VAWI_log: A Log-Transformed LSWI–EVI Index for Improved Surface Water Mapping in Agricultural Environments

Abstract

Detecting surface water beneath vegetation canopies remains a major challenge for widely used water indices, which often underestimate water obscured by vegetation. This limitation is further compounded by the scarcity of reliable in situ data needed for robust index development and validation. To address this, we introduce a Vegetation-Adjusted Water Index using a logarithmic transformation of the ratio between the Land Surface Water Index (LSWI) and the Enhanced Vegetation Index (EVI), referred to as VAWI_log. This transformation compresses high vegetation values while expanding the range typical of water surfaces, enhancing contrast in mixed land cover areas and improving class separability. The index was developed and validated using in situ water level measurements, providing a strong empirical basis for detecting surface water under variable vegetation conditions. VAWI_log consistently outperformed established indices in detecting surface water beneath vegetation, demonstrating superior detection accuracy and overall performance, with a balanced accuracy (BA) of 0.69 and a producer’s accuracy (PA) of 0.80. This reflects an average improvement of 25% over conventional methods. Benchmarking against the Dynamic World V1 dataset further confirmed the improved capability of the proposed index to detect surface water in vegetated areas, which are often missed by the dataset. This enhanced performance can be attributed to the fact that the proposed index was developed using on-the-ground observations, rather than relying solely on expert-labeled imagery as in the case of Dynamic World. Overall, VAWI_log provides a simple yet effective solution for improved surface water mapping in vegetated landscapes. Its compatibility with open-source optical satellite data enables broader applications, including irrigation monitoring and greenhouse gas assessments.

1. Introduction

Optical remote sensing satellites (e.g., Landsat, Sentinel, and MODIS) have significantly advanced the monitoring of surface water at large spatial scales by providing consistent and accurate observations from space. This monitoring is typically achieved through the use of spectral band combinations derived from top-of-atmosphere or surface reflectance data. These combinations, commonly referred to as water indices, enhance the spectral contrast between water and non-water features, facilitating more accurate detection. Numerous water indices have been extensively employed to detect open surface water across a variety of environments, including lakes, reservoirs, and rivers [1,2,3]. One of the earliest and most widely used indices for surface water detection is the Normalized Difference Water Index (NDWI), which utilizes reflectance values from the near-infrared (NIR) and green spectral bands [4]. A modified version of NDWI was later introduced [5], which replaces the NIR band with shortwave infrared (SWIR) reflectance to improve water delineation, particularly in urban areas characterized by complex land cover. Additional indices, such as the Water Ratio Index (WRI) and the Automated Water Extraction Index (AWEI), have also been developed to address the challenges of detecting surface water in heterogeneous landscapes comprising vegetation, built-up areas, and water bodies [6,7,8].

However, the detection of surface water beneath vegetation canopies remains a significant limitation of commonly used water indices. These indices often underestimate the presence of standing water obscured by vegetation, particularly in densely vegetated environments [9]. In vegetative environments, optical remote sensing has primarily been applied to estimate canopy or vegetation water content, with a strong emphasis on drought monitoring and crop yield prediction [10,11,12]. In these applications, vegetation indices, such as the Normalized Difference Vegetation Index (NDVI), are commonly used to assess vegetation health and density by exploiting the high reflectance of the NIR band and the strong absorption of the red band by healthy vegetation [13]. Similarly, the Enhanced Vegetation Index (EVI) has been developed to improve sensitivity to atmospheric conditions and reduce background noise, making it especially effective in regions with high biomass or cloud cover [14].

Nevertheless, there is a growing need to detect not only vegetation water content but also the presence or absence of surface water beneath vegetation, particularly to support irrigation management at both field and regional scales [15,16]. This need is also critical for accounting for greenhouse gas emissions resulting from prolonged flooding in agricultural systems such as paddy rice cultivation [17,18,19]. Several water indices have been developed to detect the presence of surface water resulting from natural flooding or irrigation, primarily by capturing variations in surface wetness. The Land Surface Water Index (LSWI) has been widely adopted in applications related to wetlands, irrigation monitoring, and drought assessment, owing to its sensitivity to both vegetation water content and soil moisture. It utilizes reflectance in the NIR and SWIR spectral regions [20,21,22]. LSWI, either on its own or in combination with EVI, has been extensively employed for mapping flooded rice paddies and estimating irrigation events [23,24,25,26]. The LSWI and also NDWI [27,28] are generally effective only during the early stages of crop growth, when vegetation cover is still minimal. The Normalized Difference Flood Index (NDFI), which combines red and SWIR2 reflectance bands, has been applied to estimate surface water in Italian rice-growing regions [29]. Its performance significantly declines with increasing vegetation density and vigor. In contrast, the Water in Wetlands (WIW) index was specifically designed to detect surface water in wetland environments, including vegetated areas such as forests, rice paddies, and shrublands, regardless of vegetation presence or absence [30]. However, the applicability of these indices for detecting surface water beneath vegetation remains challenging, as water indices often exhibit strong correlations with vegetation indices, thereby complicating their interpretation and increasing the risk of misclassification [31,32,33].

Furthermore, progress in this domain is constrained by the limited availability of reliable in situ data required for robust index development and validation. While satellite imagery, expert interpretation, and field surveys are commonly used for remote sensing validation [34,35,36], the presence of accurate and representative ground-based measurements remains essential for ensuring the credibility and applicability of remote sensing products, particularly in complex land cover conditions [37]. In response to these challenges, this study proposes a Vegetation-Adjusted Water Index (VAWI), developed through a systematic exploration of different formulations based on the Land Surface Water Index (LSWI) and the Enhanced Vegetation Index (EVI). The objective is to reduce the confounding influence of vegetation on water detection and thereby improve the reliability of identifying surface water under dense vegetative cover. LSWI was chosen as the basis for this development because, unlike other water indices, it incorporates information related to both soil moisture and vegetation water content. By correcting LSWI with EVI, which reflects vegetation greenness and canopy density, the proposed VAWI aims to suppress the vegetation signal while retaining the soil moisture component. A key innovation of this research is the integration of in situ water level measurements, which captures a wide range of water level variability both above and below the soil surface, collected throughout an entire rice cropping season. These measurements serve as ground truth for optimizing index thresholds and evaluating performance. Performance assessment includes a comparative analysis of the best-performing VAWI formulation against standard indices and dataset used for open surface water and wetland monitoring, as well as an evaluation of index performance across varying vegetation density.

This approach harnesses the synergy between satellite observations and field-based data to develop a modified water index tailored to improve surface water detection in rice paddies. Enhanced detection capabilities enabled by VAWI aim to extend the effective use of optical satellite data beyond early crop stages, thereby facilitating more accurate mapping of irrigation dynamics, supporting water balance assessments, and refining greenhouse gas emission estimates in paddy rice systems.

2. Data and Methods

2.1. On-the-Ground Water Level Measurements

2.1.1. Sensor Installation

The field area for on-the-ground water level sensor installation is in Pangasinan, a province located in the northern Philippines, as shown in Figure 1a. The Philippines predominantly follows a double-cropping system, characterized by distinct wet and dry seasons [38]. Wet season cropping typically occurs from June to October, while dry season cropping is generally from November to March [39], illustrated in Figure 1c. The timing of transplanting and harvesting across the country varies depending on the availability of irrigation water, which continues to be a major challenge in agricultural management [40].

A total of 100 sensors were deployed during the dry season of November 2024 to February 2025 across the study area, organized into three main clusters for comprehensive spatial coverage and variability, as shown in Figure 1b. The variation in elevation within the study site contributes to differences in water movement and retention [41]. Lower-lying areas tend to accumulate surface water, which can support crop growth but also increase the risk of flooding that may damage plants and deplete soil nutrients. Including fields at different elevations, as illustrated in Figure 1d, allows the study to capture the unique patterns of water dynamics associated with varied terrain.

The water level sensors are ultrasonic devices that measure water levels by emitting ultrasonic waves and calculating the time taken for the reflected waves to return, as shown in Figure 2a. These sensors provide high-resolution measurements at 15 to 16 min intervals, with an accuracy of ±1 cm. Regular maintenance was performed to minimize potential errors and ensure measurement accuracy. This involved removing accumulated mud beneath the perforated pipe, clearing leaf debris, and guaranteeing that sensor measurements remained consistent with manual measurements across all sensors. Each sensor is mounted on a 30 cm PVC pipe, shown in Figure 2b, with 10 cm extending above the ground and 20 cm positioned below, capturing a broad range of water level measurements. Water levels from the soil surface up to +10 cm are classified as indicative of water presence, while levels from the soil surface down to −20 cm are indicative of water absence. Thus, even if the soil appears moist when water levels are by the soil surface, it will still be considered indicative of water absence.

2.1.2. Time Series Smoothing and Dynamic Outlier Detection

Despite strict maintenance protocols, sensor malfunctions and obstructions, from leaves and other natural disturbances, can introduce outliers in the water level measurements. For data quality maintenance, a dynamic outlier detection algorithm using a sliding window is applied. For this process, a window size of n = 15 is selected to detect unnatural fluctuations in water levels. Given the 15–16 min interval between consecutive measurements, this window spans approximately 3.75 to 4 h of data. This duration is sufficient to capture short-term patterns while remaining sensitive to abrupt anomalies. Water levels, regardless of influences such as irrigation, rainfall, or drainage, are generally expected to follow a consistent rising or falling trend. By using a relatively small window, the algorithm conservatively flags outliers occurring over brief periods. The dynamic mean and standard deviation are computed for each time t using a 15-point centered window, as shown in Equations (1) and (2):

$${\mu }_t=\frac{1}{n}\sum _{i=t-\frac{n-1}{2}}^{t+\frac{n-1}{2}}{x}_i$$

(1)

$${\sigma }_t=\sqrt{\mathrm{Var}\left(\right|x_i\left|\right)}$$

(2)

$x_i$ represents the water level measurements at time t, and ${\mu }_t$ and ${\sigma }_t$ denote the dynamic mean and standard deviation at time t, respectively, calculated within the sliding window. A measurement is identified as an outlier if its deviation from the dynamic mean exceeds twice the dynamic standard deviation. The water level measurement is retained only if it falls within an acceptable range, i.e., within two standard deviations from the dynamic mean. This process is summarized in Equation (3):

$$\mathrm{WL}=\mathrm{outlier}\mathrm{if}|\mathrm{WL}-{\mu }_t|\ge 2{\sigma }_t$$

(3)

Figure 3 presents three sample water level time series from selected fields, highlighting the detection and removal of significant outliers, from abrupt spikes or drops, attributed to sensor malfunctions or temporary obstructions. By applying a short time window to dynamically compute the local mean and standard deviation, the water level time series are effectively smoothed for more consistent data. The sample time series also showcases the distinct water level trends unique to each field, emphasizing the variability in hydrological patterns across different paddy fields.

2.2. Satellite Data

2.2.1. Data Pre-Processing

This study utilized surface reflectance data from the Harmonized Landsat and Sentinel-2 (HLS) project [42], which integrates imagery from the Landsat Operational Land Imager (OLI) and the Sentinel-2 MultiSpectral Instrument (MSI). The harmonized product maximizes the availability of optical satellite observations by combining the 16-day revisit cycle of Landsat with the 5-day revisit cycle of Sentinel-2 at the equator, enabling a more frequent and consistent monitoring of land surface conditions. Both sensors offer a spatial resolution of 30 m, providing adequate detail for sub-field level analysis. The surface reflectance products are atmospherically corrected and preprocessed to reduce atmospheric noise, making them well-suited for time-series analysis and change detection applications [43,44]. Cloud and cloud shadow masking were performed using the FMask quality assurance (QA) bit flags embedded within the image collection. The spectral bands used in the study include the visible bands (Red, Green, and Blue), as well as the Near-Infrared (NIR) and Shortwave Infrared (SWIR1/2) bands, which are commonly used in vegetation and water-related indices. Following this pre-processing step, the Land Surface Water Index (LSWI) and Enhanced Vegetation Index (EVI) were calculated using their respective spectral band combinations, as defined in Equations (4) and (5):

$$LSWI={\displaystyle \frac{NIR-SWIR1}{NIR+SWIR1}}$$

(4)

$$EVI=2.5·{\displaystyle \frac{NIR-Red}{NIR+6.0·Red-7.5·Blue+1}}$$

(5)

2.2.2. Temporal Alignment of Satellite and Water Level Data

Satellite data provides field information at regular intervals. With revisit frequency of at least five days, it is much less frequent than the 15–16 min interval of water level measurements. As a result, satellite data serves as a practical basis for temporal alignment with the more frequent in situ measurements. This process involves synchronizing the date and time of both datasets to a common time zone, in this case, Japan Standard Time (JST). While any time zone of choice is acceptable, making sure that both datasets adhere to the same time zone is necessary for maintaining consistency. Temporal alignment is achieved by identifying the closest date-time from the water level dataset to each satellite date-time, then associating the corresponding water level values with the satellite band values, as shown in Equation (6):

$$\mathrm{aligned}\_\mathrm{data}=\{t_{si},x_{si},t_{wi},x_{wi}\}\mathrm{minimize}\left|t_{si}-x_{si}\right|$$

(6)

where $t_{si}$ and $t_{wi}$ represent the satellite and water level date-times, and $x_{si}$ and $x_{wi}$ are the corresponding satellite and water level values. Finally, to account for potential installation errors and sensor sensitivity, data pairs with sensor readings between −1.0 cm and 1.0 cm were removed from the aligned dataset. Since a threshold of 0 cm is critical for determining water presence from water level measurements, even subtle errors in installation or sensor sensitivity could lead to misinterpretations of the satellite-derived results.

Figure 4a presents a time series boxplot of EVI throughout the dry season cropping period, at least for the field data collection locations. To provide context for the observed EVI values, key growth stages are also illustrated in Figure 4c, reflecting actual field conditions. These correspondences were verified with field staff on the ground. Figure 4b depicts data availability throughout the cropping season. The steady increase in data from mid-November to early December 2024 corresponds to the installation of water level sensors, coinciding with the completion of land preparation and the start of transplanting. From December to January 2025, data availability remains relatively stable. By late January 2025, fields that started activities earlier begin harvesting, leading to a gradual decline in available data. By early February, the remaining unharvested fields undergo terminal drainage in preparation for harvesting within the next one to two weeks, resulting in a reduction in water presence during this period.

A total of 682 aligned data points were obtained by temporally aligning satellite observations with water level measurements, comprising 218 instances of water presence and 464 instances of water absence. Water presence was observed throughout the season, confirming that controlled flooding is essential for rice growth and must be accurately captured. This flooding pattern indicates two possible scenarios: (1) some fields practice continuous flooding, while (2) others follow a rotational flooding schedule, where different groups of fields are alternately flooded and drained on different days.

2.3. Vegetation-Adjusted Water Index Formulations and Their Interpretations

The Vegetation-Adjusted Water Index (VAWI) is proposed to address the persistent challenge of detecting surface water beneath vegetative cover by minimizing the influence of vegetation. In this study, multiple vegetation correction transformations were formulated by combining LSWI and EVI in different mathematical structures. The corrected VAWI formulation represents a simple linear difference that reduces vegetation interference by subtracting the EVI from the LSWI. This operation lowers the values for vegetation-dominated pixels while preserving higher values for water, which typically has high LSWI and low EVI. The normalized difference form scales the difference between LSWI and EVI by their combined magnitude, accounting for overall signal intensity and reducing bias in mixed pixels. This enhances the contrast between water and vegetated surfaces. In the weighted formulation, the water signal (LSWI) is adjusted by the inverse of EVI. Since vegetation has high EVI, the resulting weight becomes small, suppressing the signal in vegetated areas. For water bodies with low EVI, the weight remains high, preserving the original LSWI. This non-linear scaling effectively reduces vegetation influence. The normalized vegetation-adjusted form modifies the LSWI–EVI difference based on vegetation dominance. When EVI is high, the adjustment reduces the index value, while for low EVI, the formula amplifies the water signal. This enhances responsiveness to vegetation presence. Finally, the log-transformed variant applies a logarithmic transformation to the moisture-to-vegetation ratio. This compresses high vegetation values and expands low values typical of water surfaces, improving contrast in areas with subtle differences and making class separation more symmetric and linear.

A simple one-dimensional grid search optimization is employed to identify the optimal threshold for each vegetation-adjusted water index transformation. Five hundred (500) equally spaced candidate thresholds are generated based on the minimum and maximum values of each index. These thresholds are then individually assessed to determine their performance. Balanced accuracy is used as the optimization criterion, and the threshold that yields the highest balanced accuracy is selected as the optimal threshold. These candidate formulations were systematically evaluated based on a set of performance metrics described in Appendix A. The transformation that demonstrated the highest overall performance, as summarized in

Table 1. Mathematical formulations of vegetation-adjusted water index (VAWI) variants based on LSWI and EVI transformations.

LSWI − EVI Transformation	Formula
Simple Correction	${\mathrm{VAWI}}_{\mathrm{corrected}}=LSWI-EVI$
Normalized Difference 1	${\mathrm{VAWI}}_{\mathrm{ND}}={\displaystyle \frac{LSWI-EVI}{LSWI+EVI+\epsilon }}$
Vegetation Weighted	${\mathrm{VAWI}}_{\mathrm{weighted}}=LSWI·(1-EVI)$
Vegetation Normalized	${\mathrm{VAWI}}_{\mathrm{Norm}}={\displaystyle \frac{LSWI-EVI}{1-EVI}}$
Log Transformed 1	${\mathrm{VAWI}}_{\log }=log\left({\displaystyle \frac{1+\mathrm{LSWI}+\epsilon }{1+\mathrm{EVI}+\epsilon }}\right)$

¹$\epsilon$ is a small constant (e.g., ${10}^{-6}$) added to maintain index stability.

, is recommended as the VAWI.

2.4. Comparative Analysis Against Established Water Indices and Dynamic World V1

Following the identification of the proposed VAWI transformation, its performance is benchmarked at two levels. First, it is compared against existing water indices specifically designed for detecting surface water in vegetated environments. The formulae and detection mechanisms of these indices are summarized in

Table 2. Water indices and their corresponding detection mechanisms.

Water Indices	Formula	Detection Mechanism
NDWI	$\frac{\mathrm{Green}-\mathrm{NIR}}{\mathrm{Green}+\mathrm{NIR}}$	Optimum thresholding
LSWI	as in Equation (4)	Optimum thresholding
NDFI	$\frac{\mathrm{Red}-\mathrm{SWIR}2}{\mathrm{Red}+\mathrm{SWIR}2}$	Optimum thresholding
WIW	Landsat: $\mathrm{NIR}\le 0.1735,\mathrm{and}\mathrm{SWIR}2\le 0.1035$ Sentinel-2: $\mathrm{NIR}\le 0.1804\mathrm{and}\mathrm{SWIR}2\le 0.1131$

. To assess whether performance differences between VAWI and the benchmark indices were statistically significant, we applied a non-parametric paired significance testing framework. Classification performance was quantified using balanced accuracy. To account for sampling variability, we employed bootstrapping with 1000 iterations, resampling the data with replacement and computing metrics for each iteration. Pairwise comparisons between VAWI and each alternative index were conducted using the Wilcoxon signed-rank test [45]. This test was chosen for its robustness to non-normal distributions and its appropriateness for comparing paired performance samples. To control for multiple comparisons, p-values were adjusted using the Benjamini–Hochberg procedure [46]. Differences were considered statistically significant at an adjusted p-value threshold of 0.05.

Second, VAWI is evaluated against Dynamic World V1, a widely used near-real-time (NRT) dataset that classifies Sentinel-2 imagery into eight land use/land cover (LULC) categories: water, trees, grass, flooded vegetation, crops, shrubs and scrub, built-up areas, bare ground, and snow and ice [34]. The dataset is developed using semi-supervised learning with expert-labeled training samples. The “flooded vegetation” class in Dynamic World is annotated only when water and vegetation are visibly intermixed in the satellite imagery—typically where emergent vegetation allows surface water to remain discernible. As a result, training samples for this class predominantly correspond to low-density vegetation, often associated with low EVI values.

3. Results

3.1. Vegetation Influence Reduction and Performance of LSWI–EVI Transformations

With the primary objective of minimizing vegetation influence, Figure 5 presents the distribution of index values from each transformation plotted against vegetation density, as indicated by EVI. The baseline plot of LSWI against EVI reveals a positive correlation, indicating that LSWI values tend to increase with increasing vegetation density. All transformations demonstrate significant reductions in this correlation, evidenced by the decreased slopes of their trendlines, except for VAWI_Norm. Although VAWI_Norm reverses the correlation direction to negative, it does not exhibit a meaningful improvement in independence. This negative trend is likely due to the subtraction of EVI in both the numerator and denominator of its formulation, causing index values to decrease as vegetation increases. VAWI_log stands out among the tested transformations, exhibiting a near-zero slope and a nearly horizontal trendline, suggesting effective decoupling from vegetation influence.

The five LSWI–EVI transformation candidates exhibited varying levels of performance, as evaluated using balanced accuracy (BA), producer’s accuracy (PA), and user’s accuracy (UA). These metrics, respectively, reflect the overall classification effectiveness, the accuracy in detecting surface water under vegetation, and the reliability of detections, as illustrated in Figure 6. Among the transformations, VAWI_log achieved the highest balanced accuracy and producer’s accuracy, with BA = 0.69 and PA = 0.80, while also demonstrating a comparable user’s accuracy (UA = 0.47). A close second was VAWI_corrected, which yielded a BA of 0.68, PA of 0.73, and a slightly higher UA of 0.49.

Figure 5, particularly the distribution of labeled surface water presence and absence points, illustrates the rationale behind the formulation of VAWI_log. This index effectively reduces the influence of vegetation, thereby improving the separability between surface water presence and absence classes under vegetated conditions. This visual evidence supports the superior performance observed for this index. Moreover, Figure 5 and Figure 6 reveal a clear relationship between the slope of the index distribution trendline, representing the degree of vegetation influence, and the producer’s accuracy in detecting surface water beneath vegetation. Indices with flatter slopes indicate reduced sensitivity to vegetation density and tend to perform better in surface water detection. In particular, VAWI_log, which exhibited the flattest slope, achieved the highest producer’s accuracy, followed by VAWI_weighted (PA = 0.77) and VAWI_corrected. In contrast, VAWI_Norm, which reversed the correlation but showed limited improvement in vegetation suppression, recorded the lowest detection accuracy (PA = 0.60).

Demonstrating superior performance and stability across increasing vegetation cover, VAWI_log emerged as the most effective transformation among the five LSWI–EVI variants evaluated.

3.2. Benchmarking Against Established Water Indices

In this section, we evaluate the performance of VAWI_log in comparison to several well-established water indices, focusing on both overall detection accuracy and robustness under increasing EVI.

3.2.1. Overall Detection Performance

Figure 7a presents the performance scores of VAWI_log alongside NDWI, LSWI, WIW, and NDFI, evaluated against ground truth observations. All these conventional indices demonstrated suboptimal performance, with balanced accuracy (BA) scores ranging only from 0.54 to 0.56. For NDWI, WIW, and NDFI, this lower performance primarily stems from their limited ability to detect surface water under vegetation cover. In contrast, LSWI exhibited strong producer’s accuracy (PA), indicating effective detection of surface water presence; however, its low user’s accuracy (UA = 0.35) suggests considerable overestimation. VAWI_log, while achieving similarly high PA to LSWI, significantly improved UA due to reduced vegetation influence.

VAWI_log achieved an average of 25% improvement in balanced accuracy against the other indices, highlighting its enhanced accuracy for detecting surface water under vegetative cover. Statistical testing using the Wilcoxon signed-rank test further confirmed that these improvements are statistically significant (p-value < 0.01), as shown in Figure 7b.

Figure 8 presents box plots of NDWI, LSWI, NDFI, and VAWI_log values under surface water presence and absence conditions. WIW is excluded from this comparison because it is based on direct thresholding of NIR and SWIR reflectance and does not produce continuous values. The distributions indicate substantial overlap between presence and absence classes for NDWI, LSWI, and NDFI, highlighting their limited discriminative ability under complex surface conditions. In contrast, VAWI_log shows a more distinct separation, with higher values consistently associated with surface water presence and lower values with its absence. This suggests that the logarithmic transformation of the LSWI–EVI ratio improves class separability, particularly in mixed or vegetation-covered areas where traditional indices tend to underperform.

3.2.2. Detection Stability with Increasing EVI

Evaluating the performance of VAWI_log under varying vegetation conditions is essential to demonstrate its effectiveness and consistency, especially in areas with dense crop cover. Figure 9a,b show the Balanced Accuracy (BA) and Producer’s Accuracy (PA), respectively, for NDWI, LSWI, WIW, NDFI, and VAWI_log across increasing vegetation density (EVI). Figure 9c presents the number of surface water presence and absence instances per EVI range, confirming that surface water remains present even at high vegetation densities, highlighting the need for reliable detection beyond early crop stages.

Figure 9a shows that VAWI_log consistently outperforms all other indices across the entire range of EVI values, demonstrating superior detection capabilities regardless of vegetation density. While LSWI shows competitive performance at low to moderate vegetation levels, its accuracy declines at both ends of the spectrum, whereas VAWI_log maintains high detection performance throughout. Further insights into the limitations of each index are provided in Figure 9b. NDWI, WIW, and NDFI perform reasonably well under low-EVI conditions, corresponding to early growth stages shortly after transplanting. However, their detection capabilities decline rapidly beyond EVI = 0.2 and become ineffective (PA ≈ 0) beyond EVI = 0.41, as vegetation becomes denser and begins to obscure surface water. LSWI shows increasing PA with higher vegetation cover, but this trend is potentially misleading. Its strong correlation with EVI suggests that elevated detection rates at high vegetation densities may stem from overestimation, as the index conflates vegetation moisture with surface water. In contrast, VAWI_log stands out for its consistent and accurate detection across all vegetation levels, effectively distinguishing surface water without being confounded by vegetation influence.

These findings are reinforced by the surface water presence probability heatmaps shown in Figure 10, which are derived using the optimal threshold for each index and validated against ground-based observations. NDWI, LSWI, and NDFI fail to provide clear separability between surface water presence and absence across varying vegetation conditions, while VAWI_log maintains strong separability throughout, demonstrating its reliability and suitability for detecting surface water under vegetated environments.

3.3. Benchmarking Against Dynamic World V1 Dataset

To assess Dynamic World’s capability in detecting surface water within vegetated agricultural areas, we compare its land cover classifications with in situ water level measurements across the study period. All monitored fields are confirmed rice-growing areas, where correct classifications should predominantly fall under either “crops” or “flooded vegetation.” As shown in Figure 11a, 80.5% of observations are labeled as crops, while only 3.0% are classified as flooded vegetation. Misclassifications occur, with some fields incorrectly labeled as trees (8.8%), built-up areas (2.5%) or others. For performance evaluation, Figure 11b simplifies the classification scheme into binary categories: surface water presence (water and flooded vegetation) and absence (all other classes). This enables a direct comparison with both sensor-based measurements and VAWI_log performance.

The performance comparison in Figure 12 highlights the significant limitations of the Dynamic World dataset. Dynamic World achieves a BA of only 0.53, largely due to its severe underestimation of flooded vegetation, reflected in a low detection accuracy (PA) of just 0.11. This indicates that the dataset detects surface water presence in vegetated areas only 11% of the time. This shortcoming aligns with the nature of the training data preparation, where “flooded vegetation” is labeled only when water is visibly discernible, typically under sparse vegetation conditions. As such, Dynamic World may be of limited use beyond the early stages of the rice-growing season, such as during land preparation or transplanting, when vegetation cover is minimal. In contrast, VAWI_log demonstrates better performance, owing to its development in conjunction with ground-based water level measurements. The results underscore the value of integrating remote sensing with field observations to develop indices that outperform conventional water detection methods and datasets.

4. Discussion

In Developing Surface Water Maps

The capabilities of VAWI_log in developing high-resolution and accurate general surface water maps are compared against NDWI and Dynamic World V1 near-real time (NRT). Figure 13 presents surface water maps for an area in Pangasinan, Philippines, across two acquisition dates. The first panels show the EVI map, showing the variable vegetation conditions (e.g., no, sparse, moderate and dense) that influence water visibility. The second panels display the Dynamic World V1 land cover classification, used for visual reference. The third and fourth panels show surface water segmentation results derived from VAWI_log and NDWI, respectively. These maps cover a broad area encompassing diverse land cover types to demonstrate the consistency of surface water detection across different landscapes and seasonal conditions.

For open surface water types, particularly rivers and reservoirs, VAWI_log exhibited a tendency to underestimate surface water extent. This limitation may be attributed to its LSWI-based formulation, where the low reflectance of NIR in open water bodies can reduce the index’s sensitivity. Since LSWI relies on the difference between NIR and SWIR, its effectiveness can be diminished in detecting pure open water, where both NIR and SWIR reflectance values are typically very low, leading to weak contrast [4,47]. Thus, NDWI, which leverages strong water absorption in the green band and low reflectance in the NIR, demonstrated better performance in detecting open surface water bodies.

In forested areas, VAWI_log tends to overestimate the presence of surface water, whereas NDWI more effectively avoids false detections in these regions. This overestimation is likely due to inflated LSWI values in dense vegetation canopies, where the NIR reflectance remains relatively high and the SWIR reflectance can be suppressed by internal canopy moisture or shadows, misleadingly mimicking surface wetness. Since VAWI_log has so far been validated primarily in low-stature vegetation environments such as rice paddies, its applicability in complex, mixed vegetation landscapes, including forests, wetlands, and tree-covered floodplains, warrants further investigation. These patterns of underestimation in open water areas and overestimation in forested or densely vegetated regions are quantified in Figure 14, which illustrates the extent of surface water mapping error associated with the proposed index.

In areas with flooded vegetation and croplands, VAWI_log shows strong agreement with the flooded vegetation class in Dynamic World V1 especially during 21 January 2025 capture data in Figure 13a. In contrast, NDWI significantly underestimates the extent of flooding in these vegetated regions. Notably, VAWI_log also identifies additional flooded areas, particularly within the croplands on the right side of the image, which are not captured by Dynamic World.

Consequently, Dynamic World’s flooded vegetation detection is likely accurate only under conditions of low vegetation cover. As vegetation density increases, particularly in areas with moderate to dense vegetation, as in Figure 13b, Dynamic World tends to miss flooded regions that VAWI_log is able to detect. This highlights the enhanced capability of VAWI_log to capture surface water beneath vegetation, even in comparison to a semi-supervised product like Dynamic World. Supporting this, Venter et al. [48] reported that Dynamic World has difficulty detecting flooded vegetation, with class-specific accuracies as low as 53%, underscoring the limitations of such datasets in complex, vegetated environments.

For additional context, rice farmers in the region typically follow two main cropping seasons: one during the dry season and another during the wet season. However, in areas with reliable water access—particularly fields located near reservoirs or dams—a third cropping cycle is sometimes feasible. In such cases, farmers may adjust their schedules, initiating the dry season earlier than usual to accommodate a second dry-season crop before the onset of the wet season. Among the study sites, fields designated for general data collection adhere to the conventional two-season cropping pattern. In contrast, the location selected for the surface water mapping demonstration benefits from proximity to a major water source, allowing for third-cropping practices. The image captured on 21 January 2025 corresponds to the early stage of this third cropping cycle, with the fields advancing toward the maturity phase by the time of the 20 February 2025 acquisition.

Building on the observed behavior of VAWI_log and conventional open-water indices across different land cover types, we propose a combined approach that integrates VAWI_log, any established water index (such as NDWI, MNDWI, AWEI, WRI or others), and the Dynamic World V1 dataset to produce more comprehensive surface water maps. While NDWI is used here for demonstration purposes, the framework is designed to accommodate any open-water index suited to the target landscape. In this approach, NDWI is applied to detect open water and water under tree canopies, as identified through Dynamic World V1, whereas VAWI_log is employed for other land cover classes such as crops, flooded vegetation, and shrubs. A demonstration of this integrated method is presented in Figure 15, using two observation dates: (a) 21 January 2025 and (b) 25 February 2025.

5. Conclusions

This study systematically explored various transformations of the LSWI-EVI relationship with the primary objective of reducing the influence of vegetation on surface water detection. Among all the transformations tested, the logarithmic transformation of the LSWI-to-EVI ratio, referred to as VAWI_log, emerged as the most effective. It consistently demonstrated the highest overall performance and detection accuracy, largely due to its ability to suppress vegetation interference while enhancing the separation between water and vegetative signals.

VAWI_log was then benchmarked against several widely used water indices designed for mapping open surface water, flooded vegetation, and wetlands. Across these diverse water conditions, VAWI_log consistently outperformed the other indices, not only in terms of overall detection accuracy, but also in maintaining stable performance across increasing vegetation density. Furthermore, when compared to the publicly available Dynamic World V1 dataset, VAWI_log demonstrates superior performance in detecting flooded vegetation. Notably, it identifies water presence in areas where Dynamic World fails to do so, underscoring its enhanced sensitivity to water bodies obscured by vegetation. This contrast highlights the advantages of developing indices in conjunction with in situ water level measurements, rather than relying solely on expert annotation. This is particularly relevant in vegetated environments, where surface water can be partially or entirely obscured by canopy cover, making visual labeling less reliable. It is important to note, however, that the proposed index has thus far been formulated and validated primarily within rice paddy field environments. Future work is recommended to assess its applicability across other land cover types, including mangroves, shrubs, and non-rice agricultural systems.

Through rigorous benchmarking against conventional indices and global datasets, VAWI_log has demonstrated its effectiveness in detecting surface water under low to moderate vegetation cover. Its simplicity and compatibility with high-resolution, high-frequency open-source satellite data make it a practical tool for agricultural applications. In particular, the combined use of Landsat and Sentinel-2 imagery strengthens both the spatial and temporal monitoring of surface water, enabling improved irrigation mapping at local to regional scales. This capability supports efforts to assess the extent and timing of irrigation, helping water managers and policymakers identify areas experiencing water stress and redirect irrigation resources where and when they are most needed.

In addition, accurate surface water detection plays a critical role in estimating greenhouse gas (GHG) emissions, especially methane (CH4), from paddy fields. Process-based models for soil carbon and GHG fluxes rely heavily on water management regimes as a key input, as the duration of field flooding strongly influences methanogenesis [49,50]. The ability to monitor water regimes from space using optical satellite data and VAWI_log reduces reliance on costly and labor-intensive ground surveys. This facilitates more efficient and scalable estimation of GHG emissions under varying water management practices.