Multi-Temporal Assessment of Bimodal Monsoon Flood Dynamics and Agricultural Exposure Using Integrated Sentinel-1 SAR and Sentinel-2 Optical Data in Punjab, Pakistan

Abstract

Floods in monsoon-dominated river basins exhibit high spatio-temporal variability, necessitating high-resolution, multi-sensor approaches for reliable monitoring and impact assessment. In flood-prone agricultural regions, continuous monitoring using optical remote sensing is frequently hindered by dense monsoon cloud cover. This study establishes a comprehensive multi-sensor framework within the Google Earth Engine (GEE) to examine the spatio-temporal dynamics and land surface impacts of the 2025 monsoon floods in Punjab, Pakistan. Flood inundation mapping was executed using a 12-day Sentinel-1 Synthetic Aperture Radar (SAR) time series via a dual-threshold change detection methodology. Concurrently, Sentinel-2 imagery facilitated the derivation of land use/land cover (LULC) changes and vegetation dynamics using a Random Forest classifier, achieving overall accuracy of 93% (pre-flood), 91% (during flood), and 94% (post-flood). These accuracy levels were consistent across all three phases despite spectral confusion between water, saturated soil, and vegetation during peak inundation, indicating consistent classification performance under monsoon conditions. The analysis revealed a distinct bimodal flooding regime, characterized by an early monsoon peak in July–August and a more severe late monsoon peak in August-September. The cumulative maximum flood extent reached 9495.33 km², with peak single-date inundation reaching 5449 km². Mapped cropland declined by 6.7% (8181 km²) during peak flooding, with 3.9% (4796 km²) remaining non-cropland by the end of the observation period; 5892 km² of pre-flood cropland was identified as inundated through spatial intersection. In addition, the Normalized Difference Vegetation Index (NDVI) declined by 28.6%, from 0.28 to 0.20, indicating a substantial reduction in vegetation greenness. Spatial consistency was checked with the United Nations Satellite Centre (UNOSAT) and the Food and Agriculture Organization (FAO), independently collected data showing moderate spatial agreement. The proposed framework is highly scalable for continuous flood monitoring, offering critical insights for disaster management and climate adaptation planning in monsoon regions plagued by data scarcity and persistent cloudiness. The approach is particularly relevant for near-real-time operational monitoring, given its reliance on freely available Sentinel data and cloud-based processing that requires no specialized ground infrastructure.

1. Introduction

Floods are recognized as one of the most frequent environmental catastrophes globally, causing extensive socio-economic damage and affecting millions of people annually [1,2]. The occurrence of floods is closely associated with extreme precipitation events, which are becoming increasingly variable and unpredictable under changing climatic conditions [3,4]. This increasing climate variability, combined with rapid urbanization, population growth, and the expansion of settlements and infrastructure in flood-prone areas, has significantly heightened global flood vulnerability [5]. Furthermore, anthropogenic land-use changes, deforestation, and alterations to natural drainage systems have intensified flood impacts, leading to substantial economic and environmental losses worldwide [6,7].

These challenges are particularly pronounced in South Asia, where intense monsoon rainfall, high population density, and rapid, often unplanned, urban expansion have made the region highly vulnerable to flood hazards [8,9]. In such contexts, precise flood mapping is crucial for delineating flood extents, assessing damages, and supporting effective disaster management, particularly in agricultural regions where field-based observations are often constrained during active flood events [10,11]. Pakistan represents a critical case, being among the world’s most climate-vulnerable countries [12,13,14]. The catastrophic 2022 floods, which affected more than one-third of the country’s agricultural land, along with subsequent extreme monsoon events in 2025, highlight the growing influence of climate extremes coupled with land-use vulnerabilities on flood risks [15,16].

Despite the frequency of these events, obtaining reliable flood monitoring during the peak monsoon season remains a significant challenge due to limited field accessibility [17,18]. Although optical remote sensing is widely used for flood inundation mapping, its effectiveness is often limited by persistent cloud cover associated with the South Asian monsoon [19,20,21]. In contrast, Synthetic Aperture Radar (SAR), particularly C-band sensors like Sentinel-1, offers a robust solution by enabling flood detection under cloudy conditions and during both flood onset and recession phases [22,23,24].

Although SAR-based remote sensing has improved flood monitoring capabilities under challenging monsoon conditions, its application in Pakistan has largely focused on single-date flood assessments or the delineation of inundation extent, providing limited insight into the spatio-temporal evolution of complex flood events [25,26,27]. Furthermore, there is a distinct lack of research integrating SAR-based flood mapping with optical data to evaluate the impacts on vegetation and agriculture [28]. This gap is critical in agriculturally grown regions such as the Punjab plains, where evaluating flood duration, vegetation response, and post-flood recovery processes is crucial for understanding agricultural resilience and developing effective flood management strategies [29].

While the catastrophic 2022 floods have been extensively documented, the complex bimodal flooding regime observed in 2025, characterized by distinct early and late monsoon peaks, presents unique challenges for agricultural monitoring, particularly during the critical transition between the Kharif and Rabi cropping seasons. Understanding the spatio-temporal persistence of these sequential events is crucial for developing targeted adaptation strategies. To address these limitations, this study offers a multi-temporal and multi-sensor framework by integrating Sentinel-1 SAR and Sentinel-2 optical data within the Google Earth Engine (GEE) environment [30]. The study aims to (1) characterize the spatio-temporal evolution of the 2025 bimodal flood event across Punjab; (2) assess flood-induced land use/land cover (LULC) changes and vegetation responses using NDVI; and (3) establish an efficient and transferable approach for flood impacts on agricultural land and vegetation in data-limited, monsoon-affected regions.

2. Materials and Methods

2.1. Study Area

Punjab is the second-largest province of Pakistan, which is situated between 27°33′ N and 34°02′ N latitude and 69°18′ E and 75°23′ E longitude (Figure 1). The province is approximately 205,344 km² in area, forming a large portion of the Indus Basin, and is characterized by the fertile alluvial plains of the Indus and its major rivers—Jhelum, Chenab, Ravi, and Sutlej. It is the most populous province of Pakistan, with an estimated population of more than 120 million [31]. It has a subtropical climate type, with sweltering summers, with temperatures regularly breaking 45 °C in June, and mild to cold winters, with temperatures dropping to about 4–7 °C in January [32]. The monsoon season is from July to September, and it makes up the bulk of the yearly rainfall. Punjab is the ‘breadbasket of Pakistan’ with flat topography, which is well suited to produce wheat, rice, sugarcane, and cotton on a large scale [31]. This agricultural intensity, combined with a dense irrigation canal network fed by the Indus and its tributaries, means flood impacts in Punjab are compounded by pre-existing surface-water management infrastructure, which can both mitigate and, in cases of breach or overflow, exacerbate inundation extent.

Figure 1. Geographical location of Punjab, Pakistan.

2.2. Data

Multi-source satellite remote sensing data were used for the assessment of the 2025 flood event. This dataset consisted of C-band Synthetic Aperture Radar (SAR) Sentinel-1 data (10 m spatial resolution), provided by the European Space Agency (ESA, Paris, France), which were used to map flood inundation. Land use/land cover (LULC) classification and vegetation dynamics analysis using spectral indices were carried out using Sentinel-2 Surface Reflectance (Level 2A) data at 10 m spatial resolution provided by the European Space Agency (ESA, Paris, France). The Google Earth Engine (GEE) cloud platform was used for data processing and analysis, and ArcGIS (version 10.8) was used for post-processing, map generation, and spatial visualization. All satellite datasets used are freely and openly accessible, supporting the reproducibility of this framework in other data-scarce monsoon-affected regions without reliance on commercial imagery.

2.3. Methodology

A systematic methodological framework was established to process and analyze multi-temporal satellite data. The analysis is organized into three temporal phases: pre-flood (February–March 2025), during-flood (July–August 2025), and post-flood (September–Early October 2025) conditions. An overview of the entire methodological workflow is presented in Figure 2.

Figure 2. Methodological outline of this study.

2.3.1. Preprocessing of Satellite Data

The Sentinel-2 cloud probability band (MSK_CLDPRB) was employed in GEE for preprocessing the Sentinel-2 images. Cloud contamination was reduced by masking pixels with Cloud Probability values above 10%, thereby enhancing the radiometric consistency of the imagery [33]. Following this cloud-masking procedure, 806, 980, and 510 Sentinel-2 scene observations were retained within the pre-flood, during-flood, and post-flood temporal windows, respectively. Median compositing was performed over the defined temporal windows to generate representative cloud-masked optical composites, particularly during the high-cloud-coverage monsoon period (July–August), while making use of the available observations within each temporal interval.

To reduce multiplicative speckle noise in the Sentinel-1 SAR imagery, a 5 × 5 focal median filter was applied before flood detection [17,23]. The 5 × 5 kernel was selected to balance speckle reduction with preservation of spatial details. Although spatial filtering may smooth narrow flood boundaries, the same filter was applied consistently across all temporal intervals to ensure comparable flood delineation.

2.3.2. Spatio-Temporal Flood Inundation Mapping and Threshold Optimization

A multi-temporal change detection technique was used for flood inundation dynamics mapping in the plains of Punjab using Sentinel-1 SAR data [34]. The flood-monitoring period from 1 July to 10 September 2025 was divided into six consecutive 12-day intervals. The 12-day interval was selected based on the repeat cycle of an individual Sentinel-1 satellite and to provide consistent temporal sampling for tracking flood onset, peak inundation, persistence, and recession. The six consecutive 12-day intervals were defined as 1–12 July, 13–24 July, 25 July–5 August, 6–17 August, 18–29 August, and 30 August–10 September 2025, with 32, 32, 33, 29, 32, and 33 available VV/IW Sentinel-1 acquisitions contributing to these intervals, respectively. Both ascending and descending observations from multiple relative orbits were included, with the same preprocessing and median compositing procedure applied consistently across all intervals to maintain comparability. A temporal median composite of VV-polarized backscatter data (from 1 May to 15 June 2025) was used for a pre-flood baseline, as the antecedent surface conditions were dry. Median compositing provided a representative backscatter condition for each 12-day interval.

A hybrid approach of relative backscatter change detection and absolute backscatter constraints with enhanced robustness in various land cover conditions [35] was used to delineate the extent of inundation in the flood. The relative change in backscatter was obtained as follows:

$$\Delta {\mathit{\sigma }}^{\mathbf{0}}={\mathit{\sigma }}_{\mathbf{p}\mathbf{r}\mathbf{e}}^{\mathbf{0}}-{\mathit{\sigma }}_{\mathbf{p}\mathbf{o}\mathbf{s}\mathbf{t}}^{\mathbf{0}}$$

(1)

where σ⁰_pre represents the pre-flood baseline SAR backscatter intensity in decibels (dB), σ⁰_post is the post-event backscatter intensity (dB), and Δσ⁰ represents the backscatter difference (dB). Accordingly, a decrease in post-event backscatter leads to a positive Δσ⁰, aligning with the specular reflection from unobstructed water surfaces. A systematic sensitivity evaluation was performed to assess threshold selection by using combinations of relative backscatter reduction (2.5–4.5 dB) and absolute post-event backscatter limits (−11 to −15 dB). Four representative threshold combinations were systematically tested: (1) Δσ⁰ > 2.5 dB, σ⁰ < −11 dB; (2) Δσ⁰ > 3.0 dB, σ⁰ < −12 dB; (3) Δσ⁰ > 3.5 dB, σ⁰ < −13 dB; (4) Δσ⁰ > 4.0 dB, σ⁰ < −14 dB. Threshold performance was evaluated using Sentinel-1 backscatter distributions, spatial inundation patterns, and SRTM terrain slope constraints. Relaxed thresholds (>2.5 dB, <−11 dB) produced an inflated footprint of 19,776 km² due to false detections over smooth barren surfaces. The final thresholds were a relative backscatter reduction > 3.5 dB and an absolute backscatter threshold of σ⁰ < −13 dB, selected as the most appropriate empirical threshold based on improved separation between flooded and non-flooded areas and retention of spatially coherent inundation signals. The absolute threshold was supported by the backscatter distribution of dry-season reference imagery, corresponding approximately to the lower 5th percentile of non-flooded surface conditions. This statistical constraint improved the separation between open water and low-backscatter land surfaces while reducing false detections associated with moist soils and urban surfaces [20]. Conversely, overly restrictive thresholds increasingly excluded shallow and marginal inundation, particularly due to vegetation-induced variations in radar response that weakened the flood signal [36]. Permanent water bodies were removed using a pre-flood water mask; irrigation channels, reservoirs, and seasonal water surfaces were not separately masked, while topographic distortions (radar shadow and layover) were minimized using an SRTM-based slope mask excluding areas with slopes ≥5° [37]. The final binary flood maps were aggregated across temporal intervals to evaluate flood extent dynamics and persistence [38].

2.3.3. NDVI Analysis

The Normalized Difference Vegetation Index (NDVI) was used to assess vegetation health and greenness across the pre-flood, during-flood, and post-flood periods [39]. Multi-temporal median composites were generated for the three specific intervals. NDVI was computed using Equation (2):

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{B8-B4}{B8+B4}}$$

(2)

where B8 and B4 denote the near-infrared (NIR) and red bands of Sentinel-2 imagery, respectively. NDVI values range from −1 to +1, with higher values indicating dense vegetation and lower or negative values denoting sparse vegetation or water. Vegetation condition changes during the flood period were assessed using NDVI change detection between the pre-flood baseline and flood peak using Equation (3):

$$∆\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=\mathrm{N}\mathrm{D}\mathrm{V}{\mathrm{I}}_{\mathrm{p}\mathrm{r}\mathrm{e}}-\mathrm{N}\mathrm{D}\mathrm{V}{\mathrm{I}}_{\mathrm{d}\mathrm{u}\mathrm{r}}$$

(3)

where ΔNDVI represents the change in NDVI, while NDVI_pre represents the pre-flood baseline NDVI and NDVI_dur represents the flood-peak NDVI. Positive ΔNDVI values signify a decline in vegetation greenness due to flood stress, whereas values around zero reflect limited change, and negative values indicate higher vegetation greenness during the flood period and do not necessarily represent post-flood recovery. It is acknowledged that NDVI values computed over open water or newly submerged vegetation can be artificially suppressed independent of actual plant stress, and this effect was not separately isolated from genuine vegetation degradation in the current analysis.

To quantify the relationship between vegetation response and flood persistence, ΔNDVI was evaluated against the number of inundated observation periods (0–6) identified from the Sentinel-1 flood-frequency analysis using Pearson correlation (r) across a pixel-wise sample (n = 9997). Mean ΔNDVI and standard deviation were also calculated for each flood-persistence class.

2.3.4. Land Use/Land Cover (LULC) Classification

A multi-temporal LULC classification was performed to investigate spatio-temporal characteristics of flood impacts. The Random Forest (RF) machine learning algorithm was used for supervised classification, and it was chosen due to its resistance to noise and superior classification performance in remote sensing [40,41].

A multi-source feature stack of Sentinel-2 spectral bands, spectral indices, and topographic variables was used to train the RF classifier. The spectral inputs included six Sentinel-2 surface reflectance bands: B2 (Blue), B3 (Green), B4 (Red), B8 (NIR), B11 (SWIR-1), and B12 (SWIR-2). The spectral indices included NDVI and EVI for vegetation characterization. Surface water bodies were identified from the Modified Normalized Difference Water Index (MNDWI) [42], and built-up areas and bare soils were identified by the Normalized Difference Built-up Index (NDBI) [43] and the Bare Soil Index (BSI) [44]. Elevation and slope derived from the SRTM DEM were incorporated to account for terrain-related variability. Furthermore, a seasonal NDVI baseline for the pre-monsoon (agricultural) period was used to improve the model’s ability to differentiate active croplands from permanently barren soils [33].

Five categories of land use/land cover (LULC) were distinguished: Water, Vegetation, Cropland, Built-up, and Barren. A total of 4601 reference samples were distributed throughout the study area across the pre-flood (n = 1355; Water: 150, Vegetation: 200, Cropland: 505, Built-up: 250, Barren: 250), during-flood (n = 1675; Water: 470, Vegetation: 200, Cropland: 505, Built-up: 250, Barren: 250), and post-flood (n = 1571; Water: 366, Vegetation: 200, Cropland: 505, Built-up: 250, Barren: 250) phases. For each temporal phase, the extracted samples were randomly partitioned into 70% for RF training and 30% for internal validation. This random partitioning does not guarantee complete spatial independence between the training and internal validation subsets and may result in optimistic accuracy estimates due to spatial autocorrelation in remote-sensing data. The RF classifier was trained with 300 decision trees, providing stable classification performance [17,40]. The remaining RF parameters were retained at their Google Earth Engine defaults. Feature importance was evaluated using the RF variable-importance output to assess the relative contribution of the input predictors.

2.3.5. Accuracy Assessment and Flood Extent Validation

Independent validation reference points for multi-temporal LULC maps were created with a stratified random sampling approach over the study area for accuracy assessment. A total of 400 spatially independent validation points were assigned to each temporal phase, independently of the GEE training samples. Landsat 8 and 9 Operational Land Imager (OLI) imagery corresponding to the respective Sentinel-2 classification periods (pre-flood: February–March 2025; during-flood: July–August 2025; post-flood: September–October 2025) was used as reference imagery for labeling the validation points. The multispectral bands used for reference interpretation have a spatial resolution of 30 m, which is coarser than the 10 m Sentinel-2 classification; this difference may introduce labeling uncertainty in fragmented landscapes and should therefore be considered when interpreting the accuracy metrics. Confusion matrices were generated to calculate overall accuracy (OA), Kappa coefficient (KC), producer’s accuracy (PA), and user’s accuracy (UA). The flood extent validation was carried out using the independent inundation data provided by the Food and Agriculture Organization [45] and the United Nations Satellite Centre [46]. Also, the visual comparison of UNOSAT flood maps derived from Sentinel-1 data to the Sentinel-1 flood products was carried out to assess the spatial agreement for July and August 2025. Moreover, the Sentinel-1 maximum flood composite was compared with the FAO seasonal flood extent dataset using a quantitative area-based comparison after resampling the reference dataset to 10 m spatial resolution using the nearest-neighbor approach.

3. Results

3.1. Multi-Temporal Flood Inundation Dynamics

The multi-temporal Sentinel-1 SAR analysis revealed marked changes in flood extents in Punjab during the monsoon season of 2025 (Figure 3a–f). The analysis reveals a well-defined bimodal flood pattern, with an early flood period in July, a flood recession period in early to mid-August, and a second, larger flood inundation period in late August into early September.

Figure 3. Spatio-temporal distribution of flood inundation across Punjab during 2025 derived from Sentinel-1 SAR data: (a) 1–12 July, (b) 13–24 July, (c) 25 July–5 August, (d) 6–17 August, (e) 18–29 August, and (f) 30 August–10 September.

The inundated area in the first phase of the flood (1–12 July) was about 3167 km², mainly in the northeastern area of Punjab (Figure 3a). The extent of flooding rose to 4048 km² on 13–24 July, with expansion occurring along the main river flood corridors and surrounding agricultural floodplains (Figure 3b). Then gradually the floodwaters subsided during late July and early August, so that by 25 July–5 August the inundated area was 2443 km² and by 6–17 August the inundated area was 1443 km² (Figure 3c,d). Although the overall flooded area decreased, residual inundation persisted in poorly drained agricultural fields and topographic depressions.

A second flood peak occurred during the last days of August. Inundation increased from 4452 km² on 18 August to reach a maximum inundation of 5449 km² between 30 August and 10 September (Figure 3e,f). In this phase, significant spatial growth was seen in central and southern Punjab, with the rivers Chenab and Ravi providing a continuous flood flow. Flooding recurrence and intensification suggest the role of long-duration monsoon rainfall and delayed drainage in the flat alluvial landscape. Gujranwala, Hafizabad, and Chiniot districts exhibited relatively high flood persistence, reflecting their location within active floodplain zones. The persistence of flooding in these districts over multiple 12-day intervals, rather than a single peak, suggests that drainage congestion, rather than rainfall intensity alone, played a substantial role in prolonging inundation in these specific locations. While the peak inundation during a single 12-day interval reached 5449 km² (30 August–10 September), the cumulative maximum spatial extent of the flood over the entire monitoring period reached 9495 km², highlighting the vast geographical footprint of the bimodal event. The change in flood extent over time is further highlighted in Figure 4.

Figure 4. Temporal variation in flood inundation area (km²) across 12-day intervals.

3.2. Validation of Flood Inundation

A spatial comparison of Sentinel-1-derived flood maps to independent UNOSAT flood products showed spatial alignment in both evaluated flood phases (Figure 5a–d). In the first flood phase (Figure 5a,b), both datasets revealed similar inundation patterns in the northeastern parts of Punjab (Gujranwala, Sialkot, and Narowal districts). During the second flood peak, both products mapped lateral floodplain expansion along the Chenab and Ravi rivers through Gujrat, Hafizabad, and Chiniot (Figure 5c,d). As UNOSAT provided vector data, it served as an operational spatial cross-validation benchmark rather than in situ ground truth.

Figure 5. Spatial comparison of flood extents over Punjab, Pakistan: (a,b) The July 2025 flood event, with (a) representing the UNOSAT-based flood extent on 10 July 2025 and (b) the Sentinel-1-based flood extent from 1 to 10 July 2025. (c,d) The August 2025 flood event, with (c) representing the UNOSAT-based flood extent for 27–28 August 2025 and (d) the Sentinel-1-based flood extent from 18 to 29 August 2025.

Quantitatively, the Sentinel-1 maximum-flood composite estimated 9495.33 km² of total inundation, compared to 13,026.19 km² from the FAO seasonal product. To evaluate spatial agreement, the 30 m FAO layer (1 August–28 September 2025) was resampled to 10 m and aligned with the Sentinel-1 raster (25 July–10 September 2025). Pixel-by-pixel comparison yielded an IoU of 0.49, a Dice coefficient of 0.66, a Producer’s Accuracy of 62.0% (Omission Error 38.0%), and a User’s Accuracy of approximately 70.0% (Commission Error 30.0%). The differences may reflect the temporal mismatch between the datasets, differences in sensor characteristics and mapping methodology, and interval-based versus cumulative flood mapping.

3.3. NDVI-Based Vegetation Change Analysis

Multi-temporal NDVI analysis showed that there were significant variations in vegetation conditions in the 2025 flood event across Punjab (Figure 6). The NDVI during the pre-flood period was 0.28, while during the peak flooding period it was 0.20, with a change of −28.6% in vegetation greenness at the provincial mean level. This reduction indicates a decline in vegetation greenness during the flood period and may reflect both seasonal variability and flood-associated vegetation stress. During the post-flood period, mean NDVI was found to increase, reaching 0.26 (Figure 6c), which represents an increase in vegetation greenness relative to the flood period. However, the NDVI remained below the pre-flood value, suggesting that vegetation conditions had not fully recovered during the observation period. The spatial distribution of ΔNDVI (Figure 6d) demonstrated that areas with the strongest vegetation decline were primarily associated with prolonged inundation zones and agricultural regions exposed to excessive water accumulation, suggesting an association between inundation and reduced vegetation greenness. The quantitative analysis further demonstrated a statistically significant, weak positive association between flood persistence and ΔNDVI (Pearson’s r = 0.165, p < 0.001), indicating that pixels exposed to inundation over a greater number of observation periods generally experienced greater reductions in vegetation greenness. The slight decrease in mean ΔNDVI observed in persistence class 6 relative to class 5 may partly reflect the temporal mismatch between the NDVI_dur composite (July–August 2025), which ended on 31 August 2025, and the final flood observation period (30 August–10 September 2025). Mean ΔNDVI and standard deviation for each flood-persistence class are presented in Table 1.

Figure 6. Normalized Difference Vegetation Index (NDVI) maps of Punjab showing (a) pre-flood, (b) during-flood, and (c) post-flood conditions, along with (d) ΔNDVI (pre-flood minus during-flood) illustrating vegetation changes associated with flood impacts.

Table 1. Vegetation response (ΔNDVI) across flood-persistence classes during the 2025 Punjab flood.

Flood Persistence Class (Observation Periods)	Sample Count (n)	Mean ΔNDVI	Standard Deviation
0 (Non-flooded)	9333	0.0479	0.1838
1	450	0.1703	0.2154
2	125	0.2049	0.2238
3	45	0.2265	0.2308
4	22	0.2468	0.2131
5	10	0.3191	0.2230
6	12	0.1810	0.2177

3.4. Land Use and Land Cover (LULC) Dynamics

The Random Forest classification was applied to characterize LULC changes under pre-flood, during-flood, and post-flood conditions (Figure 7; Table 2). The Random Forest variable-importance analysis showed that the contribution of predictors varied among the three temporal phases. During the pre-flood phase, B8, DEM, B12, and MNDWI showed relatively high importance. During the flood phase, the seasonal NDVI baseline was the most influential predictor, followed by DEM, B11, and B8. In the post-flood phase, the seasonal NDVI baseline showed the highest importance, followed by DEM, B11, and B12. These results demonstrate that spectral, topographic, and seasonal vegetation information jointly contributed to the discrimination of LULC classes under different flood instances. Before the flood, cropland represented the dominant land cover class, occupying 122,047 km² (59.4%), followed by barren land (55,383 km²; 27.0%), vegetation (15,831 km²; 7.7%), built-up areas (9961 km²; 4.9%), and water bodies (2122 km²; 1.0%) (Figure 8, Table 2).

Figure 7. Classified land use/land cover (LULC) maps of Punjab showing (a) pre-flood, (b) during-flood, and (c) post-flood conditions.

Table 2. Area (km²) of land use/land cover (LULC) classes under different conditions.

LULC Class	Pre-Flood (km2)	During-Flood (km2)	Post-Flood (km2)
Water	2122	10,360	6247
Vegetation	15,831	14,068	10,910
Cropland	122,047	113,866	117,251
Built-up	9961	8089	7366
Barren	55,383	58,961	63,570
Total	205,344	205,344	205,344

Figure 8. Area distribution (km²) of the classified LULC classes across Punjab, Pakistan.

During peak flooding, surface water extent increased substantially to 10,360 km² (5.1%), representing an expansion of approximately 8238 km². This corresponded with a reduction in cropland area to 113,866 km² (55.5%), indicating that agricultural areas were the primary land cover affected by inundation. Built-up and vegetation classes also showed reductions, reflecting temporary spectral changes caused by water coverage.

After the flood recession, water extent declined to 6247 km² (3.0%). The cropland area partially recovered to 117,251 km² (57.1%). However, barren land increased further to 63,570 km² (31.0%), indicating that a substantial area remained classified as barren during the post-flood observation period. This barren-land increase was spatially concentrated in areas that had been classified as water during peak flooding (Figure 7b vs. Figure 7c), consistent with sediment deposition rather than a shift from cropland or vegetation classes. This pattern may reflect changes in post-flood surface conditions and seasonal agricultural dynamics.

3.5. Accuracy Assessment of LULC Classification

Confusion matrix analysis (Table 3) was used to assess the performance of multi-temporal LULC classification. The overall accuracy for the classification ranged between 91% and 94%, and the Kappa coefficient ranged from 0.84 to 0.89, with consistent results across all temporal phases. The best overall accuracy (94%) was achieved for the post-flood classification, while the least accurate classification was at the active flood phase (91%) because vegetation, water, and saturated soil surfaces have similar spectral characteristics at this time. Producer’s accuracy values for individual classes were high, with cropland, vegetation, and water at 96%, 95%, and 92%, respectively, during the post-flood period.

Table 3. Accuracy assessment of multi-temporal LULC mapping.

Classes	Pre-Flood (February–March)		During-Flood (July–August)		Post-Flood (September–October)
	PA	UA	PA	UA	PA	UA
Water	0.90	1.00	0.85	0.89	0.92	0.79
Vegetation	0.97	0.79	0.78	0.72	0.95	0.80
Cropland	0.92	0.98	0.94	0.92	0.96	0.94
Built-up	0.79	0.75	0.75	1.00	0.71	0.91
Barren	0.97	0.90	0.90	0.93	0.91	0.97
OA	0.93		0.91		0.94
KC	0.87		0.84		0.89

PA = producer’s accuracy; UA = user’s accuracy; OA = overall accuracy; KC = kappa coefficient.

4. Discussion

4.1. Multi-Phase Flood Dynamics

Multi-temporal analysis of the flood period of the monsoon season of 2025 has revealed the highly dynamic nature of monsoon floods in Punjab with two distinct flood peaks. The waters that had poured into the region in the early monsoon (July–August) began to recede by early September, while another wave of floodwaters reached the region, leading to further inundation of geographical areas. The second peak coincided with a period of elevated inflows recorded at multiple Sutlej River monitoring stations, including G.S. Wala, Islam, and Sulemanki, as reported in the daily flood bulletins of the Flood Forecasting Division, alongside further inundation of surrounding agricultural floodplain areas (AFE) [47,48].

The daily discharge observations show a marked increase in Sutlej inflows across the monitoring stations during late August and early September, with peak inflows of 327.0 × 10³ cusecs at G.S. Wala, 154.2 × 10³ cusecs at Sulemanki, and 120.6 × 10³ cusecs at Islam during the observation period (Figure 9). This temporal correspondence provides independent hydrological evidence supporting the occurrence of the second flood peak. However, the available observations do not establish a direct causal relationship, and the relative contributions of river discharge, rainfall, reservoir operations, and other hydrological processes were not independently quantified in this study. For the early flood peak, the reviewed FFD bulletins did not indicate similarly elevated Sutlej inflows, and therefore the available observations do not support attributing the first peak specifically to exceptional Sutlej discharge.

Figure 9. Sutlej River inflows recorded at the G.S. Wala, Sulemanki, and Islam monitoring stations from 18 August to 10 September 2025, corresponding to the period of the second remotely sensed flood peak. Source: Flood Forecasting Division (FFD), Daily Flood Bulletins, 2025.

The multi-temporal approach with a 12-day temporal resolution proved to be successful in discriminating between transiently inundated, persistently flooded, and newly affected terrain. This temporal analysis describes the multi-phased dynamics of monsoon-influenced hydrological systems, potentially reflecting the combined influence of river discharge, rainfall, upstream reservoir operations, and other hydrological processes.

A cyclical expansion and contraction pattern is observed that matches the cycle of the hydrological patterns recorded in the multi-temporal SAR flood studies, emphasizing the need to monitor the recession phase for an accurate depiction of extended inundation dynamics [2,6]. These results also confirm the ability to employ single-date flood mapping and highlight the need for high-temporal-resolution SAR observations to characterize the temporal evolution of a flood in large river basins. The use of 12-day median composites provided representative inundation estimates for each observation window but may underrepresent short-lived or instantaneous peak flooding. Thus, the mapped extents represent interval-based inundation rather than the instantaneous maximum flood extent.

4.2. Agricultural Exposure and Vegetation Response

The integration of Sentinel-2-derived NDVI with flood extent mapping reveals pronounced flood-associated changes in vegetation condition across agricultural areas. The reduction in mean NDVI from 0.28 (pre-flood) to 0.20 (peak inundation) indicates a rapid decline in vegetation greenness, which may reflect vegetation stress associated with waterlogging and inundation [9,10]. Although mean NDVI increased in the post-flood phase (0.26), it remained below the pre-flood value within the study period. This may indicate incomplete recovery of vegetation greenness. The spatial association of low NDVI values with the flooded areas suggests that the flooded areas were largely limited to the main agricultural zone of the province [15,49]. Because the periods represent different seasonal stages, some NDVI variation may reflect phenological variability that was not independently quantified.

The quantitative correlation further supports an association between flood persistence and vegetation decline, although the overall relationship across the landscape was weak (r = 0.165, p < 0.001). This suggests that pixels exposed to inundation during a greater number of observation periods generally experienced greater reductions in vegetation greenness. However, the modest magnitude of the correlation indicates that flood persistence alone does not explain most of the spatial variability in NDVI response, which may also reflect differences in phenology, pre-flood vegetation condition, crop characteristics, and other local environmental factors.

These observations are confirmed by land use analysis, which shows a decrease in the area classified as cropland during the flood period and an increase during the post-flood period. These observations are consistent with independent assessments of agricultural inundation, while differences in temporal coverage and methodology should be considered. The overall results indicate that flooding across Punjab was spatially and temporally variable, with some areas experiencing persistent inundation, and that the effects of flooding on vegetation condition may extend beyond the immediate inundation period, consistent with vulnerability patterns reported in previous Indus Basin flood assessments [10,12,50].

4.3. Flood-Associated Changes in Classified Land Cover

Classified changes in LULC indicate substantial changes in mapped cropland area during the peak inundation period, with a 6.7% (8181 km²) reduction relative to the pre-flood cropland area of 122,047 km². This reduction represents an apparent change in classified cropland extent and should not be interpreted as direct agricultural land loss. To provide a more direct estimate of flood exposure, a spatial intersection of the independently derived maximum flood extent with the pre-flood cropland layer identified 5892 km² of pre-flood cropland within the detected flood extent, equivalent to approximately 4.8% of the pre-flood mapped cropland area. This represents flooded/inundated cropland rather than confirmed permanent crop loss or agricultural damage. Following flood recession, the mapped cropland area increased by 3385 km², corresponding to an increase of 2.7% relative to the pre-flood baseline. Nevertheless, a net difference of 3.9% (4796 km²) relative to the pre-flood mapped cropland area remained at the end of the monitoring period (September–early October 2025), indicating that the pre-flood mapped cropland extent remained below the pre-flood level within the observation window. This remaining difference may reflect several factors, including soil-moisture conditions, delayed sowing of winter crops (Rabi season), and spatial or temporal differences in land-surface conditions. According to the FAO dataset [45], in geospatial flood assessment, 9% (11,251 km²) of the agricultural fields in Punjab were flooded. Notwithstanding methodological differences in temporal tracking, both datasets show that agriculture has the potential to leave strong signals at a very consistent regional spatial scale, thus supporting the macro-level patterns identified by our multi-sensor framework.

In addition to agriculture, LULC results show considerable disturbances in the peri-urban and ecological environment. The temporary reduction in the built-up classification may reflect the inundation of low-lying peri-urban areas, where floodwater alters the spectral response of built-up surfaces and can affect their classification in optical imagery [51]. One of the significant findings of this research is the progressive growth of barren areas after the flood. The post-flood increases in areas classified as barren may be associated with changes in surface conditions following flood recession, including possible sediment deposition that can modify the physical and spectral characteristics of exposed surfaces [52,53].

4.4. Methodological Contributions, Limitations, and Uncertainty Assessment

The multi-temporal Sentinel-1 SAR change-detection framework used six consecutive 12-day median composite periods, a pre-flood VV backscatter baseline, a hybrid threshold of relative backscatter change > 3.5 dB and absolute VV backscatter <−13 dB, spatial masking, and independent validation against FAO and UNOSAT datasets. This approach provides a reproducible and scalable framework for regional flood monitoring in monsoon-dominated agricultural environments. Change-detection approaches have been used in SAR-based flood mapping to address variability in backscatter responses associated with heterogeneous land-cover conditions [54].

Despite these advantages, several sources of uncertainty remain. C-band VV backscatter is sensitive to surface roughness associated with wind and precipitation, which may result in missed inundation in some areas [55]. Variations in vegetation structure and soil moisture can also produce backscatter responses similar to shallow inundation or irrigated agricultural fields, increasing classification uncertainty [22,23,26]. Irrigation channels, reservoirs, wetlands, and seasonal water surfaces may also contribute to false-positive inundation detections in intensively irrigated areas. This is a particularly relevant limitation for Punjab, given its extensive canal-irrigation network (Section 2.1), and future applications of this framework in similarly irrigated basins should consider incorporating a canal/reservoir exclusion mask derived from existing infrastructure datasets. Mixed-pixel effects may further influence flood delineation along urban–agricultural boundaries at 10 m spatial resolution. Furthermore, the absence of coordinated field observations, including flood depth and crop-level damage measurements, limited direct quantitative validation. Moderate spatial agreement with FAO primarily reflects the temporal coverage offset into late September, alongside sensor, resolution, and aggregation differences. Lacking synchronous field data, FAO and UNOSAT served as secondary satellite proxies rather than absolute ground truth.

A further limitation concerns the exclusion of shallow or intermittent flooding below the sensitivity of the selected backscatter thresholds (Section 2.3.2); such under-detection may disproportionately affect low-lying rice-growing areas where standing water depths are naturally shallow, potentially leading to a conservative estimate of true agricultural flood exposure. Overall, the proposed framework offers a scalable and transferable methodology for flood impact assessment in data-scarce, monsoon-dominated regions. However, threshold parameters require calibration to local surface, vegetation, and hydrological conditions before transferring the workflow to other regions.

5. Conclusions

This study developed a multi-temporal, multi-sensor remote sensing framework integrating Sentinel-1 SAR and Sentinel-2 optical data within Google Earth Engine to assess flood dynamics and associated agricultural cropland changes during the 2025 monsoon floods in Punjab, Pakistan. Six consecutive 12-day Sentinel-1 composite windows captured a complex bimodal inundation pattern, with the second flood peak temporally coinciding with elevated Sutlej inflows and prolonged waterlogging. The resulting hydrological stress was reflected in a decline in mean NDVI from 0.28 to 0.20, while spatial analysis indicated that 5892 km² of pre-flood cropland was directly inundated. Although cropland extent partially recovered during recession, a 3.9% net reduction persisted by the end of the observation period, indicating delayed post-flood recovery rather than permanent land-use conversion. Comparison with independent UNOSAT and FAO datasets further supported the reliability of the mapped flood extents, while also highlighting the effects of differences in temporal coverage.

Overall, the results demonstrate that the integrated Sentinel-1/Sentinel-2 framework provides a scalable and reproducible approach for regional flood and agricultural impact assessment in data-scarce, monsoon-dominated environments. The study establishes a quantitative baseline for future comparison of flood dynamics and agricultural vulnerability in Punjab. Future work should integrate satellite observations with hydrological/hydraulic modelling, field-based flood-depth measurements, and crop production data, while higher-frequency SAR observations or data assimilation could improve temporal characterization and near-real-time flood impact assessment.