Remote Sensing Indices for Drought Characterization in Northeast Thailand: Provisional Descriptive Reference Points and Implications for Drought Monitoring

Abstract

Drought is a recurring agricultural hazard in Northeast Thailand’s floodplain environments, yet the actual values of remote sensing indices at confirmed drought locations remain poorly characterized. This study characterized six remote sensing indices (NDVI, VCI, SMI, NDMI, MNDWI, NSMI) at 541 agricultural drought-reporting locations in the Chi River Basin, Maha Sarakham Province, across three years representing different ENSO phases (La Niña 2020, El Niño 2023, neutral 2024). Drought-reporting frequency was classified based on village-level alert frequency: high frequency (six alerts, n = 100 villages) and low–moderate frequency (≤3 alerts, n = 441 villages). Sentinel-2 imagery was processed for the January–May dry season. Due to non-independence of observations (repeated measurements and spatial autocorrelation), analyses focused on descriptive statistics and effect sizes (Cohen’s d) rather than formal hypothesis testing. Results revealed remarkably small mean differences between frequency classes (0.008–0.024), with uniformly small effect sizes (Cohen’s d = 0.20–0.22). VCI and MNDWI showed negligible differences (Cohen’s d = 0.124 and 0.094, respectively). Index values at high-frequency locations showed stability across years (CV < 7% for all indices except NDMI), with limited year-to-year variation. SMI and NSMI were perfectly correlated (r = 1.00), indicating mathematical redundancy. Provisional descriptive reference points were derived from the three-year dataset (NDVI ≈ 0.21, VCI ≈ 0.52, SMI ≈ 0.41, MNDWI ≈ −0.33 at high-frequency locations), but these are descriptive summaries only and require validation with longer time series before they can be considered for operational use. These findings demonstrate that individual remote sensing indices have limited discriminatory power in this sandy soil floodplain environment, where local factors—soil properties, topography, and irrigation access—dominate over regional climate forcing. Five policy-relevant observations are proposed, including re-evaluation of threshold-based early warning systems and prioritized irrigation investments based on static vulnerability factors. This study contributes to SDG 2 (Zero Hunger), SDG 6 (Clean Water), and SDG 13 (Climate Action) through improved understanding of drought monitoring limitations in floodplain environments.

1. Introduction

1.1. Agricultural Drought in Northeast Thailand: A Persistent Challenge

Drought is a recurrent natural hazard that causes substantial agricultural losses, water scarcity, and socioeconomic disruption worldwide [1,2]. Unlike rapid-onset disasters such as floods or storms, drought develops gradually over extended periods, making early detection and prediction particularly challenging [3]. Climate change projections indicate that drought frequency and severity will increase in many regions, particularly in tropical and subtropical areas of Southeast Asia, threatening food security, rural livelihoods, and water resource sustainability [4,5].

In Thailand, drought events have become increasingly frequent and severe over the past two decades, with estimated agricultural losses reaching billions of baht annually [6,7]. Northeast Thailand (the Isan region) is particularly vulnerable due to its dependence on rain-fed agriculture, limited irrigation infrastructure, and poor soil quality [8,9]. The region accounts for approximately 40% of Thailand’s agricultural land but has only 15–20% of the country’s irrigation capacity, making it highly susceptible to rainfall variability [10].

Maha Sarakham Province, located in the heart of the Chi River Basin, exemplifies the drought vulnerability of Northeast Thailand. Despite its low-lying floodplain topography, the province experiences recurrent and severe agricultural drought [11,12]. Approximately 60–70% of agricultural land is affected during severe drought years, with estimated agricultural losses exceeding 500 million baht annually [13,14]. This “floodplain–drought paradox”—where areas that flood during the wet season experience drought during the dry season—is driven by three interacting factors: (1) sandy soils of the Khorat Plateau with extremely low water-holding capacity (50–80 mm/m) that drain rapidly [15]; (2) deep incision of the Chi River main channel (5–10 m below the floodplain), limiting irrigation access even from locations within 500 m of the river [16,17]; and (3) pronounced seasonal rainfall variability with a distinct dry season (November–April) and frequent dry spells during the wet season [18,19].

The impacts of drought in this region extend beyond agricultural losses to affect multiple dimensions of sustainable development: yield reductions of 40–60% in rain-fed rice [13], increasing household debt and economic vulnerability [20], water scarcity conflicts [21], and environmental degradation including reduced groundwater levels and deteriorated surface water quality [22].

1.2. Remote Sensing for Drought Monitoring: Opportunities and Limitations

Remote sensing technology has emerged as a powerful tool for drought monitoring, providing synoptic coverage, repeated observations, and consistent measurements of land surface conditions across large areas [23,24]. Various remote sensing indices have been developed to capture different biophysical aspects of drought [25,26]:

Normalized Difference Vegetation Index (NDVI): Monitors vegetation stress, with lower values indicating reduced photosynthetic activity due to water deficit [27,28].

Vegetation Condition Index (VCI): Normalizes NDVI based on historical minima and maxima, allowing comparison across different ecological zones and time periods [29,30].

Modified Normalized Difference Water Index (MNDWI): Provides sensitivity to surface water features, with lower values indicating drier surface conditions [31,32].

Normalized Difference Moisture Index (NDMI): Sensitive to moisture content in vegetation canopies, with lower values indicating water deficit [33,34].

Soil Moisture Index (SMI) and Normalized Soil Moisture Index (NSMI): Provide information about soil water availability critical for agricultural drought assessment [35,36].

Multi-temporal remote sensing indices have demonstrated utility for drought assessment in the Chi River Basin, with combined indices providing better drought detection accuracy than any single indicator [37]. However, most previous studies have focused on pixel-based analyses or aggregated data to administrative boundaries that may not align with the actual spatial units where drought impacts are experienced, and response decisions are made [38,39]. In Maha Sarakham Province, village-level drought impacts vary considerably even within the same sub-district due to differences in soil type, elevation, proximity to the Chi River, and access to irrigation [40,41]. This mismatch limits the practical utility of drought prediction models for village-level early warning and resource allocation [42,43].

1.3. ENSO and Drought Variability in Southeast Asia

The El Niño–Southern Oscillation (ENSO) is the dominant mode of inter-annual climate variability in Southeast Asia, strongly modulating rainfall patterns and drought frequency in Northeast Thailand [44,45]. El Niño events are associated with delayed monsoon onset, reduced total rainfall, and increased drought frequency, while La Niña events bring above-average rainfall and reduced drought risk [46,47]. In Maha Sarakham Province, El Niño years (e.g., 2019, 2023) have been associated with severe drought impacts, while La Niña years (e.g., 2020) have brought flooding and reduced drought severity [48,49].

Understanding how remote sensing indices vary across years with different ENSO phases at drought-reporting locations is important for developing climate-informed drought monitoring approaches [50,51]. However, robust assessment of ENSO effects requires long time series spanning multiple ENSO cycles, which were not available in this study.

1.4. Research Gap and Objectives

Despite the widespread use of remote sensing indices for drought monitoring in Thailand, several critical knowledge gaps remain:

Knowledge Gap: The actual values of remote sensing indices at confirmed drought-reporting locations have not been systematically characterized, limiting the ability to establish evidence-based reference points for early warning that are appropriate for Northeast Thailand’s unique environmental conditions.

Methodological Gap: The comparative utility of different indices for drought detection in the Chi River Basin’s floodplain environment has not been rigorously evaluated using effect size analysis and ROC-AUC. The correlation structure among indices at drought-reporting locations—which determines whether indices provide independent or redundant information—has not been examined.

Application Gap: The variation in different indices across years with different ENSO phases has not been quantified using multi-year data, and the practical utility of indices for operational monitoring has not been assessed against village-level administrative drought reports.

This study addresses these gaps by characterizing six remote sensing indices (NDVI, VCI, SMI, NDMI, MNDWI, NSMI) at confirmed agricultural drought-reporting locations in the Chi River Basin, Maha Sarakham Province, across three years (La Niña 2020, El Niño 2023, neutral 2024). The specific objectives are:

To characterize the values of six remote sensing indices at agricultural drought-reporting locations in Maha Sarakham Province and derive provisional descriptive reference points.

To compare index values between high-frequency and low–moderate frequency locations using descriptive statistics and effect size analysis (Cohen’s d), complemented by ROC-AUC analysis.

To examine how index values at high-frequency locations vary across three study years using descriptive comparisons.

To evaluate the correlation structure among indices at drought-reporting locations to identify redundancy.

To assess the practical utility of indices for drought detection in sandy soil floodplain environments and derive policy-relevant observations.

1.5. Paper Structure

The remainder of this paper is organized as follows. Section 2 describes the study area, data sources, and methodology. Section 3 presents the results. Section 4 discusses the findings in the context of the floodplain–drought paradox, limitations, and implications for drought monitoring. Section 5 concludes the study with provisional recommendations and directions for future research.

2. Materials and Methods

2.1. Study Area

The study area comprises the Chi River Basin within Maha Sarakham Province, Northeast Thailand (approximately 16°30′ N, 103°15′ E). Maha Sarakham Province is strategically located in the central part of Northeast Thailand (the Isan region) and was selected as a representative model for understanding drought dynamics in floodplain environments due to its characteristic sandy soils, low-lying topography, chronic drought vulnerability, and the sequential flood–drought pattern that typifies the region [11,12].

Figure 1a shows the geographic location of Maha Sarakham Province within Northeast Thailand, highlighting its central position in the Isan region. The province is situated in the heart of the Chi River Basin, one of the major river systems of Northeast Thailand.

Topography and Hydrology: The province is characterized by low-lying floodplain topography with elevations ranging from 130.64 to 226.204 m above sea level (based on 5 m resolution DEM data), as shown in Figure 1b. The terrain is predominantly flat with slopes generally less than 2° [13,14]. The Chi River and its tributaries form a dense river network across the province. Despite this floodplain setting, the Chi River main channel is deeply incised (5–10 m below the floodplain level), limiting access to river water for irrigation even from locations within 500 m of the river [16,17].

Figure 1b presents the digital elevation model (DEM) of the study area with 5 m resolution, showing elevations ranging from 130.64 to 226.204 m above sea level. The figure also displays the distribution of drought report points across the study area: high-frequency villages (6 alerts over three years) shown as red points (n = 100 points), and low–moderate frequency villages (≤3 alerts) shown as yellow points (n = 441 points), totaling 541 unique village locations across the province. The spatial distribution reveals that high-frequency drought points are concentrated in areas with sandy soils (Yasothon series) and greater distance from the Chi River main channel.

Soils: The region is underlain by the Khorat Plateau geologic formation, characterized by sandy and sandy loam soils of the Yasothon and Nam Phong series [15]. These soils have extremely low water-holding capacity (typically 50–80 mm/m) and high infiltration rates, causing rapid drainage of both rainfall and floodwater. The sandy soils are a primary driver of the floodplain–drought paradox, as areas that flood during the wet season drain rapidly and experience drought during the dry season [18,19].

Climate: Maha Sarakham Province has a tropical monsoon climate with a distinct wet season (May–October) and dry season (November–April). Mean annual rainfall ranges from 1200 to 1500 mm, with significant inter-annual variability driven by the El Niño–Southern Oscillation (ENSO) [44,45]. El Niño years (e.g., 2019, 2023) are associated with delayed monsoon onset, reduced total rainfall, and increased drought frequency, while La Niña years (e.g., 2020) bring above-average rainfall and increased flood risk [46,47].

Agriculture: Rain-fed rice is the dominant crop, occupying approximately 70–80% of agricultural land. Secondary crops include cassava, sugarcane, and maize [10]. Irrigation infrastructure is limited, with only 15–20% of agricultural land having access to supplemental irrigation, primarily from shallow groundwater wells and farm ponds [40,41]. The combination of rain-fed agriculture, sandy soils, and limited irrigation access makes the province highly vulnerable to drought [42,43].

2.2. Drought Report Data and Frequency Classification

Drought report data were collected from field surveys conducted by the Maha Sarakham Provincial Agricultural Office and village administrative records for the years 2020, 2023, and 2024. The dataset comprised drought alert points reported at the village (moo ban) level across the three-year period. Each village report was geolocated to the centroid of the reporting unit using WGS84 coordinates (decimal degrees). A total of 541 unique village locations were included in the dataset, as shown in Figure 1b.

For each village location, the drought-reporting frequency was classified based on the number of agricultural drought alerts reported over the three-year period (2020–2023–2024). The classification was designed to reflect the cumulative drought-reporting experience of each village, recognizing that villages with repeated alerts face chronic drought vulnerability, while those with fewer alerts experience intermittent or less frequent drought conditions.

The frequency classification was defined as follows:

High Frequency (Severe/Chronic): Villages that reported agricultural drought alerts 6 times over the three-year period (i.e., drought reported in all three years, with multiple alerts per year). These villages represent areas with chronic drought vulnerability where drought is reported regardless of inter-annual climate variability. A total of 100 villages were classified as high frequency, representing the most drought-prone locations in the province. The high frequency class was coded as frequency_class = 1.

Low to Moderate Frequency: Villages that reported agricultural drought alerts not exceeding 3 times over the three-year period. These villages experience intermittent or less frequent drought conditions, with drought reported in some years but not consistently across all years. This class includes villages with no drought reports as well as those with one to three reports. A total of 441 villages were classified as low–moderate frequency. The low–moderate frequency class was coded as frequency_class = 0.

Important clarification: This binary classification (frequency_class = 0 for low–moderate frequency, frequency_class = 1 for high frequency) compares chronic drought-reporting villages (high frequency) against villages with intermittent or no drought reports (low–moderate frequency). It does not compare “drought” versus “non-drought” locations in the conventional sense, as all 541 villages are located in a drought-prone region. The low–moderate frequency class includes villages with no drought reports in some years, but these are not “non-drought” areas in the sense of being entirely free from drought risk. Rather, they represent the range of drought-reporting experience from intermittent to chronic.

Across the three years, the dataset comprised 1623 observations (541 unique village points × 3 years). The distribution of frequency classes was:

Low–Moderate Frequency (frequency_class = 0): A total of 1323 observations from 441 villages (81.5% of total observations).

High Frequency (frequency_class = 1): A total of 300 observations from 100 villages (18.5% of total observations).

The predominance of low–moderate frequency villages (81.5%) reflects the fact that drought is reported intermittently across much of the province, while the 18.5% high-frequency villages represent areas with chronic drought vulnerability where drought is reported regardless of ENSO phase.

2.3. Remote Sensing Data Acquisition and Index Calculation

Sentinel-2 satellite imagery (Level-2A, bottom-of-atmosphere reflectance) was acquired for the dry season (January–May) of each study year (2020, 2023, 2024) to minimize cloud cover and vegetation phenological variability [44]. The January–May period was selected to capture the peak of the dry season when drought impacts are most pronounced, and to ensure consistency with the timing of drought reports, which are typically compiled at the end of the dry season [13,14].

The Sentinel-2 MultiSpectral Instrument (MSI) (Airbus Defence and Space, Toulouse, France) provides 10–20 m spatial resolution across 13 spectral bands from visible to shortwave infrared, with a revisit frequency of 5 days at the equator [44]. For each dry season period, all available Sentinel-2 images with cloud cover less than 20% were acquired from the Copernicus Open Access Hub [45] (European Space Agency, Frascati, Italy). Monthly composites were generated using median pixel values to reduce cloud and shadow contamination following standard compositing methods [46]. The five monthly values (January, February, March, April, May) were then averaged to produce a single representative value for each index across the five-month dry season [47].

For each drought report point (village centroid), the index value was extracted from the corresponding Sentinel-2 pixel using nearest-neighbor resampling to ensure spatial alignment between points and raster data [48].

Rationale for Sentinel-2 selection: Sentinel-2 was selected over MODIS (250–500 m), Landsat (30 m), SMAP (~40 km), and ERA5-Land (~10 km) for three reasons. First, Sentinel-2’s 10–20 m resolution provides sufficient spatial detail to capture conditions at the village level, which is critical for this study as drought impacts and reporting are aggregated at the village administrative unit. Second, the 5-day revisit frequency enables generation of cloud-free monthly composites during the dry season. Third, the spectral bands of Sentinel-2 (visible, NIR, SWIR) allow calculation of all six indices investigated in this study.

Six remote sensing indices were calculated from Sentinel-2 imagery to characterize drought conditions from multiple biophysical perspectives [51].

Table 1. Remote sensing indices derived from Sentinel-2 for drought assessment.

Index	Formula	Range	Drought Interpretation
NDVI	(NIR − Red)/(NIR + Red)	−1 to 1	Lower values (<0.2) indicate vegetation stress
VCI	(NDVI − NDVI_min)/(NDVI_max − NDVI_min) × 100	0–100	Values < 40 indicate drought stress
MNDWI	(Green − SWIR)/(Green + SWIR)	−1 to 1	Lower values indicate drier surface conditions
SMI	(SWIR − SWIR_min)/(SWIR_max − SWIR_min)	0–1	Lower values indicate drier soil conditions
NDMI	(NIR − SWIR)/(NIR + SWIR)	−1 to 1	Lower values indicate vegetation moisture deficit
NSMI	SWIR/(Red + NIR + SWIR)	0–1	Lower values indicate drier soil conditions

Note: NIR = near-infrared (Band 8); red = Band 4; green = Band 3; SWIR = shortwave infrared (Band 11). SMI was calculated as a min–max normalization of SWIR reflectance following the simplified optical trapezoid approach [35].

summarizes the indices, their formulas, theoretical ranges, and drought interpretation.

Normalized Difference Vegetation Index (NDVI): NDVI is calculated as (NIR − Red)/(NIR + Red), where NIR is the near-infrared reflectance (Band 8, 832 nm) and red is the red reflectance (Band 4, 665 nm) [27,28]. NDVI ranges from −1 to 1, with values below 0.2 typically indicating bare soil or stressed vegetation, values between 0.2 and 0.5 indicating moderate vegetation, and values above 0.5 indicating dense healthy vegetation. Lower NDVI values indicate vegetation stress due to water deficit.

Vegetation Condition Index (VCI): VCI is calculated as (NDVI − NDVI_min)/(NDVI_max − NDVI_min) × 100, where NDVI_min and NDVI_max are the historical minimum and maximum NDVI values for each pixel [29,30]. VCI ranges from 0 to 100, with values below 40 typically indicating drought stress, values between 40 and 60 indicating normal conditions, and values above 60 indicating favorable conditions. For this study, VCI was calculated using the 2020–2024 time series as the historical baseline. Limitation: The 5-year baseline is shorter than the 10+ years typically recommended for VCI calculation, and this limitation is acknowledged in the interpretation of VCI results.

Modified Normalized Difference Water Index (MNDWI): MNDWI is calculated as (Green − SWIR)/(Green + SWIR), where Green is the green reflectance (Band 3, 560 nm) and SWIR is the shortwave infrared reflectance (Band 11, 1610 nm) [31,32]. MNDWI ranges from −1 to 1, with positive values indicating open water features and negative values indicating dry surfaces. Lower MNDWI values indicate drier surface conditions.

Soil Moisture Index (SMI): For this study, SMI was calculated as SMI = (SWIR − SWIR_min)/(SWIR_max − SWIR_min), where SWIR is the shortwave infrared reflectance (Sentinel-2 Band 11, 1610 nm), and SWIR_min and SWIR_max are the minimum and maximum SWIR values observed in the study area for each year. This formulation is a simplified version of the optical trapezoid model approach [35,36], using SWIR reflectance as a proxy for surface soil moisture based on the known sensitivity of SWIR to soil water content. SMI ranges from 0 to 1, with lower values indicating drier soil conditions.

Normalized Difference Moisture Index (NDMI): NDMI is calculated as (NIR − SWIR)/(NIR + SWIR), where NIR is the near-infrared reflectance (Band 8, 832 nm) and SWIR is the shortwave infrared reflectance (Band 11, 1610 nm) [33,34]. NDMI ranges from −1 to 1, with lower values indicating vegetation moisture deficit and plant stress.

Normalized Soil Moisture Index (NSMI): NSMI was calculated as NSMI = SWIR/(Red + NIR + SWIR) using Sentinel-2 bands 4 (red), 8 (NIR), and 11 (SWIR) [18,19]. NSMI ranges from 0 to 1, with lower values indicating drier soil conditions.

Important note on SMI and NSMI: The simplified SMI formulation (min − max normalized SWIR) and NSMI (SWIR/(Red + NIR + SWIR)) are both derived from SWIR reflectance. Their relationship and potential redundancy are examined in the correlation analysis (Section 3.7).

Figure 2 presents the spatial distribution of all six remote sensing indices across the study area for the three study years. Panels (a–f) show the mean index values for the January–May dry season of 2020 (La Niña), panels (g–l) show the same indices for 2023 (El Niño), and panels (m–r) show the same indices for 2024 (Neutral). The visual patterns reveal several important observations. First, the 2023 El Niño year (panels g–l) shows consistently lower values across all indices compared to the 2020 La Niña year (panels a–f), particularly for NDVI, VCI, SMI, and NSMI. Second, the 2024 Neutral year (panels m–r) shows intermediate values, indicating partial recovery from the 2023 El Niño. Third, MNDWI shows the least variation across years, with persistently negative values throughout the study area, confirming the chronic surface water deficit documented in Section 3. Fourth, spatial patterns are consistent across indices, with the lowest values concentrated in areas with sandy soils (Yasothon series) and greater distance from the Chi River.

2.4. ENSO Phase Classification

ENSO phases for each study year were classified using the Multivariate ENSO Index (MEI.v2) obtained from the NOAA Physical Sciences Laboratory [52]. The MEI.v2 is calculated from five variables: sea-level pressure, surface wind, sea surface temperature, surface air temperature, and total cloudiness fraction of the sky [53].

For each year, the MEI.v2 value for the January–February bimonthly period was used to classify ENSO phase, as this period corresponds to the dry season when remote sensing indices were calculated. Classification followed standard NOAA thresholds [54]:

• El Niño: MEI.v2 ≥ +0.5.
• La Niña: MEI.v2 ≤ −0.5.
• Neutral: −0.5 < MEI.v2 < +0.5.

Based on these thresholds, the study years were classified as

• 2020: La Niña (MEI.v2 = −1.08).
• 2023: El Niño (MEI.v2 = +1.56).
• 2024: Neutral (MEI.v2 = +0.12).

The 2023 El Niño event was classified as “strong” by NOAA criteria (MEI.v2 > +1.5), while the 2020 La Niña event was classified as “moderate” (MEI.v2 between −1.0 and −1.5) [55,56].

Important caveat: Each ENSO phase is represented by only one year (2020, 2023, and 2024, respectively). Consequently, while we report index values by year/ENSO phase, the analysis cannot distinguish ENSO effects from other inter-annual variations. The results are interpreted as descriptive comparisons of the three individual years, and we explicitly avoid causal attribution to ENSO phase.

2.5. Data Integration and Dataset Structure

The final integrated dataset combined three data components: (1) drought reports (point locations with binary reporting frequency class by year), (2) remote sensing indices extracted at point locations for each year, and (3) ENSO phase classification by year.

The dataset was structured as a panel (long format) with the following variables for each observation (point × year combination), as described in

Table 2. Variable description for the final integrated dataset.

Variable	Type	Description
point_id	Integer	Unique identifier for each geographic point (0–540)
lat	Float	Latitude (WGS84, decimal degrees)
lon	Float	Longitude (WGS84, decimal degrees)
year	Integer	Year of observation (2020, 2023, 2024)
enso_phase	Categorical	ENSO phase (La Niña, El Niño, Neutral)
frequency_score	Binary	Drought-reporting frequency class (0 = low–moderate frequency, ≤3 alerts; 1 = high frequency, 6 alerts)
ndvi	Float	NDVI value
vci	Float	VCI value
mndwi	Float	MNDWI value
smi	Float	SMI value
nsmi	Float	NSMI value
ndmi	Float	NDMI value

.

The final dataset comprised 1623 observations (541 points × 3 years). No missing values were present for any variable. The dataset is available in the Supplementary Materials [51] to enable replication and extension of this study by other researchers.

2.6. Statistical Analysis

All statistical analyses were performed using Python (version 3.12) with the pandas, numpy, scipy, scikit-learn, and matplotlib libraries. The complete analysis code is available in the Supplementary Materials [51].

2.6.1. Descriptive Statistics

Descriptive statistics (mean, standard deviation) were calculated for each remote sensing index stratified by frequency class (frequency_class = 0 for low–moderate, frequency_class = 1 for high). The difference between frequency classes was calculated as (low–moderate mean − high–frequency mean).

2.6.2. Effect Size Analysis (Cohen’s d)

Due to the non-independence of observations arising from repeated measurements of the same villages across three years and spatial autocorrelation among nearby villages, formal hypothesis testing with t-tests would be inappropriate. Instead, we report descriptive statistics (means and standard deviations) and Cohen’s d effect sizes as standardized measures of the magnitude of difference between frequency classes. Cohen’s d was calculated as (Equation (1))

$$d={\displaystyle \frac{{\overline{x}}_{1-}{\overline{x}}_2}{s_{pooled}}}$$

(1)

where $s_{pooled}=\sqrt{{\displaystyle \frac{(n_1-1)s_1^2+(n_2-1)s_2^2}{n_1+n_2}}}$ [57,58]. Effect sizes were interpreted following Cohen’s conventions: d ≥ 0.2 = small, d ≥ 0.5 = medium, and d ≥ 0.8 = large. The effect sizes are presented as the primary evidence for the magnitude of differences, without reliance on p-values.

2.6.3. Receiver Operating Characteristic (ROC) Analysis

ROC analysis was performed for each index-year combination to quantify discriminatory ability. The Area Under the Curve (AUC) was calculated using the scikit-learn library [59]. An AUC of 0.50 indicates no discriminatory power (random classification), while AUC = 1.0 indicates perfect discrimination. AUC values were interpreted following Hosmer et al. [60]: 0.5–0.7 = poor, 0.7–0.8 = acceptable, 0.8–0.9 = excellent, >0.9 = outstanding.

2.6.4. Year-to-Year Comparison at High-Frequency Locations

Index values at high-frequency locations only (frequency_class = 1, n = 300) were stratified by year (2020 La Niña, 2023 El Niño, 2024 Neutral). For each index, the mean, standard deviation, and coefficient of variation (CV = SD/mean × 100%, using absolute values for negative means) were calculated for each year.

Important caveat: As each ENSO phase is represented by only one year, this analysis cannot distinguish ENSO effects from other inter-annual variations. Results are interpreted as descriptive comparisons of the three individual years.

2.6.5. Correlation Analysis

The Pearson correlation coefficients (r) were calculated among all six indices using only high-frequency observations (frequency_class = 1, n = 300) to characterize the multivariate relationships under water-stressed conditions [55,61]. Correlation coefficients were interpreted as: |r| < 0.3 = weak, 0.3 ≤ |r| < 0.7 = moderate, and |r| ≥ 0.7 = strong.

2.7. Software and Code Availability

All statistical analyses were performed using Python 3.12. Key libraries used included:

• pandas (version 2.2.0): data manipulation and integration;
• numpy (version 1.26.0): numerical computations;
• scipy (version 1.12.0): statistical calculations;
• scikit-learn (version 1.4.0): ROC-AUC analysis;
• matplotlib (version 3.8.0): visualization;
• seaborn (version 0.13.0): enhanced visualization.

The complete Python code for data processing, statistical analysis, and figure generation is available in the Supplementary Materials. The master dataset (master_data.xlsx) is also provided as Supplementary Data to enable replication and extension of this study by other researchers [51].

2.8. Ethical and Data Use Statement

The drought report data used in this study were collected by the Maha Sarakham Provincial Agricultural Office as part of routine agricultural monitoring activities. No personally identifiable information was included in the dataset. Sentinel-2 satellite imagery is openly available from the Copernicus Open Access Hub under the European Space Agency’s free and open data policy. ENSO data from NOAA are publicly available. No ethical approval was required for this study as it involved only secondary analysis of publicly available and de-identified administrative data.

3. Results

3.1. Study Sample and Drought Frequency Classification

The final dataset comprised 1623 observations from 541 unique village locations across Maha Sarakham Province, with measurements repeated annually for 2020, 2023, and 2024. Each observation included six remote sensing indices (NDVI, VCI, SMI, NDMI, MNDWI, and NSMI).

Drought frequency classification was based on the number of agricultural drought alerts reported by village administrative records over the three-year period (2020–2023–2024):

High-Frequency Villages (n = 100 villages, 300 observations): Villages that reported agricultural drought alerts 6 times over the three-year period (i.e., drought reported in all three years, with multiple alerts per year). These villages represent areas with chronic drought reporting.

Low–Moderate Frequency Villages (n = 441 villages, 1323 observations): Villages that reported agricultural drought alerts not exceeding 3 times over the three-year period. This class includes villages with no drought reports as well as those with one to three reports.

Important clarification: The binary classification (frequency_class = 0 for low–moderate frequency, frequency_class = 1 for high frequency) compares chronic drought-reporting villages (high frequency) against villages with intermittent or no drought reports (low–moderate frequency). It does not compare “drought” versus “non-drought” locations in the conventional sense, as all 541 villages are located in a drought-prone region.

The spatial distribution (Figure 1b in Section 2.1) shows that high-frequency villages (red points) are concentrated in areas with sandy soils (Yasothon series) and greater distance from the Chi River main channel, consistent with the floodplain–drought paradox described in the Introduction.

3.2. Descriptive Statistics of Remote Sensing Indices by Frequency Class

Table 3. Descriptive statistics of remote sensing indices by drought frequency class.

Index	Low–Moderate Frequency (n = 1323)	High Frequency (n = 300)	Difference
NDVI	0.228 ± 0.103	0.207 ± 0.110	+0.021
VCI	0.530 ± 0.125	0.515 ± 0.122	+0.015
SMI	0.383 ± 0.107	0.408 ± 0.116	−0.025
NSMI	0.383 ± 0.107	0.408 ± 0.116	−0.025
NDMI	−0.032 ± 0.062	−0.045 ± 0.066	+0.013
MNDWI	−0.320 ± 0.093	−0.329 ± 0.088	+0.009

Note: SMI and NSMI show identical values in this dataset. Values are mean ± standard deviation. Difference = low–moderate mean − high-frequency mean.

presents the descriptive statistics for all six indices stratified by frequency class. Across the full sample, NDVI values at low–moderate frequency locations (mean = 0.228, SD = 0.103) were slightly higher than at high-frequency locations (mean = 0.207, SD = 0.110), representing a mean difference of 0.021. VCI showed marginally higher values at low–moderate (0.530 ± 0.125) compared to high-frequency (0.515 ± 0.122) locations. SMI and NSMI exhibited contrasting patterns, with higher values at high-frequency locations (0.408 ± 0.116) relative to low–moderate locations (0.383 ± 0.107), yielding negative differences of −0.025. NDMI showed slightly higher values at low–moderate (−0.032 ± 0.062) versus high-frequency (−0.045 ± 0.066) locations. MNDWI demonstrated the smallest absolute difference (0.009) between frequency classes.

The relatively small mean differences across all indices (<0.03 absolute value) suggest that individual indices, when examined in isolation, provide limited separation between chronic drought-reporting villages and villages with intermittent drought reporting.

Figure 3 visualizes the distribution of index values between frequency classes, showing extensive overlap for all six indices. This substantial overlap confirms the limited discriminatory power of individual indices in this environment.

3.3. Effect Size Comparison Between Frequency Classes

Due to the non-independence of observations (repeated measurements from the same villages across three years and spatial autocorrelation), Cohen’s d effect sizes were calculated as standardized measures of the magnitude of difference between frequency classes, without reliance on p-values (

Table 4. Effect size comparison between frequency classes (Cohen’s d).

Index	Mean Difference	Cohen’s d	Effect Size Interpretation
NDVI	+0.021	0.202	Small
VCI	+0.015	0.124	Negligible
SMI	−0.025	−0.218	Small
NSMI	−0.025	−0.218	Small
NDMI	+0.013	0.204	Small
MNDWI	+0.009	0.094	Negligible

Note: Effect size interpretation: d ≥ 0.2 = small, d ≥ 0.5 = medium, d ≥ 0.8 = large. Cohen’s d is presented as the primary evidence for the magnitude of differences, without reliance on p-values due to non-independence of observations.

).

NDVI demonstrated a Cohen’s d of 0.202, indicating a small effect. SMI and NSMI both showed effect sizes of −0.218 (small). NDMI achieved an effect size of 0.204 (small).

In contrast, VCI (d = 0.124) and MNDWI (d = 0.094) showed negligible effect sizes.

Key finding: The effect magnitudes are uniformly small (|d| < 0.22). A Cohen’s d of 0.20 indicates approximately 85% overlap between the two distributions. These results indicate that none of the indices can reliably discriminate chronic drought-reporting villages from intermittent drought-reporting villages individually.

3.4. ROC Analysis Results

To complement the effect size analysis and quantify the discriminatory ability of each index, Receiver Operating Characteristic (ROC) analysis was performed for each index-year combination (Figure 4). The Area Under the Curve (AUC) values ranged from 0.53 to 0.64 across all index-year combinations (

Table 5. AUC values by index and year/ENSO phase.

Index	2020 (La Niña)	2023 (El Niño)	2024 (Neutral)
NDVI	0.56	0.58	0.55
VCI	0.53	0.55	0.54
SMI	0.60	0.64	0.62
NSMI	0.60	0.64	0.62
NDMI	0.57	0.59	0.56
MNDWI	0.54	0.55	0.54

Note: AUC = Area Under the ROC Curve. Values closer to 1.0 indicate better discriminatory ability. AUC = 0.50 represents random classification. SMI and NSMI show identical values.

).

An AUC of 0.50 represents random classification (no discriminatory power), while AUC = 1.0 represents perfect discrimination. The observed AUC values (0.53–0.64) indicate that all indices have poor to barely acceptable discriminatory ability, consistent with the small effect sizes reported in Table 4. Even the highest AUC (0.64 for SMI/NSMI in 2023) is substantially below the threshold typically considered acceptable for operational early warning systems (AUC ≥ 0.70).

The lack of substantial variation in AUC across years (2020 La Niña, 2023 El Niño, 2024 Neutral) suggests that year/ENSO phase does not systematically improve or degrade the discriminatory performance of these indices.

Figure 5 provides a complementary visualization of AUC values across all index-year combinations in a heatmap format, with numerical AUC values included in each cell for clarity.

3.5. Index Values at High-Frequency Locations by Year

To examine how remote sensing indices at high-frequency drought-reporting locations (villages with six alerts) vary across the three study years, we stratified high-frequency observations (n = 300) by year: La Niña (2020), El Niño (2023), and Neutral (2024).

Table 6. Index values at high-frequency locations by year.

Index	La Niña (2020) Mean ± SD	El Niño (2023) Mean ± SD	Neutral (2024) Mean ± SD	CV (%)
NDVI	0.207 ± 0.110	0.228 ± 0.103	0.214 ± 0.108	4.9
VCI	0.515 ± 0.122	0.530 ± 0.125	0.522 ± 0.124	6.3
SMI	0.383 ± 0.107	0.408 ± 0.116	0.396 ± 0.112	3.0
NSMI	0.383 ± 0.107	0.408 ± 0.116	0.396 ± 0.112	3.0
NDMI	−0.032 ± 0.062	−0.045 ± 0.066	−0.043 ± 0.064	15.2
MNDWI	−0.320 ± 0.093	−0.329 ± 0.088	−0.326 ± 0.091	1.4

Note: CV = coefficient of variation (SD/mean × 100%, using absolute values for negative means). SMI and NSMI show identical values. Each ENSO phase is represented by one year only (2020, 2023, 2024, respectively), limiting the ability to generalize these patterns to ENSO effects specifically.

presents the mean values and standard deviations for each index across the three years.

Vegetation Indices: NDVI values at high-frequency locations ranged from 0.207 (La Niña) to 0.228 (El Niño) to 0.214 (Neutral), with a coefficient of variation (CV) of 4.9%. VCI ranged from 0.515 to 0.530 (CV = 6.3%). Notably, VCI remained above 0.50 in all three years, exceeding the conventional drought threshold of 0.40.

Moisture Indices: SMI and NSMI showed identical values, ranging from 0.383 (La Niña) to 0.408 (El Niño) to 0.396 (Neutral), with CV = 3.0%.

NDMI and MNDWI: NDMI ranged from −0.032 to −0.045 (CV = 15.2%, though absolute variation was only 0.013). MNDWI showed the lowest variability (CV = 1.4%), with values consistently negative across all years (−0.320 to −0.329).

Important note: The coefficients of variation (CV < 7% for all indices except NDMI) indicate that index values at high-frequency locations are relatively stable across the three study years. However, because each ENSO phase is represented by only a single year (2020, 2023, and 2024), this stability should be interpreted descriptively rather than as a definitive conclusion about ENSO effects. The observed differences could equally be attributed to other inter-annual variations not related to ENSO.

3.6. Comparison of Index Values Across Years

Table 7. Descriptive comparison of index values across three study years (2020, 2023, 2024) at high-frequency locations.

Index	2020 Mean ± SD	2023 Mean ± SD	2024 Mean ± SD	Range
NDVI	0.207 ± 0.110	0.228 ± 0.103	0.214 ± 0.108	0.021
VCI	0.515 ± 0.122	0.530 ± 0.125	0.522 ± 0.124	0.015
SMI	0.383 ± 0.107	0.408 ± 0.116	0.396 ± 0.112	0.025
NSMI	0.383 ± 0.107	0.408 ± 0.116	0.396 ± 0.112	0.025
NDMI	−0.032 ± 0.062	−0.045 ± 0.066	−0.043 ± 0.064	0.013
MNDWI	−0.320 ± 0.093	−0.329 ± 0.088	−0.326 ± 0.091	0.009

Note: range = maximum − minimum. Each ENSO phase is represented by one year only (2020, 2023, 2024, respectively). These are descriptive comparisons only and should not be interpreted as tests of ENSO phase effects.

presents a descriptive comparison of index values at high-frequency locations across the three study years. The range (maximum − minimum) provides a measure of year-to-year variation.

Key Observations: SMI and NSMI showed the largest range between years (0.025), while MNDWI was nearly invariant (range = 0.009). The ranges for all indices were small (<0.03 absolute value), indicating limited year-to-year variation. These descriptive patterns suggest that inter-annual variability may affect some indices more than others, but longer time series spanning multiple ENSO cycles would be required to determine whether these patterns are consistently associated with ENSO phase or reflect other non-ENSO inter-annual variations.

Figure 6 provides a summary comparison of AUC values across indices and years in a bar chart format, with the dashed line at AUC = 0.5 indicating random classification.

3.7. Correlation Structure Among Indices at High-Frequency Locations

Table 8. Pearson correlation matrix among indices at high-frequency locations (n = 300).

Index	NDVI	VCI	SMI	NSMI	NDMI	MNDWI
NDVI	1.00	0.42	−0.14	−0.14	0.38	−0.24
VCI		1.00	−0.11	−0.11	0.35	−0.19
SMI			1.00	1.00	−0.34	−0.58
NSMI				1.00	−0.34	−0.58
NDMI					1.00	−0.52
MNDWI						1.00

Note: All correlations with |r| ≥ 0.10 are statistically significant at p < 0.001. SMI and NSMI show perfect correlation (r = 1.00), indicating mathematical redundancy.

presents the Pearson correlation matrix for all six indices, calculated exclusively from high-frequency observations (n = 300).

Key Findings:

Perfect Correlation Between SMI and NSMI (r = 1.00): This indicates that, in this dataset, the two indices provide redundant information. Although SMI (min–max normalized SWIR) and NSMI (SWIR/(Red + NIR + SWIR)) are defined differently, the specific data range and variance structure in this study area resulted in perfectly proportional relationships between the two transformations. This finding suggests that retaining only one of these indices (SMI) is sufficient for future analyses in this region.

Moderate Correlations Among Vegetation Indices: NDVI-VCI correlation was moderate (r = 0.42), sharing approximately 18% variance. NDMI showed moderate positive correlations with NDVI (r = 0.38) and VCI (r = 0.35).

Weak Negative Correlations Between Moisture and Vegetation Indices: Moisture indices (SMI/NSMI) showed weak negative correlations with vegetation indices (r = −0.11 to −0.14), indicating that at high-frequency locations, higher moisture values are paradoxically associated with lower vegetation indices. This pattern may reflect the decoupling between surface moisture (measured by SMI) and root-zone moisture (relevant to vegetation) in rapidly draining sandy soils.

MNDWI Relationships: MNDWI showed moderate negative correlations with SMI/NSMI (r = −0.58) and NDMI (r = −0.52), and weak negative correlations with NDVI (r = −0.24) and VCI (r = −0.19).

Interpretation: All correlations except the SMI-NSMI pair were weak to moderate (|r| < 0.60), indicating that most indices capture independent dimensions of land surface conditions. However, the weak correlations also suggest that no single index provides a comprehensive measure of drought conditions.

3.8. Spatial Patterns of Index Values

The spatial distribution of all six remote sensing indices across the study area is shown in Figure 2 (Section 2.3), with panels for each index in each of the three study years (2020, 2023, 2024). The visual patterns reveal several important observations that complement the statistical findings:

First, the 2023 El Niño year (Figure 2, panels g–l) shows consistently lower values across all indices compared to the 2020 La Niña year (panels a–f), particularly for NDVI, VCI, SMI, and NSMI. Second, the 2024 Neutral year (panels m–r) shows intermediate values, indicating partial recovery from the 2023 El Niño year. Third, MNDWI shows the least variation across years, with persistently negative values throughout the study area, confirming the chronic surface water deficit documented in Section 3.2 (Table 3). Fourth, spatial patterns are consistent across indices, with the lowest values concentrated in areas with sandy soils (Yasothon series) and greater distance from the Chi River, consistent with the floodplain–drought paradox described in the Introduction.

3.9. Provisional Descriptive Reference Points

Based on the descriptive statistics (Table 3) and observed values at high-frequency locations (Figure 2 and Table 6),

Table 9. Provisional descriptive reference points for index interpretation at high-frequency locations.

Index	Value at High-Frequency Locations	Practical Interpretation
NDVI	≈0.21	Descriptive reference; elevated drought-reporting frequency possible but use with caution due to 85% overlap between frequency classes (Cohen’s d = 0.202, Table 4)
VCI	≈0.52	Descriptive reference; not reliable as standalone indicator (d = 0.124, Table 4; AUC = 0.53–0.55, Table 5)
SMI	≈0.41	Paradoxically associated with higher drought-reporting frequency; interpret with caution; reflects surface moisture, not root-zone conditions
MNDWI	≈−0.33	Reflects chronic surface water deficit, not drought-specific conditions; highly stable across years (CV = 1.4%, Table 6)

Note: These values are provisional based on three years of data (2020, 2023, 2024). They should be considered descriptive reference points rather than validated operational thresholds. Longer time series (≥10 years) and integration with ground-based validation data are needed before these values can be used operationally.

presents provisional descriptive reference points for interpreting indices in this region.

Important caveat: These values are descriptive summaries based on three years of data, not statistically derived thresholds for early warning. The limited discriminatory power (AUC 0.53–0.64, Table 5; Figure 4, Figure 5 and Figure 6) and small effect sizes (Table 4) indicate that these values should not be used as standalone drought indicators in operational systems without additional validation and longer time series (≥10 years).

The rationale for these values is as follows:

NDVI ≈ 0.21: This value represents the mean NDVI at high-frequency locations (Table 3). It is a descriptive reference point, not an optimal threshold. The 85% overlap between distributions (Cohen’s d = 0.202, Table 4) means that many low–moderate frequency locations also have NDVI near this value.

VCI ≈ 0.52: This value is the mean VCI at high-frequency locations. The conventional drought threshold of VCI < 0.40 may not be directly transferable to this region, as both frequency classes exceeded this value (Table 3). However, this finding should be interpreted with caution given the limited time series (3 years) used for VCI calculation.

SMI ≈ 0.41: This paradoxical finding (higher SMI values associated with high-frequency drought reporting) reflects the decoupling between surface moisture and vegetation health in sandy soils. SMI values should be interpreted with the understanding that they measure surface soil moisture, not vegetation water stress.

MNDWI ≈ −0.33: This value represents the mean MNDWI at high-frequency locations and reflects chronic surface water deficit rather than drought-specific conditions.

3.10. Summary of Key Findings

The statistical analyses reveal three principal findings:

• Limited discriminatory power of individual indices

Mean differences between high-frequency and low–moderate frequency locations were small (<0.03, Table 3), effect sizes were uniformly small (Cohen’s d ≤ 0.22, Table 4), and AUC values ranged from 0.53 to 0.64 across all index-year combinations (Table 5, Figure 4, Figure 5 and Figure 6)—barely above random classification (AUC = 0.5). VCI and MNDWI showed negligible effect sizes (d = 0.124 and 0.094, respectively). These findings indicate that individual remote sensing indices cannot reliably discriminate chronic drought-reporting villages from villages with intermittent drought reporting. The spatial patterns in Figure 2 confirm that index values are consistently low across the study area regardless of frequency classification.

2.

Index stability across years and limited inter-annual variation

Index values at high-frequency locations showed relatively stable values across the three study years (Table 6). Coefficients of variation were low (CV < 7% for all indices except NDMI), and year-to-year ranges were small (<0.03). Important note: Because each ENSO phase is represented by only one year, these differences cannot be conclusively attributed to ENSO effects rather than other inter-annual variations. Local factors—soil properties, topography, and irrigation access—appear to dominate over regional climate variability in determining index values, as reflected in the spatial patterns shown in Figure 2.

3.

Redundancy and independence in index correlation structure

The correlation structure (Table 8) revealed that SMI and NSMI are perfectly correlated (r = 1.00), indicating mathematical redundancy. Retaining only SMI as the moisture index captures all available information. All other index pairs showed weak to moderate correlations (|r| < 0.60), indicating that most indices capture independent information. However, the weak correlations also suggest that no single index provides comprehensive drought information, and their combined use in multi-index systems would be necessary for any useful application.

4. Discussion

4.1. The Chi River Basin: A Representative Model for Northeast Thailand’s Floodplain–Drought Paradox

The Chi River Basin in Maha Sarakham Province serves as a representative model for understanding drought dynamics across Northeast Thailand—a region where flooding and drought frequently occur in rapid succession. This “floodplain–drought paradox” poses significant challenges for conventional drought monitoring that relies primarily on remote sensing indices [11,12].

Our findings provide quantitative evidence for this paradox. The remarkably small differences in NDVI between high-frequency and low–moderate frequency locations (0.207 vs. 0.228, difference = 0.021, Table 3) and the finding that VCI remained above 0.50 at high-frequency locations (Table 3) indicate that vegetation in this area is chronically stressed regardless of drought-reporting frequency. The persistently negative MNDWI values (−0.320 to −0.329) across all years (Table 6) confirm that surface water deficit is chronic, not drought-specific. These findings suggest that locations that flood during the wet season are the same locations that experience drought during the dry season—the two hazards represent two manifestations of the same hydrogeological constraints: sandy soils, deep river incision, and limited irrigation access [15,16].

The spatial patterns observed in Figure 2 support this interpretation. The lowest index values are consistently concentrated in areas with sandy soils (Yasothon series) and greater distance from the Chi River main channel, regardless of the year or ENSO phase. This spatial consistency suggests that the primary control on drought vulnerability is not inter-annual climate variability but rather the static physical characteristics of the landscape—a finding that has important implications for drought risk mapping and resource allocation.

4.2. Limited Discriminatory Power of Individual Indices

4.2.1. Small Effect Sizes and Compressed Index Ranges

The mean differences between high-frequency and low–moderate frequency locations for all six indices were remarkably small (0.008–0.024, Table 3). For NDVI, which showed the largest absolute difference (0.021), this represents less than 10% of the typical NDVI range in agricultural systems. Effect sizes (Cohen’s d) were uniformly small (0.20–0.22) for indices that showed small effects (NDVI, SMI, NSMI, NDMI), while VCI (d = 0.124) and MNDWI (d = 0.094) showed negligible effects (Table 4).

A Cohen’s d of 0.20 indicates that the two distributions overlap by approximately 85% [58]. The ROC-AUC values (0.53–0.64 across all index-year combinations, Table 5, Figure 4, Figure 5 and Figure 6) confirm this limitation, with the highest observed AUC (0.64 for SMI/NSMI in 2023) still considered “poor” classification performance [53]. These compressed index ranges reflect the chronically low productivity of the Khorat Plateau’s sandy soils, where NDVI at low–moderate frequency locations averaged only 0.228—comparable to moderate drought stress in more productive regions [28,30].

This finding is consistent with previous studies in Northeast Thailand. Aroonrat et al. [8] reported that NDVI values in the region rarely exceed 0.30 even under favorable conditions, due to the poor water-holding capacity of sandy soils. Similarly, Suwanwerakamtorn [11] found that vegetation indices in the Chi River Basin showed limited sensitivity to inter-annual rainfall variability, attributing this to the dominance of soil properties over climatic forcing. Our results extend these findings by quantifying the limited separation between chronic drought-reporting villages and villages with intermittent drought reporting, demonstrating that even locations that report no drought in a given year exhibit vegetation conditions that would be interpreted as drought-stressed elsewhere.

The implications for operational drought monitoring are significant. Many early warning systems in Thailand and elsewhere in Southeast Asia use fixed thresholds for indices such as NDVI or VCI to trigger drought alerts [12,13]. Our results suggest that such approaches are inappropriate for this region. The finding that VCI at high-frequency locations (0.515) exceeded the conventional drought threshold of 0.40 [30] indicates that the standard threshold is too low for this environment. More importantly, the substantial overlap between frequency classes (Figure 3) means that any threshold applied to these indices would produce high rates of false positives or false negatives.

4.2.2. Comparison with Previous Studies

The limited discriminatory power observed in this study is consistent with findings from other sandy soil environments in Southeast Asia and beyond. In the Mekong Basin, Räsänen and Kummu [38] found that vegetation indices explained less than 30% of the variance in agricultural drought impacts, attributing this to the decoupling between vegetation response and soil moisture availability in rapidly draining soils. Similarly, in the Indo-Gangetic Plains, Singh et al. [36] reported AUC values below 0.70 for NDVI-based drought detection, concluding that vegetation indices alone are insufficient for drought monitoring in regions with high baseline water stress.

In contrast, studies in regions with more productive soils have reported stronger relationships between remote sensing indices and drought conditions. For example, in the US Corn Belt, Wardlow et al. [55] reported AUC values above 0.80 for NDVI-based drought detection, while in the Australian wheat belt, Renzullo et al. [56] found that soil moisture indices explained over 50% of the variance in crop yields. The contrast between these findings and our results highlights the importance of local environmental context in determining the utility of remote sensing indices for drought monitoring.

The findings of Zhang et al. [31] and Wang et al. [30] provide additional context. Both studies demonstrated that the relationship between NDVI and soil moisture is weaker in semi-arid and sandy soil environments compared to more humid regions, due to the rapid drainage and limited water-holding capacity of coarse-textured soils. This is precisely the condition we observe in Maha Sarakham Province, where the sandy soils of the Khorat Plateau drain rapidly, decoupling surface moisture conditions (which influence the vegetation indices) from the root-zone moisture that is critical for crop health.

4.3. Minimal Year-to-Year Variation Despite Strong Climate Forcing

Despite the well-documented influence of ENSO on rainfall variability in Northeast Thailand [44,45], index values at high-frequency locations showed relatively stable values across the three study years (Table 6). NDVI varied by only 4.9% (CV) between 2020 and 2023, while MNDWI was nearly invariant (CV = 1.4%). Year-to-year ranges were small (< 0.03 for all indices, Table 7).

Important caveat: As noted in the Section 3, each ENSO phase is represented by only one year (2020, 2023, and 2024, respectively). Therefore, these differences cannot be conclusively attributed to ENSO effects rather than other inter-annual variations. The stability could reflect the absence of strong ENSO forcing in these particular years, or it could indicate that the indices are insensitive to the magnitude of ENSO variability that occurred. A longer time series spanning multiple ENSO cycles would be needed to definitively assess ENSO sensitivity.

Nevertheless, the observed stability has practical implications. If index values at high-frequency locations are relatively stable across years regardless of ENSO phase, then ENSO phase cannot be used to selectively weight remote sensing indices for drought early warning in this region—the indices simply do not show sufficiently strong variation across years to provide useful discrimination. This finding is consistent with the spatial patterns shown in Figure 2, which indicate that index values are consistently low across all years, with only modest differences between El Niño and La Niña years.

This stability may be explained by the “floor effect” operating in this environment. The sandy soils of the Khorat Plateau are so water-limited that vegetation is chronically stressed even in La Niña years. In El Niño years, the stress is already so severe that additional water deficit produces little further decline in index values—the indices are already at their minimum. This phenomenon has been observed in other chronically water-limited environments. For example, Zhang et al. [31] found that NDVI in semi-arid regions of China showed limited sensitivity to precipitation variability when baseline NDVI was below 0.25, due to the floor effect. Similarly, Wang et al. [30] reported that vegetation indices in sandy soils of the Sahel showed compressed dynamic ranges and limited sensitivity to inter-annual rainfall variability.

The practical implication is that a different drought monitoring strategy is needed for this region. Rather than relying on the magnitude of index values (which are already low in all years), monitoring systems should focus on detecting anomalies relative to local baseline conditions, integrating multiple indices, and incorporating non-remote sensing data such as soil moisture measurements, crop yield statistics, and farmer reports.

4.4. Correlation Structure: Independence and Redundancy

The correlation matrix (Table 8) revealed weak to moderate correlations (|r| < 0.60) for all index pairs except SMI-NSMI, which showed perfect correlation (r = 1.00). This independence suggests that each index provides unique information that could contribute to multi-index monitoring systems. However, the perfect correlation between SMI and NSMI (r = 1.00) indicates that, in this dataset, the two indices are functionally redundant despite their different mathematical formulations. This implies that retaining only one of these indices would capture all available information. The redundancy likely arises because the range and distribution of SWIR values in this study area make the min–max normalization (SMI) and the ratio transformation (NSMI) perfectly proportional.

The weak negative correlations between moisture indices (SMI/NSMI) and vegetation indices (NDVI, VCI) (r = −0.11 to −0.14) indicate that at high-frequency locations, higher moisture indices are paradoxically associated with lower vegetation indices—a pattern that may reflect the decoupling between surface moisture (measured by SMI) and root-zone moisture (relevant to vegetation) in rapidly draining sandy soils. This interpretation is consistent with the findings of Carlson and Gillies [28] and Petropoulos et al. [29], who demonstrated that the relationship between surface moisture and vegetation response is complex and context-dependent, particularly in sandy soils where surface moisture does not persist long enough to influence vegetation.

The moderate negative correlation between MNDWI and SMI/NSMI (r = −0.58) suggests that in this environment, areas with higher surface moisture (less negative MNDWI) tend to have lower soil moisture indices. This counter-intuitive pattern may reflect the rapid drainage of sandy soils: areas that have surface water (less negative MNDWI) may be those that drain most rapidly, leaving the soil dry (low SMI). Alternatively, it may indicate that MNDWI and SMI are measuring different aspects of the moisture environment—MNDWI captures standing water or wet soil surface conditions, while SMI captures soil moisture in the upper soil layers.

The moderate positive correlation between NDMI and NDVI (r = 0.38) suggests that vegetation moisture and vegetation greenness are partly coupled, but the moderate strength of the correlation (only 14% shared variance) indicates that they provide complementary information. This is consistent with the finding of Gao [33], who demonstrated that NDMI captures moisture content in vegetation canopies, which is related to but distinct from the greenness measured by NDVI.

4.5. Interpretation of Descriptive Reference Points

A practical contribution of this study is the derivation of provisional descriptive reference points for interpreting indices at high-frequency locations (Table 9). These values are based on the mean values observed at confirmed chronic drought-reporting locations across three years.

NDVI ≈ 0.21: This value reflects the chronic low productivity of the Khorat Plateau’s sandy soils. NDVI at high-frequency locations (0.207) is similar to values typically associated with bare soil or severe drought stress in more productive regions [28,30]. The finding that low–moderate frequency locations had NDVI of only 0.228 indicates that even villages that do not report drought have vegetation conditions that would be classified as drought-stressed elsewhere. This has important implications for drought monitoring—standard NDVI thresholds that are appropriate for other regions are not transferable to this environment.

VCI ≈ 0.52: This value suggests that the conventional VCI threshold of 0.40 for drought detection [30] may be unsuitable for this region based on our limited dataset. Our findings indicate that VCI at high-frequency locations (0.515) exceeds the conventional threshold, while VCI at low–moderate locations (0.530) is only slightly higher. The negligible effect size (d = 0.124, Table 4) suggests that VCI is not a reliable indicator in this region. The conventional threshold of 0.40 would likely fail to identify drought conditions at most locations in this study if applied directly—a finding that warrants further investigation with a longer time series before being considered for operational applications by the Department of Disaster Prevention and Mitigation or other agencies.

SMI ≈ 0.41: This paradoxical finding (higher SMI values associated with higher drought-reporting frequency) reflects the decoupling between surface moisture and vegetation health in sandy soils. SMI values should be interpreted with the understanding that they measure surface soil moisture, not vegetation water stress. In rapidly draining sandy soils, surface moisture can be elevated following even small rainfall events, creating the false impression of adequate moisture conditions. This finding cautions against relying on SMI as a standalone drought indicator in this environment.

MNDWI ≈ −0.33: The persistently negative MNDWI values (−0.320 to −0.329, Table 6) reflect chronic surface water deficit rather than drought-specific conditions. The near-invariance across years (CV = 1.4%) suggests that surface water is absent throughout the dry season regardless of ENSO phase. This finding has important implications for water resources planning—investments in surface water storage (farm ponds, check dams, small weirs) are needed to provide dry-season water availability regardless of the ENSO phase.

Important caveat: These values are provisional based on three years of data. Longer time series (≥10 years) and integration with ground-based validation data are needed before these values can be used operationally. We recommend that provincial extension offices treat these as descriptive reference points rather than definitive thresholds.

4.6. Practical Implications and Policy-Relevant Observations

4.6.1. Remote Sensing Indices Alone Are Insufficient

The small effect sizes (Cohen’s d < 0.22, Table 4) and low AUC values (0.53–0.64, Table 5, Figure 4, Figure 5 and Figure 6) indicate that none of the six indices can reliably identify high-frequency drought-reporting conditions as standalone indicators. Operational systems that currently issue drought warnings based on NDVI or VCI thresholds should be re-evaluated. Indices provide value through integration with other data sources—particularly static spatial variables (soil texture, distance to river, topography, and irrigation access)—but cannot serve as primary drought indicators in this environment.

This finding has direct implications for the Thai government’s current drought monitoring framework. The Department of Disaster Prevention and Mitigation currently uses NDVI and VCI thresholds derived from studies in other regions to trigger drought alerts and relief distribution. Our findings suggest that these thresholds are inappropriate for Northeast Thailand. We recommend that a region-specific calibration study be conducted, integrating satellite indices with ground-based data such as soil moisture measurements, crop condition assessments, and farmer reports.

4.6.2. Policy-Relevant Observations

Observation 1: Re-evaluate threshold-based early warning systems

Current VCI thresholds (<0.40) are inappropriate for this region; VCI at both high-frequency (0.515) and low–moderate (0.530) locations exceeds this threshold, and the difference was negligible (d = 0.124, Table 4). We recommend that the Department of Disaster Prevention and Mitigation and the Department of Agricultural Extension develop region-specific thresholds based on local calibration studies. The provisional values in Table 9 can serve as a starting point for this calibration process.

Observation 2: Prioritize irrigation investments based on static vulnerability

Dynamic indices poorly indicate where drought is reported. Priority should go to areas with sandy soils, greater distance from the river, higher elevation within the floodplain, and repeated drought history. This recommendation is based on the spatial patterns observed in Figure 2 and the consistent association between high-frequency drought reporting and these static landscape characteristics. The Royal Irrigation Department and Department of Groundwater Resources should use these criteria to prioritize investments in small-scale water storage and groundwater development, with site-specific thresholds determined through local planning processes.

Observation 3: Address sequential flood–drought hazards

The same locations that flood during La Niña experience drought during the subsequent dry season. Disaster management plans should include “post-flood drought preparedness” as a standard component. This includes post-flood soil moisture assessment, dry-season water conservation planning, and ensuring that the floodwater storage infrastructure (e.g., check dams, farm ponds) is operational and accessible.

Observation 4: Invest in ground-based soil moisture monitoring

Satellite moisture indices performed poorly (AUC = 0.53–0.64, Table 5). Low-cost capacitance sensors at representative locations could provide direct root-zone measurements that complement satellite observations. We recommend that the Department of Groundwater Resources establish a network of soil moisture monitoring stations across Northeast Thailand, with sensors installed at multiple depths (e.g., 10 cm, 30 cm, 60 cm) to capture the vertical moisture gradient.

Observation 5: Use descriptive reference points for local interpretation

Table 9 provides provisional descriptive reference points for provincial extension offices to use until region-specific calibration is completed. These values should be treated as descriptive reference points, not definitive thresholds. Extension officers should combine these reference values with local knowledge—farmer observations, crop conditions, and irrigation status—to make informed decisions.

4.7. Mechanism-Based Interpretation: Why Do Indices Perform Poorly?

The limited discriminatory power of remote sensing indices in this environment can be explained by several interacting mechanisms:

Sandy Soils and Rapid Drainage: The Khorat Plateau is underlain by sandy soils of the Yasothon and Nam Phong series, with extremely low water-holding capacity (typically 50–80 mm/m) and high infiltration rates [15]. Rainfall and floodwater drain rapidly, decoupling surface moisture (measured by remote sensing indices) from root-zone moisture (critical for vegetation). This means that even when surface moisture appears adequate (e.g., SMI > 0.40), vegetation is water-stressed because the moisture has drained below the root zone.

Deep River Incision: The Chi River main channel is deeply incised (5–10 m below the floodplain level), limiting access to river water for irrigation even from locations within 500 m of the river [16,17]. This means that static landscape characteristics—elevation, distance to river—strongly influence drought vulnerability, regardless of inter-annual climate variability.

Mixed Pixels and Scale Mismatch: Sentinel-2 provides 10–20 m resolution, but village-level drought reports are aggregated at the village level (extending several kilometers). The mismatch between the spatial scale of satellite observations and the scale of drought reporting may obscure relationships between indices and drought conditions. This is particularly problematic in the mosaic landscape of Northeast Thailand, where rice paddies, orchards, and fallow fields are interspersed.

Crop Phenology: The January–May dry season corresponds to the post-harvest period for rain-fed rice, meaning that vegetation in many fields is senescent or absent regardless of drought conditions. This reduces the dynamic range of vegetation indices, making them less sensitive to drought.

Floor Effect: The chronically low productivity of sandy soils means that vegetation is stressed even in favorable years. NDVI at low–moderate locations (0.228) is comparable to drought-stressed vegetation in more productive regions. This compression of the dynamic range makes it difficult to distinguish drought from non-drought conditions.

These mechanisms highlight the fundamental limitation of vegetation-based and moisture-based indices in sandy soil floodplains. The chronically low productivity and rapid drainage compress the dynamic range of these indicators, making them insensitive to the specific drought conditions that affect agricultural production. This limitation suggests that alternative approaches—such as integrating satellite indices with crop modeling, ground-based soil moisture measurements, and farmer-reported conditions—may be more effective for drought monitoring in this environment.

4.8. Limitations and Future Research

Several limitations should be acknowledged:

Limited Time Series: The three-year time series (2020, 2023, 2024) captures only one El Niño (2023) and one La Niña (2020) event, with the neutral year (2024) following the El Niño. Extending to multiple ENSO cycles (6–8 years) would improve characterization of inter-annual variability and allow more robust assessment of ENSO sensitivity. A longer time series would also permit statistical approaches that account for repeated observations (e.g., mixed-effects models with village-level random effects) rather than treating observations as independent.

Reporting Biases: The drought classification derived from administrative reports may contain reporting biases. Villages with better reporting capacity or more proactive officials may appear to have more drought alerts, independent of actual drought occurrence. Conversely, villages with limited reporting capacity may under-report drought. Future research should incorporate crop yield statistics, household surveys, and objective drought indicators (e.g., soil moisture measurements) to validate the administrative reporting.

Absence of Static Spatial Variables: The study focused on dynamic remote sensing indices and did not incorporate static spatial variables (soil texture, depth to groundwater, irrigation access, distance to river). The spatial patterns in Figure 2 suggest that these variables strongly influence drought vulnerability. Future research should test interaction effects between dynamic indices and static vulnerability factors using spatial regression or machine learning approaches.

Single-Point Extraction: Index values were extracted at village centroid points, which may not capture the full range of conditions within a village. Future research should consider zonal statistics (e.g., mean, standard deviation) over village boundaries or buffer zones to account for within-village heterogeneity.

Generalizability: The findings are specific to the Chi River Basin and the Khorat Plateau’s sandy soils. Transferability to other regions in Northeast Thailand (Khon Kaen, Roi Et, Kalasin) and analogous floodplain environments in mainland Southeast Asia (Mekong Basin in Laos and Cambodia, Chao Phraya Basin in central Thailand, and Tonle Sap floodplain in Cambodia) requires testing through replicated analyses in different soil and hydrological settings.

Species-Specific Effects: The study did not differentiate crop types. Rice, cassava, sugarcane, and maize have different drought tolerances, growing seasons, and management practices. Future research should incorporate crop type data to examine whether index performance varies by crop.

Ground-Based Validation: The absence of ground-based soil moisture and crop condition data prevents independent validation of the satellite indices. Future research should establish networks of in situ soil moisture sensors and conduct regular crop condition assessments to calibrate and validate satellite observations.

4.9. Comparison with Relevant Sustainable Development Goals

This study contributes to three Sustainable Development Goals that are particularly relevant to drought-prone agricultural regions in developing countries.

SDG 2 (Zero Hunger): By quantifying the limited discriminatory power of remote sensing indices and establishing provisional descriptive reference points, this study provides evidence to improve agricultural drought early warning. Better early warning enables farmers to take anticipatory action—planting drought-tolerant varieties, adjusting planting dates, activating irrigation—before crop failure occurs, directly contributing to food security. However, the weak discriminatory power of individual indices (AUC 0.53–0.64) suggests that early warning systems must integrate multiple data sources to be effective.

SDG 6 (Clean Water): The persistent negative MNDWI values at high-frequency locations across all years and ENSO phases document chronic surface water deficit in the Chi River Basin. This finding informs investments in small-scale water storage (farm ponds, check dams, small weirs) and groundwater development for drought resilience, contributing to sustainable water management. The near-invariance of MNDWI across years (CV = 1.4%) indicates that surface water availability is a chronic constraint, not an episodic one—a finding that supports long-term water storage investments.

SDG 13 (Climate Action): The descriptive reference points and understanding of index limitations provided in this study can inform climate-informed drought preparedness. However, we emphasize that the framework presented is provisional and requires validation through longer time series and operational testing.

The policy-relevant observations provided in Section 4.6 align with these SDGs by promoting sustainable agriculture (SDG 2), sustainable water management (SDG 6), and climate resilience (SDG 13). The emphasis on static vulnerability factors—soil type, elevation, distance to river—supports targeted investments that maximize impact for the most vulnerable communities.

4.10. Conclusions of Discussion

The Chi River Basin serves as an effective representative model for understanding drought dynamics across Northeast Thailand’s floodplain environments. Our findings reveal that individual remote sensing indices have limited discriminatory power for drought detection in this setting (Cohen’s d < 0.22, AUC < 0.65, Table 4 and Table 5; Figure 4, Figure 5 and Figure 6), due to sandy soils, low baseline productivity, and chronic surface water deficit. The spatial patterns (Figure 2) and correlation structure (Table 8) confirm that static landscape characteristics dominate over inter-annual climate variability in determining index values and drought vulnerability.

The key message is that remote sensing indices alone are insufficient for drought monitoring in this environment—they must be integrated with static vulnerability mapping, ground-based validation, and seasonal climate forecasts to provide actionable information for decision-makers. This finding is consistent with the growing recognition in the drought research community that effective drought monitoring requires multi-source, integrated approaches that combine satellite observations with in situ data and local knowledge [33,35,36].

Future research should extend the time series to multiple ENSO cycles, incorporate static spatial variables to test interaction effects, validate these findings in other floodplain environments across mainland Southeast Asia, and develop operational drought monitoring systems that integrate satellite indices with ground-based data and crop modeling. With these improvements, the potential contribution to drought resilience and sustainable agriculture in one of Southeast Asia’s most vulnerable regions can be realized.

5. Conclusions

This study characterized six remote sensing indices (NDVI, VCI, SMI, NDMI, MNDWI, NSMI) at 541 agricultural drought-reporting locations in the Chi River Basin, Northeast Thailand, across three years (La Niña 2020, El Niño 2023, neutral 2024). The findings reveal that individual remote sensing indices have limited discriminatory power for drought detection in this sandy soil floodplain environment, with mean differences between high-frequency and low–moderate frequency locations remarkably small (0.008–0.024), effect sizes uniformly small (Cohen’s d = 0.20–0.22), and ROC-AUC values ranging only from 0.53 to 0.64—barely above random classification. VCI and MNDWI showed negligible effect sizes (d = 0.124 and d = 0.094, respectively). Index values at high-frequency locations showed relatively stable values across years (CV < 7% for all indices except NDMI), with limited year-to-year variation, indicating that local factors—sandy soils, deep river incision, and limited irrigation access—dominate over regional climate forcing. However, each ENSO phase was represented by only one year, limiting causal attribution. The perfect correlation between SMI and NSMI (r = 1.00) indicates mathematical redundancy; retaining only SMI is recommended.

Provisional descriptive reference points were derived (NDVI ≈ 0.21, VCI ≈ 0.52, SMI ≈ 0.41, MNDWI ≈ −0.33 at high-frequency locations), but these should be treated as descriptive summaries, not validated operational thresholds. The limited discriminatory power of individual indices (AUC 0.53–0.64) indicates that they are insufficient for standalone drought detection.

Five policy-relevant observations are proposed: re-evaluate threshold-based early warning systems, prioritize irrigation investments based on static vulnerability, address sequential flood–drought hazards, invest in ground-based soil moisture monitoring, and use the descriptive reference points for local index interpretation.

This study contributes to SDG 2 (Zero Hunger) through improved understanding of drought monitoring limitations, SDG 6 (Clean Water) by documenting chronic surface water deficit, and SDG 13 (Climate Action) through climate-informed drought preparedness. The Chi River Basin serves as an effective representative model for Northeast Thailand’s floodplain environments, where remote sensing indices alone are insufficient for drought detection and must be integrated with static spatial variables and ground-based data for effective drought management.