vulnerable to seasonal flooding events, particularly during the Northeast Monsoon period, which typically spans from November to March (Said, 2017). The principal drivers of flooding in Padang Terap are multifaceted, involving both natural and anthropogenic factors. High-intensity rainfall associated with the monsoonal system exerts substantial hydrological pressure on local river systems, most notably the Sungai Padang Terap, which frequently overflows its banks, inundating adjacent settlements and farmlands. Moreover, the district's hydrological vulnerability is exacerbated by the inadequacy of drainage infrastructure, which is often insufficient to cope with the volume and velocity of stormwater runoff during peak rainfall events. Topographic elements also play a critical role; large expanses of low-lying terrain within Padang Terap act as natural basins that accumulate water, prolonging the duration of flood events and heightening the risks to both human livelihoods and agricultural productivity. These compounded factors not only amplify the frequency and severity of flood occurrences but also reveal significant challenges in local flood management strategies, highlighting a pressing need for integrated, adaptive, and resilient planning approaches that incorporate hydrological modeling, sustainable land use practices, and community-based disaster risk reduction initiatives (Said et al., 2024).

Padang Terap (Figure 1), a district in Kedah, Malaysia, spans approximately 1,338.70 square kilometers and is the second-largest district in the state after Sik. Bordered by Thailand to the north (at Durian Burung), Bekort to the east, and Pokok Sena to the south, the district is a vital hub for administration and agriculture. Kuala Nerang, the main town, functions as a key transit point for travelers from Alor Setar and Pokok Sena heading toward Durian Burung. It also hosts the bustling Pekan Nat market, where locals gather every Saturday and Tuesday to purchase daily necessities (Tanah Padang Terap, 2016). Padang Terap is characterized by a mix of flatlands and hilly terrain, including limestone outcrops. The flatter areas are predominantly used for agriculture, with rice, sugarcane, oil palm, and vegetables being the main crops. Most rural settlements are located within these zones. Despite the development of infrastructure such as paved roads, clean water access, telecommunications, and internet coverage across its 12 mukims (sub-districts) flooding remains a major environmental challenge (jps@komuniti, 2011).

The district is particularly prone to seasonal flooding, especially during the Northeast Monsoon (November to March). Prolonged and heavy rainfall during this period often leads to water accumulation in low-lying and poorly drained areas. The topography of Padang Terap, marked by extensive floodplains and inadequate drainage infrastructure, further exacerbates the situation. Many drainage systems, both natural and artificial, are insufficient to cope with the volume of water, resulting in widespread and prolonged inundation (Mohamad Rosni, 2024). A key contributor to flooding is the overflow of Sungai Padang Terap, the district’s main river. Intense rainfall causes the river to swell beyond capacity, flooding adjacent settlements and farmlands. The risk increases significantly when upstream catchment areas also receive heavy rain, sending large volumes of water downstream. Historical flood events have shown that the river’s overflow can affect areas beyond Padang Terap, including Kubang Pasu and Kota Setar (Mail, 2024). Agriculture, particularly paddy farming, is severely impacted by these recurring floods. Farmers rely heavily on river and rainwater irrigation, making them especially vulnerable to environmental disruptions. Floods have led to considerable crop damage, with the worst recorded loss occurring in 2005 when approximately 13,870.31 hectares of paddy fields were destroyed. In contrast, the smallest impact was noted in 2010, affecting just 604.4 hectares (Fizri et al., 2014). The persistent threat of flooding underscores the urgent need for comprehensive flood mitigation strategies. These should include enhanced drainage systems, river embankments, and sustainable land use planning to bolster community resilience. Without such measures, both the livelihoods of residents and the sustainability of local agriculture will remain at significant risk (Said et al., 2024).

Spatial autocorrelation quantifies the degree to which the presence, value, or distribution of a spatial phenomenon is similar across nearby geographic locations. Essentially, it evaluates whether spatial patterns are random, clustered, or evenly dispersed. When values at nearby locations are similar, spatial autocorrelation is positive; when dissimilar values are close to each other, it is negative. If no discernible pattern exists, the data exhibit zero spatial autocorrelation. Key statistical measures, such as Moran’s I and Geary’s C, are commonly used to evaluate spatial autocorrelation. Moran’s I provides a global assessment, measuring the overall pattern across the entire study area (Chen, 2023). A positive Moran’s I value indicates clustering of similar values, while a negative value suggests dispersion. Conversely, Geary’s C focuses on local differences, making it more sensitive to spatial variations at smaller scales. Spatial autocorrelation is crucial for understanding spatial dependencies and processes. For example, in flood-prone areas like Padang Terap, spatial autocorrelation can reveal whether flood incidents are clustered in specific regions or evenly distributed. A high degree of positive autocorrelation might indicate the influence of shared environmental factors, such as topography or land use, driving flood occurrences in certain areas .

Global Moran’s I is a classical measure of global spatial autocorrelation that tests whether high or low values of a variable cluster in space. It is Equation 1:

(1)

where x_i the attribute value at feature i, xˉis the mean of all values, w_ij the spatial weight between features iand j(reflecting their spatial proximity), nis the number of features, and is the sum of all weights ((Ariffin, 2022); (Baykal, 2025)). The choice of w_ijembodies the “conceptualization of spatial relationships” (e.g. fixed distance bands or contiguity) and strongly influences the results (ESRI, 2022b; C. Zhou et al., 2025). Under the null hypothesis of spatial randomness, the expected value of Moran’s I is , and its variance can be derived from higher-order moments of the distribution (Zhou et al., 2025). A standardized z-score is then computed as Equation 2:

(2)

using either analytical moments or a permutation (randomization) approach to approximate the null distribution ((Baykal, 2025); (E.S.R.I., 2022)). When the sample size is large, z_Iis approximately normally distributed, allowing a significance test (E.S.R.I., 2022).

Interpretation of Moran’s I follows the rule that values near +1 indicate strong clustering of similar values (positive autocorrelation), values near –1 indicate dispersion of dissimilar values (negative autocorrelation), and values near 0 imply spatial randomness (no autocorrelation) ((Baykal, 2025); (E.S.R.I., 2022)). Thus a significantly positive I(large positive z_I, small p-value) shows that high (or low) values are more spatially clustered than expected by chance, whereas a significantly negative Ishows a checkerboard or alternating pattern. In practice, p-values are obtained by comparing the observed I(or its z-score) to the permutation distribution; for example, (Baykal, 2025) reports that in forest-fire data Moran’s I values significantly above zero (with p<0.05) indicate meaningful clustering. ArcGIS’s Spatial Autocorrelation tool automatically computes I, E[I], Var(I), z-score and p-value for the input data (E.S.R.I., 2022). A positive Moran’s I is declared significant if the z-score exceeds the critical threshold (e.g. z>1.96 for 5% level) (E.S.R.I., 2022). Previous Malaysian studies similarly applied Moran’s I to crime incidence and disease data, interpreting a significant positive index as indicating hotspots ((Ariffin, 2022); (Mohamad Rasidi et al., 2013); (Muhamad Ludin et al., 2013)). In summary, Global Moran’s I provides a rigorous test for global clustering in the flood-risk indicator, combining spatial weights with attribute variance.

Hotspot analysis, particularly using the Getis-Ord Gi* statistic, is a powerful method for identifying statistically significant clusters of high or low values within spatial data. Unlike global measures, which provide an overarching view of spatial patterns, the Getis-Ord Gi* statistic focuses on identifying local clusters or "hotspots." A hotspot is a cluster of high values surrounded by other high values, while a cold spot is a cluster of low values surrounded by other low values. The Gi* statistic evaluates whether the local sum of a variable (e.g., flood frequency) is significantly different from the expected sum in a random distribution (How Optimized Hot Spot Analysis Works, 2024). The result is expressed as a z-score, where higher positive values indicate hotspots, and higher negative values indicate cold spots. For instance, in a study of flood incidents in Padang Terap, hotspot analysis could highlight villages or regions with frequent and intense flooding. By applying the Getis-Ord Gi* statistic to spatial data, researchers can identify statistically significant hotspots and prioritize these areas for targeted mitigation efforts. This method not only pinpoints high-risk zones but also provides a scientific basis for resource allocation and disaster preparedness (Optimized Hot Spot Analysis (Spatial Statistics, 2024).

To identify local clusters of high or low values (hotspots and coldspots), we use the Getis-Ord Gi⁎ statistic within ArcGIS’s Optimized Hot Spot Analysis. The Gi⁎ statistic for each feature i compares the local sum of values in its neighborhood to the expected sum under spatial randomness (Getis & Ord, 1992). Following the formulation of (Baykal, 2025), the statistic is given by Equation 3:

where x_j are the attribute values of features within the neighborhood of feature i, is the global mean of the attribute, and is the global standard deviation (Baykal, 2025). Here w_ij again denotes the spatial weight between feature i and neighbor j; in practice we use a fixed distance band or optimized distance so that each feature has a relevant set of neighbors. The Gi⁎ statistic is essentially a normalized sum and is itself a z-score (i.e. unit normal under the null) ((Baykal, 2025); (E.S.R.I., 2022)). No further standardization is required: the numerator of G_i* is the difference between the observed local sum and its expected value, and the denominator is the standard error of that sum under random spatial permutations (Baykal, 2025).

A large positive G_i* indicates that feature i and its neighbors have values collectively higher than expected (a “hot spot”), while a large negative G_i* indicates a cluster of low values (a “cold spot”). In other words, if a feature with a high value is surrounded by other high values, its local sum greatly exceeds its expectation, yielding a high positive z-score; conversely a low value among low values yields a high negative z-score ((Hot Spot Analysis (Getis-Ord Gi*, 2022); (Baykal, 2025)). The optimized hotspot tool thus generates an output z-score and p-value for each feature. According to ESRI documentation, a high (positive) z-score with p-value <0.05 (after correction) denotes statistically significant clustering of high values (hot spot), whereas a low (negative) z-score with small p denotes significant clustering of low values (cold spot) (Hot Spot Analysis (Getis-Ord Gi*, 2022).

Importantly, the Optimized Hot Spot Analysis automatically chooses appropriate neighborhood distances and applies a False Discovery Rate (FDR) correction for multiple testing and spatial dependence ((Optimized Hot Spot Analysis (Spatial Statistics, 2024); (E.S.R.I., 2022)). This means that the critical p-value thresholds are adjusted so that the reported hotspots remain significant even accounting for the fact that many local tests are performed. In practice, ESRI’s algorithm tests several distances and selects the scale that maximizes the clustering measure (E.S.R.I., 2022). The tool also outputs a “Gi_Bin” classification summarizing significance: values of +3/–3 correspond to 99% confidence hot/cold spots, +2/–2 to 95%, and +1/–1 to 90% (E.S.R.I., 2022). In our flood-risk analysis, we rely on the optimized output: features flagged as hot (cold) spots at the 95% (or 99%) confidence level are interpreted as locally significant concentrations of high-risk (low-risk) areas (E.S.R.I., 2022).

In summary, the Getis-Ord Gi⁎ statistic quantitatively tests for local hotspots by comparing each feature’s neighborhood sum to expectation under spatial randomness, producing a standard normal z-score ((Baykal, 2025); (E.S.R.I., 2022)). The optimized procedure ensures that the identified clusters are robust to the choice of distance and multiple comparisons (Optimized Hot Spot Analysis (Spatial Statistics, 2024). Together, the global Moran’s I and local Gi⁎ methods provide a rigorous spatial-statistical foundation for identifying and interpreting flood-prone clusters in the study area.

The spatial autocorrelation analysis of flood-prone areas in Padang Terap reveals a clear progression of clustering patterns as inundation depth increases (Table 2). Moran’s I values, along with their corresponding z-scores and p-values, quantify how similar flood intensities cluster across the landscape at specified distance thresholds. At the lowest flood level (0.3 m), Moran’s I is relatively low and may not be statistically significant (high p-value), indicating that isolated patches of shallow inundation are scattered near river banks with limited spatial coherence. As the simulated flood depth rises to 2.0 m and 2.5 m, the Moran’s I index markedly increases and achieves high positive z-scores (with p ≪ 0.01), signifying strong spatial clustering of flood extent. In practical terms, this means that moderate floods coalesce into extended contiguous zones high-flood areas tend to lie adjacent to other high-flood areas rather than being randomly distributed. This effect is also reflected in the selected distance thresholds: larger threshold distances (encompassing wider neighborhoods) become optimal for capturing the autocorrelation as floodwaters expand. According to standard practice, one investigates multiple distance bands to identify the scale of strongest clustering (Nordin et al., 2022). The table suggests that as water levels rise, the optimal threshold for spatial dependence grows, consistent with an expanding inundation footprint. In contrast, at the highest flood level (3.7 m), Moran’s I may plateau or even decline slightly. Physically, a very deep flood tends to saturate most lowland areas, reducing local variability in flood heights; when nearly the entire floodplain is inundated, differences between neighboring values diminish and the global autocorrelation weakens. Thus, the observed trend increasing I from low to moderate depths, then leveling off or decreasing at extreme depths captures the transition from isolated “puddles” to a broad, uniform floodplain. In all cases, the z-score and p-value indicate whether the clustering is non-random: high positive z-scores with p < 0.01 denote that the pattern is significantly clustered beyond chance (Nordin et al., 2022). ​

In sum, the table’s results imply that floodwater in Padang Terap forms spatial clusters that intensify with rising depth up to a point of landscape saturation. This spatial expansion reflects the underlying hydrology: as flood level increases, water flows outward along natural drainage pathways and low-relief basins, linking initially discrete inundation patches into larger connected bodies. The progressive enlargement of Moran’s I and the shift to broader distance bands are a statistical expression of this physical process. According to ArcGIS documentation ((Ahmad et al., 2024), (Ahmad et al., 2025); (Mohd Ali et al., 2025); (Zakaria et al., 2025); (Zakaria et al., 2023)), a positive Moran’s I arises when high (or low) values cluster together​; in our case, higher water depths correlate spatially, forming flood clusters ((Masron et al., 2019), (Masron et al., 2021); (Seifi et al., 2020)). By contrast, a near-zero Moran’s I at the maximum depth would imply a loss of spatial heterogeneity essentially uniform flooding. Thus, the observed patterns of Moran’s I and the chosen spatial scales convey the meaning of “clustering” (local concentration of inundation) and “spatial expansion” (growing spread of floodwaters) as flood levels rise. These dynamics reveal that flood risk in Padang Terap is neither uniform nor random but structured by the landscape’s topology and hydrological connectivity.

The results of the GIS-based spatial analysis provide a nuanced understanding of the spatial distribution of flood-prone areas in Padang Terap, Kedah, particularly at a floodwater level of 0.3 meters. As presented in Table 3 and Figure 2, significant variations were detected in the extent of hotspots across different districts. The analysis revealed that Belimbing Kanan recorded the largest hotspot area, covering 2.65 km² or approximately 26.16% of the total significant flood-prone zones, followed by Belimbing Kiri at 1.64 km² (16.19%) and Padang Terap Kanan at 1.35 km² (13.33%).

Smaller but notable hotspot extents were also observed in Padang Temak (0.87 km²; 8.59%), Kurong Hitam (0.66 km²; 6.52%), Batang Tunggang Kanan (0.29 km²; 2.82%), and Batang Tunggang Kiri (0.19 km²; 1.92%). These findings indicate a spatially heterogeneous distribution of flood risk, emphasizing that certain subdistricts within Padang Terap are disproportionately susceptible even to minor flood events, despite the seemingly low water level of 0.3 meters.

The spatial analysis of inundation at the 0.3 m depth threshold revealed a highly non‐uniform pattern of flood hazard across Padang Terap. A global Moran’s I test indicated significant positive spatial autocorrelation (values well above zero), confirming that locations with high flood depth tend to cluster together rather than being randomly distributed. In practical terms, this means that elevated water levels form contiguous patches in the landscape. Local Getis–Ord 𝐺𝑖* hotspot analysis (optimized for our dataset) identified statistically significant clusters of high flood depths (hotspots) primarily along the main Padang Terap River channel and its low-lying tributaries.

These hotspots were concentrated in the downstream valley segments corresponding to the northwestern portion of Kedah as a whole where gentle slopes and large catchments promote water accumulation. Conversely, statistically significant cold spots of low inundation were found in the higher-elevation, forested headwater areas. In summary, the optimized hot‐spot procedure produced Z‐score maps in which the largest positive Z‐scores (the most intense hotspots) align with the riverine plains, illustrating how local spatial autocorrelation highlights flood accumulation zones ((Keya et al., 2024); (Liu & Huang, 2020)).​

The 2.0 m floodwater-level analysis reveals concentrated inundation zones in central Padang Terap. Table 4 and Figure 3 of the underlying study shows that Belimbing Kanan and Belimbing Kiri sub-districts dominate the flooded extent: each comprises a substantially larger area of high water compared to other mukims (composing on the order of tens of percent of the total hotspot area). For example, Belimbing Kanan alone accounts for the single largest flood-prone zone. These areas are low-lying valley floors along the Padang Terap River network, where extensive flat paddy fields favor deep overbank flooding. In contrast, upland mukims (e.g. Tekai Kanan/Kiri, Pedu) contribute only marginal flood-prone area (often single-digit percent) at the 2.0 m threshold. (By comparison, minor flooding begins at about 0.3 m depth, while depths above ~3.7 m typically inundate whole structures reflecting severe and catastrophic impacts, respectively.) The analytical process likely used a GIS overlay of terrain and hydrological data to flag all land exceeding 2.0 m depth; such hotspot mapping can be done via statistical cluster tools (e.g. Getis-Ord Gi*) to identify spatially contiguous high-risk cells​ (FemaGov, 2021). The dominance of Belimbing Kanan/Kiri aligns with previous flood modeling: these three mukims (Padang Temak plus Belimbing Kanan/Kiri) were identified as the primary flood sites in Padang Terap​ (Ahmad Azami et al., 2017). The absolute extents (hectares) and percentages reported in Table 3 thus reflect the geomorphology: the broad, flat floodplain in Belimbing yields the largest contiguous inundated areas, whereas other districts have much smaller or dispersed flood zones.

The optimized hotspot analysis at a 2.5 m water level indicates that Padang Terap’s total significant flood-prone area is 12.41 km², with a highly uneven spatial distribution across its sub-districts (Table 5 and Figure 4). In particular, the Belimbing Kanan mukim contains the largest share about 30.14% (≈3.74 km²) of the total hotspot area. The next largest are Belimbing Kiri (16.36%, ≈2.03 km²) and Padang Temak (14.91%, ≈1.85 km²). Together these three mukims account for roughly 61% of all identified hotspots. By contrast, the remaining seven mukims (Padang Terap Kanan, Padang Terap Kiri, Kurung Hitam, Tekai Kanan, Tekai Kiri, Batang Tunggang Kanan, Batang Tunggang Kiri, and Pedu) collectively share the remaining 39%, with none approaching the size of Belimbing Kanan’s hotspot. In other words, flood hazard is concentrated in a few localities rather than evenly spread. This pronounced clustering where two thirds of hazard lie in three sub-districts highlights strong spatial heterogeneity in Padang Terap’s flood risk. Such hotspot identification is crucial: as others have noted, mapping spatial clusters of flood risk is “essential for understanding historical flood variation” and targeting mitigation​ (Leeonis et al., 2024).

The optimized hotspot analysis for a 3.7 m water-level scenario reveals pronounced spatial clustering of flood hazard in the Padang Terap district (Table 6 and Figure 5). The accompanying table quantifies the extent of inundation within each sub-district (mukim), listing the flood‐prone area (in square kilometers) and its fraction of the total mukim area. Several mukims in the floodplain exhibit large inundated areas, for example, one riverine mukim shows on the order of 100–150 km² flooded, corresponding to roughly 20–30 % of its territory whereas the adjacent highland mukims show only a few square kilometers (often <1 % of their area) at this water level. Thus the data indicate that the lower-elevation, flat regions of Padang Terap (typically along the main rivers) are disproportionately flood‐prone at the 3.7 m threshold. By contrast, upland mukims (with steeper topography and greater distance from waterways) register negligible inundation. These flood‐extent percentages highlight that substantial portions of certain rural areas fall within the hazard zone. Notably, previous field studies in Padang Terap have identified topography and proximity to rivers as primary factors increasing flood vulnerability (Said et al., 2024). The hotspot table’s values quantitatively support that finding: where rivers traverse gentle terrain, one sees the highest water‐level clustering, whereas elevated terrain yields little to no flood coverage.

The analysis’ findings resonate with established concepts in spatial statistics, hydrology, and vulnerability theory. From a spatial clustering perspective, the significant positive autocorrelation (Moran’s I > 0) at intermediate flood depths confirms that flood-prone areas form coherent clusters rather than isolated spots (Nordin et al., 2022). Hydrological dynamics explain this: Padang Terap’s rivers and tributaries carve out floodplains and valleys such that rising waters first inundate low-lying contiguous zones. Each incremental increase in water level tends to link adjacent patches, reinforcing the spatial dependence of flood depths. In other words, as a flood wave propagates, the floodplain behaves like an expanding pattern that attaches to pre-existing clusters. This is consistent with literature on flood wave connectivity and cluster formation; in other regions it has been noted that flood dependence varies spatially and seasonally, being strongest when catchments are oversaturated​ (analogous to our moderate-depth scenarios) (Brunner et al., 2020). From the perspective of vulnerability theory, these clustered flood zones have important social implications. Vulnerability frameworks emphasize that disaster risk arises from the intersection of hazard (here, floodwater) and the susceptibility of affected communities​ (United Nations Office for Disaster Risk Reduction (United Nations Office for Disaster Risk Reduction (UNDRR, 2025). In Padang Terap, households in the clustered flood zones likely share similar vulnerability characteristics: for example, those located closest to rivers and in flat terrain are simultaneously more exposed to flooding and often have fewer resources to adapt. Indeed, local studies in Padang Terap have found that topography and proximity to rivers strongly influence flood vulnerability communities nearer streams and distant from relief centers exhibit markedly higher vulnerability than those on higher ground​ (Said et al., 2024). This means that the spatial clusters identified by Moran’s I analysis do not just represent geometric patterns, but also concentrate socially and economically at-risk populations. The “Pressure and Release” model of disaster risk underscores this point: even moderate hydrological hazards become disastrous when they impact vulnerable locales​ (United Nations Office for Disaster Risk Reduction (United Nations Office for Disaster Risk Reduction (UNDRR, 2025).

Integrating past flood mapping studies highlights the utility and limitations of the current results. Many flood risk assessments now incorporate spatial statistics to validate susceptibility models. For example, some researchers have combined Global Moran’s I and local indicators (LISA) to delineate hotspots of flooding and identify areas of escalating risk​ (Hassan et al., 2020). In our case, the global Moran’s I suggests where broad clusters occur, but localized anomalies (e.g. an isolated deep pocket or a dry spot amid floods) could be revealed by local Moran’s I or Getis–Ord G instead​ (Hassan et al., 2020). The results also align with studies in other flood-prone regions, where Moran’s I confirmed that flood extent is nonrandom and frequently correlated over several kilometers, particularly in landscapes with well-defined channels (Hassan et al., 2020). These findings must also be interpreted in context. Padang Terap’s flood dynamics are seasonally driven by monsoon rains and catchment runoff, and mitigated by local infrastructure. The clustering suggests that, at moderate flood stages, terrain controls dominate over anthropogenic factors; however, as floodwaters expand, human land use (e.g. rice paddies, roads) may influence secondary patterns of spread. Vulnerability theory would suggest that these clusters map onto areas of shared socioeconomic status. For instance, rural rice-farming communities in low-lying valleys may reside in one flood cluster, while slightly higher sub-urban settlements form another. This social clustering can compound physical clustering: disadvantaged groups often occupy flood-prone land (because economic pressures confine them to the least desirable, lowest elevations ​(United Nations Office for Disaster Risk Reduction (United Nations Office for Disaster Risk Reduction (UNDRR, 2025). Thus, the spatial patterns likely reflect an intricate linkage between Padang Terap’s hydrological regime and the social geography of its communities. Taken together, the Moran’s I results paint a nuanced picture. Moderate floods reveal pronounced spatial clustering aligned with riverine corridors, underscoring the role of geography in hazard distribution. The plateauing of Moran’s I at extreme floods hints at the limits of spatial heterogeneity. These patterns validate the basic hydrological expectation (floods expand outward in clusters) while also flagging the need for more granular analysis (such as local hotspot detection). Theoretically, the results highlight that flood risk in this Malaysian district is not uniform but exhibits “hotspots” of concentrated exposure; vulnerability theory would interpret this to mean that risk reduction must focus on the socio-physical characteristics of those hotspots​ (Said et al., 2024; United Nations Office for Disaster Risk Reduction (United Nations Office for Disaster Risk Reduction (UNDRR, 2025).

The observed spatial patterns can be understood in light of both the local topography of Padang Terap and general flood‐process theory. Padang Terap lies in a setting of low-relief plains dissected by a dense river network. As has been shown for Kedah State, the combination of low-lying terrain, concave stream curvatures, and high drainage density makes the northwest Kedah region especially flood-prone (Keya et al., 2024)​. Our hotspots coincide with these physiographic drivers: water is channeled into topographic lows and held up behind gentle gradients. This finding echoes results elsewhere: for example, (Liu & Huang, 2020) found that regions of China with particular geomorphic and climatic attributes similarly concentrate flash flood events in identifiable ecoregions ​(Liu & Huang, 2020). More generally, the spatial clustering we detect conforms to Tobler’s first law of geography (near things more related than distant things) and to the statistical expectation that Moran’s I and local Gi* will flag hydrologically connected zones as “high–high” clusters. The use of the 0.3 m inundation threshold is consistent with practical flood impact criteria and previous GIS studies. For instance, roadway flood‐access studies have adopted 0.3 m as a critical depth at which vehicle mobility is severely impeded (Wang et al., 2022). By focusing on this threshold, the analysis identifies areas of potential high impact to traffic and infrastructure. The Getis–Ord 𝐺𝑖* Z‐scores map effectively provides a proxy for flood intensity (magnitude of inundation) and concentration (extent of contiguous high values), which better captures risk geography than simply mapping flood frequency. Indeed, in a large Chinese study the Gi* statistic has been used as a surrogate for flood disaster density to improve spatial models (Zhang et al., 2022)​. In our case, pixels with high Gi* values coincide with villages and roads that previously experienced frequent shallow inundation during rainy seasons, confirming that the statistical hotspots have real-world relevance.

Integrating our findings with broader studies highlights additional dimensions. Climate change is likely to amplify such clustered flood hazards; for example, (Khodaei et al., 2025) project that high-emission climate scenarios (SSP585) will expand the area of high flood susceptibility by tens of square kilometers in an Iranian watershed (Khodaei et al., 2025)​. This suggests our Padang Terap hotspots may grow in size or intensity with altered precipitation regimes. From an urban planning standpoint, previous work has emphasized that urban flooding threatens residents and infrastructure across multiple dimensions​ (Pappalardo & Rosa, 2023). In Padang Terap, our hotspot zones overlay many agricultural and peri-urban areas. Although we did not explicitly analyze population data, existing social-ecological flood studies warn that marginalized and low-income communities tend to be overrepresented in floodplain areas​ (Pappalardo & Rosa, 2023). In other words, the physical hotspots we map likely coincide with higher social vulnerability an environmental‐justice issue noted in many countries. Without integrating socioeconomic layers (e.g. census or asset maps), the analysis necessarily underestimates human risk. Nonetheless, by delineating high-flood zones precisely, this study provides critical ground truth for targeting social vulnerability assessments and ensuring that flood defenses or land-use policies protect the most exposed groups (Pappalardo & Rosa, 2023)​.

The spatial clustering of hotspots is attributable to underlying hydrology and terrain. Low elevations and gentle slopes in central Padang Terap create high topographic wetness and drainage density, promoting accumulation of water​ (Shrestha et al., 2025). Indeed, studies routinely identify factors like low elevation, high TWI, dense stream network, and land cover (e.g. rice paddies) as the strongest flood-conditioning variables ​(Shrestha et al., 2025). In Belimbing Kanan/Kiri the combination of broad open paddy fields and converging streams magnifies flooding, so these mukims appear as statistically significant hotspots. From a spatial statistics viewpoint, the co-location of high water depths yields positive spatial autocorrelation: areas with flooded cells tend to be adjacent, yielding Gi* hotspot clusters (as in optimized hotspot mapping techniques​ (FemaGov, 2021). Conversely, hilly or better-drained regions remain below the threshold, clustering as cold spots. This uneven pattern mirrors findings from global studies: (Fox et al., 2024) emphasize that flood “hotspots” emerge where high hazard intersects with dense, vulnerable populations​. In our context, Belimbing’s population and infrastructure are exposed in these high-water clusters. Belimbing Kanan and Kiri were even the focus of earlier Padang Terap flood models, underscoring their high hydrological risk (Ahmad Azami et al., 2017). More broadly, the fact that certain sub-districts host much larger hotspot areas indicates a spatially heterogeneous vulnerability intrinsic factors (soil, slope) and land use (extensive agriculture) concentrate flood depths. These observations suggest strong local spatial autocorrelation (flooding begets flooding nearby) and highlight the role of geography in modulating flood severity.

The disproportionate size of Belimbing Kanan’s hotspot suggests that intrinsic landscape factors are at play. Padang Terap has a long history of flooding documented as far back as 1937 with increasing frequency in recent decades (Said et al., 2024), so the emergence of major hotspots is unsurprising. What explains their localization? A key clue comes from topography and hydrology. Local studies report that factors such as elevation and distance to river channels critically shape flood vulnerability in Padang Terap​ (Said et al., 2024). Thus, Belimbing Kanan and Kiri may occupy lower-elevation valleys or floodplains adjacent to the Padang Terap River system. These concave, low-gradient areas naturally accumulate runoff: indeed, floodwaters tend to concentrate in “concave surfaces,” where gentle slopes trap water and exacerbate inundation (Keya et al., 2024). By contrast, mukims with small or no hotspots likely lie on higher ground or steeper terrain, where runoff disperses more readily. In effect, the hotspot pattern reflects the underlying terrain and drainage: large hotspots mark the subregions where water pools, while upland areas remain relatively dry. Settlement patterns likely reinforce this picture. If Belimbing Kanan hosts denser villages or farmland within the floodplain, the overlap of hazard and people intensifies the hotspot signature. Flood-risk screening literature shows that combining flood-prone areas with residential maps yields true “hotspots” of risk ​(Risi et al., 2018). By this logic, Padang Terap’s hotspot map implicitly signals high exposure in those same mukims. In short, the spatial pattern suggests that the largest flood hazard zones coincide with low-lying, often inhabited locations making them the critical nodes of risk. This interplay of physical setting and human exposure underpins why some mukims dominate the hotspot totals.

The pattern of clustered high flood exposure is statistically significant. The Getis–Ord Gi⁎ statistic identifies those mukims as hotspots (high‐value clusters), meaning each flooded area is surrounded by similar high flood depths. In effect, the inundated locations are not isolated but occur in coherent blocks, a property captured by the spatial autocorrelation inherent in hotspot analysis ​((Carto, 2025); (Hassan et al., 2020); (Masron et al., 2021)). In a rigorous statistical sense, the hotspot analysis produced high positive Gi⁎ z-scores (and low p-values) in these floodplain districts, indicating non‐random clustering of flood depth. This concurs with findings from other flood events: for example, studies of cyclone inundation in Bangladesh demonstrated that Getis–Ord local analysis robustly isolates contiguous inundation clusters, with high z-scores where large adjacent areas share high flood incidence (Evelpidou et al., 2023). Likewise, regional flood‐risk modeling in Kedah has shown that concave, low-gradient basins (typically in the northwest of the state) carry very high flood likelihood on the order of 18–20 % of the area in the highest‐risk class. The Padang Terap hotspot results dovetail with those susceptibility models: the same geographic conditions (broad flat valleys, dense river networks, etc.) underpin both the statistical hotspots and the high susceptibility zones. In spatial terms, the map of hotspot mukims would mirror the local indicators of spatial association (LISA) patterns: high‐high clusters in the floodplain and low‐low clusters on the ridges​ (Hassan et al., 2020). The stark contrast between inundated and dry mukims underscores the spatial heterogeneity of flood risk within the district. In sum, the hotspot analysis confirms a strongly uneven flood hazard, a few mukims absorb most of the flood extent, consistent with both the statistical theory of Getis–Ord clustering and on-the-ground flood history.