Analysis of the Agricultural Applicability of Sentinel-2 NDVI Anomalies: Comparison with Onion Yield

Abstract

As climate change intensifies yield variation, the South Korean government is developing the Compact Advanced Satellite 500-4 (CAS500-4), whose planned products include vegetation index anomalies. This study offers a pre-operational evaluation of that vegetation index anomaly. Sentinel-2 Level-2A imagery was used to compute year-to-year and five-year normal normalized difference vegetation index (NDVI) anomalies for crop areas in Jeollanam-do, and these were compared with concurrent precipitation and provincial onion-yield departures (2022-2024). The results indicated a general correspondence in the directional trends of the indicators, but also suggested that the strength and nature of this relationship appear to be context-dependent. For instance, in one year, negative NDVI signals were not associated with below-average long-term yields, while in another year, positive environmental indicators coincided with a regional yield decline. Conversely, the long-term anomaly aligned well with the most significant yield loss in a year of presumed sustained pressure,suggesting its utility for diagnosing cumulative stress. The year-to-year NDVI anomaly sensitively detects short-term fluctuations but can be misleading due to base effects or dominant local factors. The five-year normal anomaly, in contrast, provides a more stable diagnosis of cumulative stress. These findings underscore the importance of employing both baselines concurrently to capture short-term impacts and long-term trends as CAS500-4 data become operational.

1. Introduction

The increasing frequency of droughts and localized heavy rainfall, widely attributed to climate change, contributes to the volatility of the global food-supply chain (Lobell et al., 2011; IPCC, 2022; Zhu et al., 2022; Hultgren et al., 2025). Empirically, drought years are associated with average reductions of 5–15% in major crop harvests, and cumulative losses by 2050 are projected to exceed the additional demand expected overthe same period (Leng and Hall, 2019; OECD, 2025). In response, international agencies and agricultural-commodity markets now recognize weeks-ahead, high-resolution crop monitoring and yield forecasting as a core capability (Bolton and Friedl, 2013; Funk et al., 2015). For this purpose, normalized difference vegetation index (NDVI) anomalies arewidely used. By calculating the difference between the current NDVI and its long-term normal, these anomalies are designed to isolate abnormal vegetation growth from seasonal effects. This approach can enable early detection of drought- or pest-induced yield losses and help improve the accuracy of crop-forecast models (Anyamba et al., 2001; Liu et al., 2001; Bayarjargal et al., 2006; Krishna et al., 2009; Li et al., 2014; Meroni et al., 2019).

Under its Third National Space Development Master Plan in 2018, the Government of the Republic of Korea is developing the Compact Advanced Satellite-500 4 (CAS500-4), with a planned launch in 2025. The satellite will provide 10 m spatial-resolution imagery with a nominal three-day revisit, offering a planned product of NDVI anomalies relative to a five-year climatology. Although the MODIS-based NDVI departure (250 m, 8-day) from the United States Department of Agriculture (USDA) - Global Agricultural Monitoring (GLAM) system has proved effective for early detection of crop-sowing failure (Whitcraft et al., 2015; Rembold et al., 2019), its 250 m resolution is often inadequate in the Korean context, where 32% of fields are smaller than 0.5 ha and more than 70% are smaller than 1 ha (Kim et al., 2003). Accordingly, the 10 m NDVI-anomaly product from CAS500-4 is anticipated to meet the demand for high-resolution national agricultural monitoring, yet its detection sensitivity and operational suitability still require quantitative assessment.

The literature on agricultural monitoring using NDVI anomalies has largely followed two parallel paths. One stream focuses on near-real-time applications, such as yield forecasting, where short-term (year-to-year) anomalies are primarily used to capture the impact of acute meteorological events on specific crop growth stages (Bolton and Friedl, 2013; Meroni et al., 2019). The other stream addresses broader seasonal productivity and drought assessment, typically employing long-term anomalies calculated against a multi-year climatology to identify sustained deviations from the norm (Ji and Peters, 2003; Funk et al., 2015). Although numerous studies have successfully utilized these distinct approaches, a direct, systematic comparison of their performance against regional yield data within a single framework remains scarce. Asystematic, quantitative comparison is therefore necessary to clarify their complementary value and operational scope.

As CAS500-4 is a new platform, historical data required to evaluate its NDVI-anomaly algorithm are unavailable. Sentinel2 imagery,which sharessimilarspectral and spatialspecifications and has provided continuous observations since 2015, serves as an effective surrogate dataset for such an evaluation (Kim et al., 2023). Thisstudy focuses on onion (Allium cepa), whose bulbing and harvesting period (April–July) coincides with the domestic drought season. The crop is also highly sensitive to water stress; physiological responses such as reduced photosynthesis and stomatal conductance under water-deficit conditions can lower yields by 20% to 65%, depending on the cultivar (Ayas, 2019; Wakchaure et al., 2021; Sansan et al., 2024). These characteristics make the onion an appropriate test crop for assessing NDVIbased stress detection.

The objectives of this study are to (i) generate short-term (year-to-year) and long-term (five-year mean) NDVI-anomaly datasets for crop areas in Jeollanam-do using 30-day composite Sentinel-2 imagery; (ii) compare and validate these anomalies with monthly precipitation departures and provincial onion-yield departures in Jeollanam-do for 2022–2024; and (iii) evaluate the accuracy and utility of the CAS500-4 NDVI-anomaly algorithm. The findings are expected to provide scientific evidence for designing an early-warning and disaster-response system based on CAS500-4 data.

2. Materials

2.1. Study Area

Among Korea’s 17 first-tier administrative divisions, Jeollanamdo contains the largest extent of upland crop areas. According to Statistics Korea (2024), the province accounted for approximately 18% of the national upland area in 2023 and produced over 31% of the country’s onions, the highest proportion nationwide. This makes the province a critical region for monitoring onion production dynamics in Korea. The upland-crop mask used in this study was derived from the Smart-Farm Map provided by the Ministry of Agriculture, Food and Rural Affairs (MAFRA). Fig. 1 illustrates the spatial distribution of upland crop areas in Jeollanam-do, along with the Sentinel-2 tiles covering the province (51SYT, 51SYU, 51SYV, 52SBC, 52SBD, 52SBE, 52SCC, 52SCD, 52SCE).

Fig. 1. Study area crops and Sentinel-2 tiles.

2.2. Sentinel-2 Data

This study selected Sentinel-2 as a surrogate dataset for the upcoming CAS500-4 satellite, given the similarities in its key spectral characteristics. As detailed in Table 1, while CAS500-4 is designed to provide a higher spatial resolution (5 m) than Sentinel-2 (10 m), the two sensors possess comparable core spectral bands in terms of number and wavelength range. This spectral similarity makes Sentinel-2 data highly suitable for this pre-operational evaluation of the NDVI anomaly algorithm.

Table 1. Comparison of spectral band specifications between CAS500-4 (planned) and Sentinel-2

¹⁾ Source:Lim et al. (2022) and Kim and Kim (2024),

²⁾ Source: NASA Earthdata (2025).

To encompass the study area, all available Sentinel-2 Level-2A images (10 m) were acquired for the period 2017–2024. The spatial footprint comprises nine tiles across UTM Zones 51 and 52 (51SYT, 51SYU, 51SYV, 52SBC, 52SBD, 52SBE, 52SCC, 52SCD, 52SCE), which collectively cover over 99% of the upland crop area in Jeollanam-do (Fig. 1). With a nominal 5-day revisit cycle, the Sentinel-2A/B pair provides sufficient temporal density for 30-day compositing (Drusch et al., 2012).

Bottom-of-atmosphere reflectance and associated metadata (e.g., acquisition date, solar geometry) for each tile were batchdownloaded via the Copernicus Open Access Hub. Subsequent preprocessing—described in Section 3—included the application of the upland-crop mask, cloud-quality filtering, and 30-day compositing. Owing to its high spatial and temporal resolution, Sentinel-2 enhances the detection of growth dynamics in small (<1 ha) and fragmented fields relative to MODIS 250 m data. It is therefore well suited to crop-monitoring studies in the mosaicked, intensively managed landscapes of Jeollanam-do (Kim et al., 2003).

2.3. Climate and Yield Data

To characterize the climatic context, monthly accumulated precipitation (mm) from January 2017 to December 2024 was compiled from 16 Automatic Synoptic Observation System (ASOS) and Automatic Weather System (AWS) stations located within the provincial boundary and operated by the Korea Meteorological Administration (KMA). Monthly totals are derived from daily sums that meet the KMA completeness criterion of ≥80% valid observations or have appropriate substitute data.

Onion-production statistics for Jeollanam-do were obtained from annualreports published by the agricultural statistics service of the Jeollanam-do provincial government, in coordination with the MAFRA. The reports provide data on cultivated area (ha), yield per unit area (kg/10 are, where 1 are = 100 m²), and total production (t) at the provincial level. For this study, yield per unit area was selected for comparison, as total production can be influenced by temporal price variations and farm-management factors. Using the unit-area metric helps to minimize such confounding effects in joint analyses with NDVI anomalies and precipitation departures.

3. Methods

3.1.Image Pre-Processing

The Sentinel-2 time series(2017–2024) used in this study followed two distinct preprocessing paths. As images from 2017–2018 are distributed only as Level-1C (top-of-atmosphere reflectance), they were converted to Level-2A(bottom-of-atmosphere reflectance) using the Sen2Cor module from the European Space Agency (ESA). From 2019 onward, data were already available as Level2A and were used in their original form. A notable change occurred in February 2022 when the Sentinel-2 processor was upgraded to Processing Baseline (PB) 4.00 (European Space Agency, 2021). This update introduced a radiometric offset of +1,000 digital numbers (DN) to all spectral bands to minimize the loss of pixels with negative reflectance values over very dark surfaces, such as water bodies, after atmospheric correction. To ensure the radiometric consistency of the time series, this offset was removed from all post-PB 4.00 images, a necessary step recommended by ESA to prevent discontinuity. Any resulting values below zero were then clipped to zero, thereby normalizing the reflectance range.

Vegetation status was then quantified using theNDVI, calculated from near-infrared (ρNIR) and red (ρRED)reflectances as Eq.(1):

NDVI = (ρNIR – ρRED)/(ρNIR + ρRED)       (1)

A high NDVI value typically indicates expanded leaf area and vigorous photosynthesis, whereas a low value may suggest moisture stress or poor growth (Tucker, 1979; Myneni et al., 1995). As described in Section 3.2, this NDVI time series was subsequently composited at 30-day intervalsto generate the base layers for anomaly calculations.

3.2. Composite NDVI Generation

Using Sentinel-2 Level-2A data from 2017 to 2024, a 30-day composite NDVI was generated as the base layer for anomaly computation. After NDVI calculation, the scene classification layer (SCL) was used to mask clouds (classes 8–9), haze (10), cloud shadow(3), and defective pixels(0–1). The remaining valid pixels were composited with the maximum value composite (MVC) technique, which selects the highest NDVI within the 30-day window and effectively removes noise from clouds and haze (Holben, 1986; Goward et al., 1994). Fig. 2 illustrates how individual Level-2A images are converted to NDVI and merged into a single 30-day composite, resulting in the disappearance of clouds apparent on the daily scenes. These composites serve as the input for the year-to-year (short-term) and five-year normal(long-term) NDVI-anomaly calculations described below

Fig.2. Example of composite NDVI creation process(April 2024,52SCE).

A 30-day interval was chosen because the KMA provides accumulated precipitation at monthly resolution; matching the temporal scale facilitates joint analysis of NDVI anomalies and precipitation departures. Using monthly composite NDVI thus reduces noise in the source data while clarifying relationships between crop-growth patterns and climate variables.

3.3. NDVI-Anomaly Generation

Two types of NDVI anomaly were derived from the composite time series: (i) short-term (year-to-year) anomaly. This metric is calculated by subtracting the composite NDVI of the same period in the previous year from that of the target year. It is designed to reveal how vegetation vigor has changed relative to the prior season, allowing for sensitive tracking of acute events such as drought shocks, disease outbreaks, or shifts in management practices within a single year. (ii) Long-term (five-year normal) anomaly: this metric is calculated by subtracting the median composite NDVI of the preceding five years from the currentyear value for the same period. A five-year reference period incorporates recent climate-trend information while allowing the baseline to update annually. The use of the median, rather than the mean, provides a more robust baseline by reducing potential distortion from outliers caused by extreme weather events or sensor errors (e.g., cloud contamination) in any single year (Ser and Shin, 2024). For example, the anomaly for May 2024 was computed using the median NDVI of May 2019–2023 as the baseline.

To prevent bias, pixels that were fully cloud-obscured in either the target or reference image were excluded from the calculation and remained as no-data values. Fig. 3 summarizes the step-wise procedure for calculating both the short- and long-term NDVI anomalies.

Fig. 3. Conceptual workflow for short- and long-term NDVI-anomaly calculation.

3.4. Calculation of Precipitation and Yield Departures

To facilitate a direct comparison with the NDVI anomalies, departures for both precipitation and onion yield were computed using the same short-term (previous year) and long-term (fiveyear normal) baselines. For precipitation, monthly totals for Jeollanam-do were used to calculate two departure metrics: (i) the short-term precipitation departure was defined as the current-month precipitation minus that of the same month in the preceding year; and (ii)the long-term precipitation departure was the current-month precipitation minus the five-year normal for that month. A negative (−) value indicates drier-than-normal conditions, whereas a positive (+) value indicates wetter-thannormal conditions. To align with the 30-day NDVI composites, only precipitation departures from the corresponding months were used in the analysis.

Yield departures for Jeollanam-do were derived from the provincial agricultural statistics described in Section 2.3. Using the yield per 10 are (kg/10 are), (i) the short-term yield departure was defined as the current-year yield minus the previous-year yield, and (ii) the long-term yield departure as the current-year yield minus the five-year normal. A negative (−) value denotes a decrease relative to the reference, and a positive (+) value an increase. These short- and long-term precipitation and yield departures serve as reference indicatorsfor assessing the explanatory power of their corresponding NDVI anomalies. Mean values for the critical growth period (April–July) were subsequently calculated to enable a quantitative, year-by-year comparison.

3.5. Analytical Approach

After aligning the time series for all six indicators—short- and long-term NDVI anomalies, precipitation departures, and yield departures—a monthly trend graph was plotted for the period 2017–2024. For the key growth period of April-July, the mean, maximum, andminimum values of each indicator were calculated, and precipitation was averaged over the same interval. A Python script was developed to automate this workflow and generate an annual summary table of the six indicators for 2022–2024.

The graphical and tabular outputs were interpreted in a threestep process. First, the concordance of signs among the NDVI anomaly, precipitation departure, and yield departure was assessed. To filter out minor statistical noise and focus on agriculturally meaningful variations, specific thresholds were established. The threshold for the NDVI anomaly was set to |NDVI anomaly| ≥ 0.025, a value identified in previous research as representing a mild to moderate level of vegetation change (Ser and Shin, 2024).

To maintain a consistent analytical framework, the precipitation departure threshold was aligned based on its statistical significance. The Z-score for an NDVI anomaly of ±0.025, considering both short- and long-term distributions, was approximately 0.5. The corresponding precipitation departure for a Z-score of 0.5 was calculated to be approximately 30 mm. Therefore, only variations meeting the criteria of |NDVI anomaly| ≥ 0.025 or |precipitation departure| ≥ 30 mm were considered in the sign-concordance analysis. Second, within each year, the amplitudes of the shortterm and long-term NDVI anomalies were compared to evaluate the relative sensitivity of each metric. Finally, the degree to which the sign patterns of the NDVI anomalies and precipitation departures mirrored those of the yield departures was examined to explore potential linkages among the indicators.

4. Results

This section first presents a time-series analysis of short- and longterm NDVI anomalies and precipitation departures in Jeollanamdo upland crop areas from 2022 to 2024 (Fig. 4). Subsequently, mean values for the April–July period(Table 2) are used to quantify the relationships among NDVI anomalies, precipitation departures, and onion yield departures for Jeollanam-do, thereby examining crop responses to weather variability.

Fig. 4. Time-series pattern analysis of NDVI anomalies and precipitation departures.

Table 2. April–July means and departures of each indicator (Units: NDVI, mm, kg/10 are)

4.1. Time-Series Pattern Analysis

Fig. 4 superimposes monthly short- andlong-term NDVI anomalies (lines) on their corresponding precipitation departures(bars)for 2022–2024. The critical growth period (April–July) is shaded in light green. The left Y-axis represents NDVI anomalies, and the right Y-axis represents precipitation departures, allowing for a visual comparison of the sign and magnitude among the four indicators.

During the 2022 growth period, both precipitation departures remained negative, with the short-term departure falling below −100 mm in May and July. Correspondingly, the short-term NDVI anomaly decreased to −0.18 in July, while the long-term anomaly was consistently negative. The absolute value of the short-term NDVI anomaly was more than twice that of the longterm anomaly, which reflects the higher sensitivity of the yearto-year metric to meteorological fluctuations.

In 2023, a different pattern emerged. Short-term precipitation departures surged to +253 mm in May and +343 mm in July, while the short-term NDVI anomaly rose to +0.13 in July and +0.21 in August. In contrast, the long-term precipitation departure hovered around +100 mm, and the long-term NDVI anomaly remained within a narrow range of −0.02 to +0.02. This led to a large amplitude gap between the two NDVI metrics.

In 2024, the indicators diverged in sign and magnitude. Although both precipitation departures were positive in February-March, the short-term departure shifted to approximately −80 mm during April–July, whereas the long-term departure remained slightly positive (≈+10 mm). Both NDVI anomalies, however, were negative from April onward. Notably, the long-term anomaly fellto−0.10, exceeding the short-term anomaly (−0.08)in absolute value.

In summary, the 2022 and 2023 growth periods exhibit concordant sign patterns between precipitation departures and NDVI anomalies. In 2024, however, the two precipitation indicators diverged in sign, and the long-term NDVI anomaly showed a greater decrease than its short-term counterpart. This suggests that while the patterns of vegetation anomalies and precipitation departures generally show a concordant trend, the data suggest the presence of vegetation stress that may not be fully explained by concurrent precipitation alone.

4.2. Comparison with Crop Yield

Table 2 lists the mean values and departures for the six indicators —short- and long-term NDVI anomalies, precipitation departures, and onion yield departures—averaged over the April–July period for 2022–2024.

In 2022, a complex relationship between the indicators was observed. The short-term NDVI anomaly was −0.068, and the short-term precipitation departure was −80.0 mm. This was associated with a substantial short-term yield loss of −1,533.0 kg/10 are. This alignment in the negative direction across all three short-term indicators suggests a concordant response to the year-on-year environmental conditions. This large year-on-year decrease can be partly contextualized by the baseline year, 2021, was a bumper crop year for onions. In contrast, from a long-term perspective, while the NDVI anomalywas negative (−0.025), the yield departure relative to the five-year normal was positive at +251.0 kg/10 are. Thus, although conditions worsened compared to the previous exceptional year, the 2022 yield remained above the five-year average. This observation suggests a decoupling between vegetation greenness and long-term yield outcomes.

In 2023, the indicators in Jeollanam-do showed a notable divergence from yield trends. Despite a positive short-term NDVI anomaly (+0.031) and a strong positive precipitation departure (+183.8 mm), a corresponding yield recovery was not observed in the province. Instead, both the short-term (−536.0 kg/10 are) and long-term (−572.0 kg/10 are) yield departures were negative. This regional downturn differed from the national trend forthe same year, where average onion yields per unit area showed a slight increase.

In 2024, the pattern of a cumulative decline became more apparent. Both NDVI anomalies were negative, with the longterm anomaly (−0.101)showing a greater magnitude of decrease than the short-term anomaly (−0.079). This was concurrent with the most substantial yield losses of the study period, with a shortterm departure of −1,029.0 kg/10 are and a long-term departure of −1,910.0 kg/10 are. The alignment of the largest negative long-term NDVI anomaly with the largest long-term yield loss indicates a potential utility for this metric in identifying conditions of presumed accumulating crop stress. Such an alignment is indicative of a pathway where apparent sustained environmental pressure may contribute to a decline in vegetation health (NDVI), which in turn is associated with a correspondingly severe loss in crop yield.

A cross-comparison of the three years highlights that the relationship between satellite-derived vegetation indices and regional crop yield is complex and context-dependent. The 2022 results underscore the importance of the baseline year in interpreting short-term changes and indicate that negative vegetation signals do not always correspond to below-average long-term yields. The 2023 data show that regional yield dynamics can diverge from both local environmental indicators and broader national trends, pointing to the influence of province-specific factors. Finally, the 2024 findings suggest that the long-term anomaly may be a relevant indicator when stress conditions are sustained and cumulative. Collectively, while the three-way interaction among these indicators proved intricate, the pairwise relationships between NDVI anomalies and precipitation departures, and between NDVI anomalies and yield departures, generally showed concordant trends.

5. Discussion

The findings of this study first lend support to the generally expected relationship between satellite-derived vegetation indices and agricultural yield. However, these same findings, grounded in region-specific yield data for Jeollanam-do, suggest that this relationship is more complex than often assumed. The year-to-year comparisons from 2022 to 2024 indicated that this relationship is not always linear and is highly dependent on multiple contexts, including baseline conditions and local agronomic factors. This discussion interprets the year-specific patterns observed in the results and considers the broader implications for agricultural monitoring using CAS500-4.

The 2022 results highlight the importance of baseline context in interpreting short-term anomalies and suggest the potential for decoupling between vegetation greenness and yield. The substantial short-term yield decrease (−1,533.0 kg/10 are) appears to have been influenced not only by drought conditions but also by the base effect of the preceding year, 2021, which was a bumper crop year with an exceptionally high production baseline. More notably, a decoupling was observed from a longterm perspective: despite a negative long-term NDVI anomaly, the yield was above the five-year average (+251.0 kg/10 are). A possible explanation is that while drought conditions stressed the vegetation'sfoliage (lowering NDVI), other factors, such as high solar radiation combined with sufficient farmer-led irrigation, may have supported the development of the onion bulb itself. This suggests that for bulb orroot crops, canopy greenness alone may not always be a direct proxy for final yield.

The 2023 data provided insight into region-specific dynamics, showing that local yield outcomes can diverge from both satellite indicators and broader national trends. In Jeollanam-do, positive shifts in precipitation and short-term NDVI did not correspond with a yield recovery; instead, yields declined, a pattern contrary to the national average increase for that year. This suggests that the primary limiting factors for yield in Jeollanam-do during 2023 were likely not captured by the monthly precipitation and NDVI data. Plausible region-specific factors include meteorological conditions not reflected in monthly totals, such as poor timing of rainfall or insufficient solar radiation during critical growth stages. Additionally, localized agronomic issues, such as a higher prevalence of onion downy mildew or differing soil and cultivar characteristics, could have contributed to this regional downturn. The influence of management decisions, such as adjustments in farming inputs based on fluctuating onion market prices, also cannot be ruled out.

Inc ontrast, the 2024 results present a scenario where cumulative stress appears to become the dominant signal. The strong alignment between the most negative long-term NDVI anomaly (−0.101) and the most severe long-term yield loss (−1,910.0 kg/10 are) suggests that after several years of climate volatility and other pressures, the agricultural system's resilience may have been diminished. In such chronic stress situations, the long-term anomaly, which is robust against single-year base effects, appears to be a more relevant indicator of the overall production trajectory than the more volatile short-term anomaly. In this context, 2024 can be seen as an important case study, suggesting that under conditions of sustained stress, the underlying linkage between vegetation health and agricultural output can become more pronounced and readily observable.

These year-to-year analyses underscore the complementary roles of the two anomaly baselines. The short-term (year-to-year) anomaly can be highly sensitive for detecting acute inter-annual changes but can be misleading if the baseline is anomalous or if non-climatic local factors are dominant. Conversely, the longterm (five-year normal) anomaly offers a more stable baseline for assessing cumulative trends and can become particularly insightful for diagnosing chronic stress over multiple seasons. Ultimately, these findings suggest that an effective monitoring strategy involves not choosing one anomaly metric overthe other but employing them concurrently. Researchers and analysts could leverage both the high sensitivity of the short-term anomaly and the stability of the long-term anomaly, selecting the most relevant indicator according to the specific context and analytical purpose, whether it is capturing acute weather impacts or assessing cumulative stress.

The implications of these findings are directly relevant to the operational application of CAS500-4. The satellite’s planned fiveyear normal NDVI anomaly product is expected to be valuable for providing a stable, long-term assessment of vegetation conditions. However, this study suggests that the additional generation of a year-to-year anomaly product would be highly beneficial, providing essential context and sensitivity to acute events. It is also important to position these NDVI anomalies appropriately within the data processing chain: they serve as powerful Level 3 diagnostic indicators of vegetation stress, not as standalone Level 4 predictors of final yield. A noteworthy finding is that while a qualitative correspondence was frequently observed between the patterns of NDVI anomalies, precipitation, and yield, translating this into a direct, quantitative correlation appears to be confounded by the complex interplay of other influencing factors. An effective national crop monitoring system could utilize these NDVI anomaly products as key inputs into more sophisticated yield models that also integrate other variables, including temperature, solarradiation, soil data, and management information.

6. Conclusions

This study conducted a pre-operational evaluation of the five-year normal NDVI-anomaly algorithm for the upcoming CAS500-4, with its launch scheduled for 2025. Given the lack of a historical archive for CAS500-4, 10 m Sentinel-2 imagery served as a surrogate. A year-to-year NDVI anomaly was also computed to assess the utility of a dual-baseline monitoring system in the context of increasing climate-related agricultural risks.

By aligning six indicators—NDVI anomalies, precipitation departures, and onion yield departures—at a consistent regional scale, this study moved beyond proxy-based analysis to reveal a complex, context-dependent relationship between vegetation indices and actual yield. Despite revealing this complexity, the analysis also consistently found a general correspondence between the directional trends of the indicators. The principal findings are threefold. First, the 2022 results indicated the influence of baseline conditions(the base effect). They showed that a negative NDVI signal does not always correspond to a below-average long-term yield, revealing a potential decoupling between the two. Second, the 2023 analysis identified a region-specific divergence, where local yield in Jeollanam-do declined despite positive environmental indicators and contrary to nationaltrends, pointing to the significant role of local agronomic factors. Third, the 2024 case suggested that under conditions of presumed sustained pressure, the long-term anomaly aligns well with severe, cumulative yield losses.

These results support the view that the two NDVI anomaly baselines offer complementary strengths. The year-to-year anomaly effectively captures acute inter-annual fluctuations but can be susceptible to base effects and may be a poor indicator when nonclimatic local factors are dominant. Conversely, the long-term anomaly provides a more stable measure for assessing cumulative trends. Therefore, an effective monitoring strategy could employ both baselines concurrently—a dual-baseline approach—to enable a more nuanced assessment that distinguishes short-term variability from long-term trajectories.

This study has several limitations. First, the analysis did not include other critical meteorological variables, such as temperature and solar radiation, which restricts a full explanation of vegetation responses to compound climate conditions. Second, it did not account for the spatial heterogeneity of soil properties or differences in crop cultivars, which can also influence yield outcomes. Third, the three-year analysis period is insufficient to characterize long-term variability comprehensively. Finally, inherent differences in sensor specifications between Sentinel-2 and CAS500-4 may affect long-term consistency without robust cross-calibration.

Future research should aim to address these limitations. This includes integrating temperature and solar radiation data from sources like the KMA’s ASOS network to model compound stresses, as well as incorporating finer-scale data on soil types and management practices. Looking ahead to the operational phase of CAS500-4 following its launch, it is also crucial to define the role of its products within an operational framework. The Level 3 NDVI anomaly products evaluated in this study are intended to serve as foundational data for more advanced applications. They can serve as valuable inputs for advanced Level 4 products, such as sophisticated yield forecast models, which can integrate multiple data layers.Acritical nextstep following the launchwill be to validate these anomaly-generation methods using native CAS500-4 imagery and begin their integration into an advanced agricultural monitoring system for Korea.

Author Contributions

Conceptualization: Ser JH, Shin JI; Data curation: Ser JH; Methodology: SerJH, Shin JI; Formal analysis: SerJH; Validation: Ser JH; Project administration: Shin JI; Funding acquisition, Supervision: Shin JI; Writing–original draft: Ser JH; Writing– review & editing: Ser JH, Shin JI.

Conflicts of Interest

No potential conflict of interest relevant to this article was reported.

Funding

This paper was conducted at the Seoul Institute (2025-ER-06, Infratechnology and Software Development of Agriculture Satellite Information), with support from the Rural Development Administration research project (RS-2021-RD009991).

Data Availability Statement

The data presented in this study are available on reasonable request from the corresponding author. Because the dataset containssensitive information, it is not publicly archived;requests will be evaluated individually and released only after aninstitutional security review.