Hail Damage Detection: Integrating Sentinel-2 Images with Weather Radar Hail Kinetic Energy

Highlights

What are the main findings?

• Integration of radar-derived Hail Kinetic Energy with Sentinel-2 NDVI enables accurate hail damage detection.
• Multi-temporal ΔNDVI analysis reveals a clear decline of vegetation impact over time after hail events.

What are the implications of the main findings?

• The main findings demonstrate the critical role of time in assessing vegetation recovery after severe hailstorms.
• The study provides a rapid, scalable method for operational crop monitoring and agricultural insurance assessment.

Abstract

Hailstorms represent one of the most damaging convective hazards for agriculture, yet quantifying their impacts at a landscape scale remains challenging due to their localized and short-lived nature. In this study, we combine weather radar parameters and Sentinel-2 multispectral imagery to assess vegetation damage caused by two major hail events in northeastern Romania: Rădăuți (17 July 2016) and Dolhasca (30 July 2020). Radar-derived hail kinetic energy (HKE) was used as a rapid temporal indicator of hail occurrence, with a threshold of 300 J m⁻² applied to delineate potentially affected areas. Sentinel-2 Level-1C imagery, selected under strict temporal and cloud cover criteria, was processed to generate pre- and post-event Normalized Difference Vegetation Index (NDVI) maps, from which NDVI differences (ΔNDVI) were computed. Thresholds of 0.10 and 0.20 were applied to identify moderate and severe vegetation stress, respectively. The results demonstrate strong spatial correspondence between radar-derived HKE cores and Sentinel-2 ΔNDVI reductions. In Rădăuți, where only one post-event image was available, ΔNDVI thresholds identified between 2236 and 5856 ha of affected vegetation within the HKE > 300 J m⁻² zone. In Dolhasca, where three post-event images were available (5, 8, and 15 days), the analysis revealed 6200–9100 ha affected at 5 days, decreasing to 4800–7200 ha at 8 days, and further to 3100–5600 ha at 15 days post-event. This temporal gradient highlights both the recovery of vegetation and the diminishing sensitivity of the ΔNDVI signal with increasing time elapsed since the event. Analysis by land use classes showed arable fields to be the most sensitive, followed by orchards and pastures, while forests exhibited smaller but persistent declines. This study demonstrates the robustness of integrating radar-derived hail kinetic energy with Sentinel-2 NDVI differencing for the spatiotemporal assessment of hail damage. The approach provides both rapid detection and temporally resolved mapping of hail damage, underlining the critical role of time as a determining factor in impact assessments. These findings have strong implications for operational crop monitoring, disaster response, and risk management in hail-prone regions.

1. Introduction

In recent years, the economic impact of hailstorms has become increasingly complex. Numerous studies suggest that certain regions worldwide are experiencing a rise in hailstorm frequency, largely associated with climate change [1,2,3]. For instance, between 2006 and 2018, eight large-hail events across central and southern Europe caused losses exceeding 1 billion USD each, primarily due to extensive crop destruction, damaged infrastructure, and severe property losses [4]. Similarly, Romania has faced significant hailstorm-induced damages in recent years, particularly to agricultural crops. Notably, on 30 May 2013, hailstorms in southeastern Romania resulted in damages estimated at 2 million euros [5], while a storm on 18 June 2016, in the country’s northeast caused damages of approximately 14 million euros [6]. All currency values are reported as nominal amounts from the year of the event or source publication.

The most recent hail climatology for Romania, based on data from 105 meteorological stations collected between 1961 and 2014, highlights northeastern Romania as having a moderate hailstorm frequency, with 6–9 hail days annually [7]. However, hail reports from the European Severe Weather Database (ESWD) for Romania [8] reveal areas of higher hailstorm density, particularly in the northeastern and northwestern regions of the country. The land-use practices in these regions heighten vulnerability to hail damage [9,10], and Romania’s status as one of Europe’s largest producers of grapes, cereals, and fruits further amplifies the risk [11].

After a hailstorm, insurance claims adjusters assess crop yield losses by inspecting the damaged area. The precision of these evaluations depends significantly on accurately identifying damage zones, which enables extrapolation of sampled data to the entirety of the affected field. Currently, these assessments are largely conducted through visual inspection, leaving the adjuster without prior information about the extent or severity of the damage before arriving at the site. One potential improvement involves employing remote sensing technologies, such as meteorological radar and Sentinel-2 satellite data, to detect hailstorm-induced changes in crops. The effectiveness of remote sensing in identifying and classifying alterations in crops’ physical, physiological, and chemical characteristics depends on the type of sensor used and the algorithms applied for data analysis. Satellite imagery is invaluable due to its ability to serve as a versatile information source, enabling analysis from multiple perspectives. These include studies on land use dynamics [12,13] and the impacts of land use changes on protected areas [14,15]. Satellite data also play a crucial role in vegetation monitoring, covering natural ecosystems, pastures [16,17], forests [18,19], and agricultural crops, including production estimates [20,21].

Vegetation, being directly affected by meteorological and climatic conditions, exhibits changes that can be effectively captured through satellite remote sensing. For instance, during drought conditions, satellite data have been extensively used to document vegetation stress. Examples include studies on soil moisture dynamics [22], drought monitoring in northeastern Romania’s orchards [23], and broader drought analyses [24,25]. Similarly, satellite imagery has been instrumental in assessing areas affected by wildfires in natural vegetation [26,27].

Hail, due to its destructive nature, can inflict physical damage on vegetation, which is detectable through satellite imagery. Hail scars or swaths—visible consequences of hailstorms—can be identified in both agricultural crops and natural vegetation. For example, using visible spectrum data from GOES-8, Klimowski et al. (1998) [28] detected hail-damaged areas as narrow, bright streaks. Optical data from ground-based, airborne, and satellite platforms [29,30,31,32] have proven effective in estimating hail-induced defoliation in corn crops. Remote sensing approaches have also been proposed to reduce the cost of in situ inspections for mapping damage [33].

Several advancements highlight the integration of remote sensing and other datasets for hail damage detection. Bell (2016) [34] and Gallo et al. (2012) [33] linked MODIS-derived NDVI reductions with radar data, demonstrating that radar footprints correspond to regions with decreased Normalized Difference Vegetation Index (NDVI) values. Prabhakar et al. (2019) [35] used Landsat 8 satellite imagery to identify pre- and post-hail differences across various crop types. Gobbo et al. (2021) [36] developed an innovative method combining Landsat 7 and 8 imageries with crop modeling to improve hail damage estimation in maize, a promising tool for the insurance industry. Ha et al. (2022) [37] introduced an algorithm using Sentinel-1 and Sentinel-2 satellite data to automatically delineate hail damage zones with high accuracy, enabling more transparent and comprehensive damage assessments without reliance on extrapolated in situ observations.

The integration of weather radar data with spectral imagery has been proposed as an effective approach for monitoring large-scale hail events [33,38]. However, typically coarse spatial resolution (ranging from 0.5 km near the radar to several kilometers at longer ranges) of these products presents significant challenges when analyzing highly fragmented agricultural landscapes. To evaluate the destructive potential of hail, studies have established correlations between radar measurements and hail-related damage. For instance, several authors [5,39,40,41] have linked hail size to convective cloud parameters detected by the weather radar, such as maximum reflectivity and VIL density. Hohl et al. (2002a, 2002b) [42,43] developed a relationship between hail-induced damage to vehicles and buildings and the hail kinetic energy (HKE) derived from radar observations. The concept of HKE was originally introduced by Waldvogel et al. (1978a, 1978b) [44,45] and has since been widely applied in studies examining hail damage [5,43,44,46,47].

In Romania, research on hail damage assessment is relatively scarce. Cică et al. (2015) [5] investigated the relationship between single-polarization radar-derived parameters and the detection of severe hailstorms and their surface impacts in southern Romania. Their findings suggested that radar-derived products, including composite reflectivity (CR), vertically integrated liquid (VIL), echo top heights, vertically integrated liquid density (VILD), and hail kinetic energy (HKE), combined with ground-based observations, are insufficient on their own for accurately assessing hailstorm damage potential. This limitation arises primarily from the inherent difficulty of single-polarization radar parameters to robustly link storm intensity to localized surface impact, compounded by the scarcity of comprehensive ground-based validation data. More recent research by Angearu et al. (2022) [48] has demonstrated the utility of remote sensing for detecting and evaluating hailstorm damage to crops. Their case study, focusing on southeastern Romania on 20 July 2020, utilized data from Sentinel-2A, Landsat-8, Terra MODIS, Sentinel-1 SAR, and weather radar. The analysis identified 3142.98 hectares of affected land based on an NDVI threshold (>0.05). The severity of hail damage was shown to correlate with higher Land Surface Temperature (LST) differences (>12 °C) and reductions in NDVI values (0.4–0.5). Soil Water Index (SWI) differences further validated damage patterns in the western part of the study area, consistent with radar reflectivity values exceeding 55 dBZ, which indicated medium-to-large hail.

In this study, we exploit the complementary strengths of weather radar and high-resolution optical imagery to assess short-term vegetation damage caused by hail events. Radar-derived kinetic energy (HKE) provides near-real-time evidence that hail fell over broad areas, while Sentinel-2 NDVI enables the fine-scale (10 m spatial resolution) quantification of vegetative impact. By using HKE polygons exceeding 300 J m⁻² as an initial temporal anchor (indicating a high likelihood of damaging hail), and subsequently mapping NDVI difference within these areas at multiple post-event intervals, we demonstrate how the time dimension—both the timing of detection and the temporal evolution of vegetation response—critically shapes damage assessment and interpretation.

2. Materials and Methods

This study integrates two complementary data sources in order to assess hailstorm impacts on vegetation. Weather radar observations provide high-frequency information on storm intensity and structure, allowing the rapid identification of hail-producing cells and the delineation of areas with high kinetic energy. These data serve as a temporal and spatial indicator of hail occurrence. In contrast, Sentinel-2 multispectral imagery offers high-resolution optical measurements that capture vegetation conditions before and after hail events. By computing spectral vegetation indices such as NDVI, the satellite data enable the quantitative estimation of crop and canopy stress at parcel scale. The combination of radar-derived hail kinetic energy contours and Sentinel-2 vegetation indices thus allows both the rapid detection of hail hazard and the detailed mapping of its environmental consequences.

2.1. Radar Data and Products

The study area, located in northeastern Romania, is adequately covered by the S-band weather radars Soroca and Cornești (Republic of Moldova). These radars, operated by the Department for Active Influences on Hydrometeorological Processes of the Republic of Moldova during the operational season, have been upgraded to provide digital data transmission (Figure 1) [49,50]. To minimize measurement interference from intervening clouds, which can attenuate reflectivity through scattering and absorption, a radar with an unobstructed line of sight to the hailstorm was selected. The key specifications of the employed radars are summarized in

Table 1. Weather radars from those data were used.

Radar ID	Equipment Type	Elevation (m)	Beam Width (Degree)	Volume Scan Length (Minutes)	Nr. of Scan	Minimum and Maximum Antenna Elevation Angles
MDKR	MRL -5 S-band (2950 MHz)	351	1.5	3	18	0–84
MDSR	MRL -5 S-band (2950 MHz)	196	1.5	3	18	0–84

.

To accurately determine the timing and location of hail occurrences, radar-derived parameters directly associated with the likelihood of hail formation were employed. Six such parameters were used to characterize the selected hailstorms (

Table 2. Radar parameters used in study.

No.	Parameter Acronym	Parameter Name	Unit of Measure
1.	Zmax	Maximum reflectivity	dBZ
2.	H_Zmax	Heights of the maximum reflectivity level	Km
3.	dH45	Height of the 45dBZ echo above the environmental melting level	Km
4.	VIL	Vertically integrated liquid	kg/m2
5.	H35	35 dBZ Storm Echo top	Km
6.	HKE	Hail kinetic energy	J/m2

), which are commonly applied in radar-based hail diagnostic studies. The storm echo top parameter represents the vertical extent of the cloud. Previous studies have demonstrated its effectiveness in diagnosing hail within clouds using various reflectivity thresholds, including 7, 12, 15, 18, 20, 30, and 35 dBZ [5,51,52,53]. In the present study, the echo top at 35 dBZ (H35), corresponding to moderate precipitation, was employed. This threshold was chosen because the height it reaches serves as a robust proxy for the strength of the storm’s updraft, which is a key mechanism for transporting precipitation mass to freezing levels essential for hail growth.

Radar reflectivity values exceeding 55 dBZ generally indicate the presence of hydrometeors larger than 5 mm in diameter, often associated with hail [54]. However, this relationship is modulated by factors such as hydrometeor type, density, and radar system characteristics [55]. Although high reflectivity typically signals hail, it may also result from dense rainfall or melting ice particles, underscoring the need for careful interpretation in severe storm analyses [56]. The height of maximum reflectivity within a storm (H_Zmax) is strongly linked to hail occurrence. Larger maximum reflectivity heights are associated with intense updrafts capable of suspending large hailstones. Above specific thresholds, this parameter reliably indicates the presence of hail [54,57].

The height of the 45 dBZ radar echo above the environmental melting level is another critical metric for assessing hail potential. When the 45 dBZ echo extends more than 4 km above the melting level, it indicates robust updrafts that support hail formation and growth above the freezing level [40,56]. This metric is widely used operationally to enhance hail detection, particularly in regions prone to severe convective storms [40,50,56,57,58,59].

The VIL (Vertically Integrated Liquid) parameter and its derivatives are also reliable indicators of hail presence [5,39,57]. Together, these parameters provide a robust framework for diagnosing and monitoring hail potential in convective storms. Hail Kinetic Energy (HKE) is a parameter derived from radar reflectivity and is used to evaluate the destructive potential of hail. In this study, hail kinetic energy was expressed in J/m², following the equation proposed by Abshaev et al. (2014). [49] HKE was computed using the ASU-MRL software 2022 version, designed for monitoring active atmospheric influences [57], by applying a reflectivity threshold of 55 dBZ, which is commonly associated with hail occurrence in radar-based assessments:

$$HKE={\sum }_{m=1}^{\infty }{10}^{(0.086Z_{10}-5.25)}∆t_m$$

where Z₁₀—maximum radar reflectivity on 10 cm wavelength; $∆t$—time step; Sum—time integral.

The ASU-MRL system is an automated radar processing suite designed to detect and characterize hail-producing convective clouds. It analyzes key radar parameters to identify storm cells with a high probability of hail formation. The software automatically detects and tracks convective cells by grouping radar echo points into coherent structures, then computes their movement, velocity, intensity, and hazard level. The system generates three-dimensional reflectivity charts that depict storm structure and precipitation intensity at multiple altitudes, supporting both hail risk assessment and anti-hail rocket operations [49,50,57]. The algorithms integrated within ASU-MRL facilitate the identification of regions with a high likelihood of hail formation and enable the estimation of hail characteristics, including size and intensity [58,60]. We used the radar-derived hail kinetic energy mask at HKE > 300 J m⁻² to delimit the potentially affected area prior to Sentinel-2 NDVI analysis. The HKE metric is a well-established proxy for hail damaging potential [42,45] and isolines around 300 J m⁻² have been applied in regional and operational studies to identify broad hazard swaths [46,61]. We note, however, that the relationship between HKE and damage depends on radar band, storm microphysics and local exposure, so regional calibration against hailpad/ground reports is desirable where available. In practice, the 300 J m⁻² contour was chosen as a pragmatic mask because this value is strongly associated with the onset of severe crop damage in the literature [42,43,47] and, for the Romanian domain, is specifically linked to hail diameters greater than 300 J m⁻² [5]. This threshold effectively reduces the optical search area while retaining both moderate and severe impacts; higher thresholds (e.g., 400–600 J m⁻²) can be used to isolate the most severe cores.

2.2. Sentinel 2 Satellite Data

From each image, the Normalized Difference Vegetation Index (NDVI) was computed using the red (B4) and near-infrared (B8) bands, according to the standard formula:

$$NDVI={\displaystyle \frac{B8-B4}{B8+B4}}$$

The difference between pre- and post-event NDVI was then calculated to obtain ΔNDVI values:

$$\Delta NDVI=NDV{I}_{pre}-NDV{I}_{post}$$

The resulting spatial patterns highlighted extended areas with significant reductions in NDVI values. The irregular geometry of these affected patches, superimposed over the otherwise regular shapes of crop parcels, provided the first evidence that the hail impact was detectable from satellite imagery. Such irregular patterns cannot be attributed to crop harvest, since harvested fields typically display uniform, regular boundaries, whereas in the affected parcels only portions of the fields showed decreased NDVI while adjacent portions remained intact. Within the damaged areas, multiple levels of NDVI reduction were observed, reflecting different intensities of hail impact.

For the optical analysis, we used Sentinel-2 multispectral imagery (Level-1C products) obtained from the Copernicus Data Space Browser [62], specifically the red (B4, 665 nm) and near-infrared (B8, 842 nm) bands, which were employed to compute the Normalized NDVI [62]. Sentinel-2 provides systematic global coverage at 10 m spatial resolution for the visible and near-infrared bands, with a revisit time of 5 days. These characteristics make Sentinel-2 particularly suitable for detecting localized vegetation disturbances, such as those caused by hailstorms, where fine spatial detail is essential [63,64]. For both case studies, only image pairs fulfilling two conditions were selected:

(i) Minimal time gap between pre- and post-event acquisitions, with the first post-event image being acquired between 5 and 10 days after the event. This minimal gap captures the initial, acute vegetative impact, while subsequent images were used to analyze the damage’s temporal evolution.

(ii) Cloud cover below 5% over the hail-affected area. While the surrounding areas sometimes exhibited cloud fractions of 10–12%, pixels inside the hail swath were cloud-free.

No additional atmospheric corrections were applied to the L1C images, since the analysis focused on relative differences rather than absolute reflectance. From each image, the NDVI was computed using the red (B4) and near-infrared (B8) bands. Subsequently, ΔNDVI maps were derived as the difference between pre- and post-event indices. Two ΔNDVI thresholds (0.10 and 0.20) were applied to differentiate between moderate and severe canopy damage, consistent with values reported in previous hail impact studies [37,48] and confirmed by a preliminary sensitivity analysis performed for the study area. The optical analysis was spatially constrained to the hail kinetic energy (HKE) > 300 J m⁻² mask derived from radar, ensuring consistency between radar-detected hail hazard and Sentinel-based impact mapping.

2.3. Land Use Classification and Statistical Analysis

In order to evaluate the sensitivity of different land use classes to hail damage, we derived a land cover layer from Sentinel-2 imagery. Eight main categories were identified: built-up areas, arable land, complex agriculture, pastures, orchards, forests, water bodies and wetlands, and unused land. For the hail-affected areas, the analysis focused on the five dominant categories relevant to vegetation response: arable land, complex agriculture, orchards, pastures, and forests.

The ΔNDVI maps were thresholded at values >0.10 and >0.20 to represent moderate and severe canopy damage, respectively. Binary rasters with a spatial resolution of 10 m (corresponding to the Sentinel-2 pixel size) were generated for each threshold. These rasters were then intersected with the land-use parcel polygons extracted from Sentinel-2–based classification to compute the mean ΔNDVI and the proportion of affected area per parcel. For the Dolhasca case study, the analysis was conducted across three post-event time points (5, 8, and 15 days) to evaluate temporal variations in vegetation response. It is acknowledged that Sentinel-2 imagery may include small geolocation and ground-projection uncertainties, typically within ±10–12 m at Level-1C processing [62]. However, given the spatial resolution of the dataset (10 m) and the parcel-level scale of analysis, these positional errors are considered negligible for the purposes of ΔNDVI comparison and intersection with land-use polygons.

To test the statistical significance of temporal differences across the three post-event intervals, non-parametric Friedman tests were applied within each land use category. Post hoc Wilcoxon signed-rank tests were then used to identify significant pairs of intervals (e.g., 5d > 8d, 8d > 15d). This approach enabled the detection of both the overall temporal attenuation of the hail signal and the differential sensitivity of vegetation types to hail damage (Figure 2).

3. Results

3.1. Case Study from 17 July 2016—The Radauti Area

The first case study is that from 17 July 2016 in the northern part of Suceava county, where two satellite images were available from 16 July 2016 and from 23 July at a distance of only 7 days one from another. Also, the vegetation was in a good state prior to the hailstorm.

3.1.1. Synoptic Drivers of the Hailstorm

The instability on 17 July 2016, was driven by an upper-level depression situated over Romania. The cold air core present in the mid-troposphere gradually shifted eastward. Cut-off lows developed in the middle troposphere represent the most important main synoptic pattern for hail occurrence over northeastern Romania [65]. Convective parameters exhibited moderate to high values during the afternoon hours, highlighting the potential for significant atmospheric convection. The Skew-T diagram analysis reveals a highly favorable environment for severe convection and hailstorm development on 17 July 2016, at 1200 UTC (Figure 3). The moderate to strong Convective Available Potential Energy (CAPE) of 1347 J/kg at the surface and the relatively low Convective Inhibition (CIN) of −5 J/kg suggest significant instability and minimal inhibition to storm development, which is conducive to deep convection [66]. The 0–6 km bulk shear of 23.9 m/s indicates sufficient deep-layer shear to support organized convection and possibly supercellular structures [67]. The melting level at approximately 4.2 km and the presence of a well-defined hail growth layer (HGL) above the freezing level provided a favorable thermodynamic profile for large hail formation. These thermodynamic and kinematic conditions support the development of severe storms with large hail, likely associated with supercells. The storm motion of 13.5 m/s further suggests rapid storm development and potential for damaging hail. These factors align with findings in the literature on the role of instability, shear, and hail growth layers in hailstorm dynamics [67,68].

3.1.2. Radar-Derived Hail Parameters

Based on the synoptic and thermodynamic environment described above, radar observations from the MRL-5 system at 15:04 UTC revealed a highly organized convective cell characterized by intense reflectivity values exceeding 60 dBZ (Figure 4a). The vertical cross-section along the storm axis indicated a deep convective structure extending up to 11 km, with the main reflectivity core located between 3 and 8 km, overlapping the hail growth layer. Such a structure is consistent with strong updrafts and efficient hailstone formation processes.

Complementary ground observations provide direct evidence of the storm’s severity (Figure 4b–d). Eyewitness photographs captured the visual morphology of the convective system, showing a supercell-like organization with a well-defined shelf cloud. Reports and images collected from local observers indicated maximum hailstones diameters reaching 4–5 cm. Field images also documented extensive crop damage, confirming the significant agricultural impact of the event.

Storm tracking analysis revealed that the storm had a total lifespan of 2 h and 40 min, from the initial radar echo to complete dissipation. According to radar parameter values, the period during which the storm posed a high risk of hailfall was between 14:15 and 15:15. During this interval, Z_max remained at values equal to or greater than 60 dBZ, and VIL exceeded 30 kg/m² (Figure 5b). High VIL values are correlated with the likelihood of hail occurrence, although exact thresholds depend on factors such as local climate, freezing level height, and storm dynamics. Studies suggest that VIL > 30 kg/m² is often associated with small hail, while VIL > 40–45 kg/m² indicates a high probability of medium- to large-sized hail [39]. Moreover, VIL > 50 kg/m² has been identified as an indicator of large hail, frequently associated with supercell storms and strong convective structures [40,56]. During this period, H35 exceeded 8 km, signifying vigorous vertical development of the storm. Similarly, dH45 remained above the critical threshold of 4 km, which is associated with hailfall [40,58] (Figure 5c).

A peak in all radar parameters (Z_max > 65 dBZ, H_Zmax > 5 km, VIL > 50 kg/m², dH45 > 5 km) was observed between 15:00 and 15:15. During this interval, reports of large hail, ranging from 2 to 5 cm, were recorded in multiple locations in the ESWD database. Three of these reports correspond to areas with maximum HKE values (Figure 5a). However, it is important to note that the localization of hailfalls is not always precise and often depends on witnesses or media sources from which the reports are collected.

3.1.3. Hailstorm Effects Assessment

Figure 6 illustrates the temporal distribution and quality of Sentinel-2 acquisitions around the hailstorm event of 17 July 2016. White boxes indicate dates without Sentinel-2 imagery, light green boxes correspond to usable acquisitions with less than 5% cloud cover, and dark green boxes represent acquisitions affected by cloud contamination. The red box marks the date of the hail event. For this case, only two images fulfilled both temporal and quality conditions—13 July (pre-event) and 23 July (post-event)—which were subsequently used to derive the ΔNDVI map presented in the following section.

The radar-derived HKE > 300 J/m² contour covered approximately 13,700 ha, while the innermost contour (>600 J/m²) encompassed about 3870 ha. These areas were calculated directly from the vectorized HKE isolines using GIS-based polygon area computation. Similarly, hail-damaged areas extracted from Sentinel-2 ΔNDVI maps were quantified by summing the pixel areas (10 m resolution) exceeding the defined ΔNDVI thresholds (>0.10 and >0.20). Within the high-intensity hail core, ΔNDVI values exceeded 0.15–0.20, indicating substantial canopy loss, whereas peripheral zones (300–400 J/m²) showed smaller reductions (ΔNDVI 0.05–0.10) (Figure 7). Despite the limitation of having only one post-event image, the spatial correspondence between zones of maximum HKE and the strongest NDVI reductions confirms the robustness of combining radar and optical data for hail damage detection.

The hailstorm that occurred on 17 July 2016 in the Rădăuți area generated visible vegetation stress detectable from Sentinel-2 imagery. NDVI values derived from pre-event (13 July) and post-event (23 July) acquisitions showed a consistent decrease across the swath impacted by hail. For this analysis, only the area delimited by hail kinetic energy (HKE) exceeding 300 J/m², as derived from MRL5 radar scans, was considered.

Applying a ΔNDVI threshold of >0.10 indicated that approximately 5856 ha of vegetation were affected within the HKE > 300 J/m² zone. The spatial distribution of this disturbance formed an irregular pattern that overlapped cropland parcels, a characteristic consistent with hail impact rather than with harvesting, which usually follows parcel boundaries. At the more conservative threshold of ΔNDVI > 0.20, the estimated affected area decreased to 2236 ha, concentrated in the central part of the hail swath where radar-derived HKE values were highest (>300 J/m²). This gradation suggests that while a broad area was moderately stressed by hail, severe vegetation damage was restricted to a smaller core region (Figure 8 and

Table 3. Estimated hail-affected vegetation in the Rădăuți study area (17 July 2016) for two ΔNDVI thresholds, expressed as both absolute surface (ha) and as a percentage of the total area encompassed by the radar-derived HKE > 300 J/m² contour.

ΔNDVI Threshold	Affected Pixels	Area (ha)	Percent from Total Area with Radar Defined HKE > 300 J/m2 Zone
>0.10	547,032	5856	40.0
>0.20	223,597	2235	16.3

).

The results support the robustness of NDVI differencing as a proxy for hail damage detection in agricultural landscapes, in line with previous studies that reported ΔNDVI reductions between 0.1 and 0.2 as reliable indicators of crop stress following hail events, e.g., [33,48,69,70]. Even when only one pre- and one post-event image pair are available, Sentinel-2 imagery provides sufficient spectral and spatial detail to discriminate hail-affected vegetation. These findings set a baseline for comparison with the Dolhasca case study, where multiple post-event images allowed a more detailed temporal analysis of vegetation recovery.

3.2. Dolhasca Case Study (30.07.2020)

Apart from this Radauti area, we found another area around the Dolhasca village and the previous scenario was found here, as well, although a smaller affected area but with devastating effect.

3.2.1. Synoptic Drivers of the Hailstorm

Synoptically, this day was characterized by a large-scale circulation pattern over Europe, marked by the extension of a ridge toward the southeastern regions of the continent, associated with a high-pressure area centered west of the Scandinavian Peninsula. An anticyclonic circulation dominated central Europe, while a low-pressure area extended over eastern Scandinavia and western European Russia. The interaction between these baric formations facilitated the advection of cold maritime-polar air toward Romania. The arcuate morphology of the Eastern and Forested Carpathians further enhanced this flow, channeling the cold air southeastward. Under these conditions, the most pronounced cold air advections were recorded across northeastern Romania. In July, such northwesterly upper-level flow aloft is among the primary synoptic patterns associated with severe hail-producing convection in this region [64]. The Skew-T analysis (Figure 9) for 1200 UTC indicates an environment moderately favorable for severe convection, with CAPE = 1665 J/kg and Convective Inhibition (CIN) = −10 J/kg, suggesting moderate instability and a relatively weak convective cap. The 0–6 km bulk shear was 11.7 m/s, sufficient to support multicellular or marginally organized convection, while the 0–8 km shear reached 16.4 m/s. Steep lapse rates (8 K/km in the 700–500 hPa layer) further enhanced the potential for large hail formation. Observations from the ESWD database confirm the occurrence of hail with diameters of 4–5 cm reported at four locations between 18:20 and 18:30 local time, verifying the intensity of the event.

3.2.2. Radar-Derived Hail Parameters

To complement the radar-based quantitative analysis of the 30 July 2020 hailstorm, Figure 10 provides a combined radar–ground verification overview of the convective system responsible for the Dolhasca event. Panel (a) shows the maximum composite reflectivity field alongside the vertical cross-section along the storm’s main axis, highlighting a deep convective core with reflectivities exceeding 60 dBZ, consistent with hail-producing updrafts. Panel (b) presents a photograph captured by local observers, illustrating the storm’s well-defined shelf cloud and the associated outflow structure characteristic of intense convective systems in their mature stage. Panels (c) and (d) document ground evidence of the event: 4–5 cm hailstones and impact marks on buildings and vehicles, confirming the occurrence and severity of the hailfall within the radar-identified swath. Together, these visual and radar-based observations strengthen the interpretation of storm intensity and provide independent verification of the hail-producing nature of the convective cell.

At 17:10, the first radar echo of the storm was recorded approximately 50 km northwest of the affected area. The total lifespan of the storm was approximately 95 min. The graph in Figure 11b depicts the temporal evolution of radar-derived parameters Zmax and VIL, both critical for assessing convective storm severity and hail potential. Between 17:16 and approximately 18:40 UTC, both Zmax and VIL demonstrate a progressive increase, with Zmax exceeding 60 dBZ at its peak, indicating the presence of high-density hydrometeors such as hailstones. Simultaneously, VIL attains a maximum value of approximately 50 kg/m², which is strongly correlated with the production of significant hail [40]. These trends indicate the intensification of the convective system, characterized by vigorous updrafts and substantial hydrometeor growth. Following the peak, both Zmax and VIL decline steadily, signaling a transition to the storm’s dissipative stage as updraft intensity weakens and hail growth diminishes.

Figure 10c provides insight into the vertical structure of the storm by illustrating the evolution of three key height parameters: H_Zmax, dH45 and H35. All three parameters exhibit a significant rise during the storm’s intensification phase, peaking around 18:40 UTC. The H_Zmax reaches approximately 8 km, indicative of strong updrafts capable of lofting large hailstones into higher altitudes. The dH45 peaks at approximately 4 km above the melting level, exceeding the threshold typically associated with severe hailstorms. Concurrently, the H35 reaches a maximum height of about 13 km, reflecting the storm’s robust vertical development and the potential for severe convection. After the peak, all height parameters decline, marking the storm’s weakening phase. The synchronization of these height and intensity metrics strongly supports the presence of a mature convective storm with severe hail potential during its maximum development.

The kinetic energy map indicates a well-defined area with values exceeding 300 J/kg along the storm’s path. The region with HKE values above this threshold begins 20 km northwest of the location of the first report of large hail (Figure 11a).

3.2.3. Hailstorm Effects Assessment

For the Dolhasca case study, suitable Sentinel-2 images were available at three post-event intervals: 5, 8, and 15 days after the hailstorm (Figure 12).

Figure 13 shows the overlap between radar-derived hail kinetic energy (HKE) contours and vegetation response expressed as ΔNDVI for the Dolhasca hailstorm of 30 July 2020. The HKE > 300 J/m² contour delineates a potentially affected area of ~7626 ha, while the innermost contour (>600 J/m²) corresponds to ~986 ha. Within this core, the majority of pixels display ΔNDVI > 0.2, indicating severe vegetation damage. By contrast, in the outer HKE zones (300–400 J/m²), large areas show ΔNDVI values below 0.1, reflecting little or no impact on vegetation. These results suggest that while the 300 J/m² threshold is effective for quickly flagging hazard-affected regions, only part of this area corresponds to actual vegetation damage observable in Sentinel-2 imagery. The spatial concordance between the higher HKE contours and strong ΔNDVI values confirms the utility of integrating radar and optical data for robust hail damage assessment.

The spatial patterns of vegetation affected by hail are shown in Figure 14 for two ΔNDVI thresholds (>0.10 and >0.20) at 5, 8, and 15 days after the event, within the radar-derived mask of HKE > 300 J m⁻². At all temporal intervals, patches of decreased NDVI were concentrated along the central and northern parts of the study area, consistent with the radar-indicated hail swath. The extent of affected vegetation, however, varied substantially with both the temporal distance from the storm and the applied ΔNDVI threshold.

At 5 days post-event, the affected area was the largest, covering ~2581 ha when using the ΔNDVI > 0.10 criterion, which corresponds to ~33.8% of the total area considered (

Table 4. Summary statistics of estimated hail-affected vegetation area for different ΔNDVI thresholds and temporal intervals post-event.

Post Event Interval	ΔNDVI Threshold	Affected Pixels	Area (ha)	Percent from Total Area	Temporal Sensitivity Index (%)
5 days	>0.10	258,054	2580.5	33.8	65.3
	>0.20	89,536	895.3	11.7	–
8 days	>0.10	189,083	1890.8	24.7	65.2
	>0.20	65,643	656.4	8.6	–
15 days	>0.10	151,053	1510.5	19.8	62.9
	>0.20	56,017	560.1	7.3	–

). A stricter threshold of ΔNDVI > 0.20 reduced the affected area to ~895 ha (~11.7%). By 8 days post-event, the estimated affected area declined to ~1891 ha for ΔNDVI > 0.10 (24.7%), and ~656 ha for ΔNDVI > 0.20 (8.6%). At 15 days, the affected area further decreased to ~1511 ha (19.8%) and ~560 ha (7.3%) for the two thresholds, respectively.

Figure 15 shows the spatial distribution of land use categories within the Dolhasca study area, overlaid with radar-derived hail kinetic energy (HKE) contours (>300 J/m²). The land use pattern is dominated by arable land and complex agriculture, interspersed with patches of pastures, orchards, and forested areas, while built-up and industrial areas are concentrated along the main settlements. The overlap between the radar-indicated hail swath and these land use classes highlights that agricultural areas were most exposed to hail impacts, with arable fields covering the largest share of the affected zone.

This spatial context provides the basis for understanding how different land use categories responded to hail damage. By combining radar-derived HKE contours with Sentinel-2 NDVI differences, it becomes possible to evaluate not only the overall extent of affected vegetation but also the varying sensitivity and recovery potential of specific land use types. In particular, arable land, complex agriculture, orchards, pastures, and forests exhibited distinct ΔNDVI trajectories across the three post-event intervals, reflecting heterogeneous responses to the hailstorm impact.

Figure 16 illustrates the progressive decline of affected area across the three-time steps. The largest difference between consecutive intervals was observed between 5 and 8 days, with a reduction of ~65% relative to the initial estimate for ΔNDVI > 0.10. A similar sensitivity index was obtained for the 8–15-day comparison (~63%). This demonstrates that the vegetation response to hail is most detectable shortly after the event, with a marked attenuation of the NDVI signal over time.

Notably, while the overall affected area decreased with time, the spatial patterns remained consistent: areas initially flagged as damaged continued to display reduced NDVI, though the magnitude of change diminished. This pattern suggests that part of the vegetation recovered or that canopy reflectance stabilized after initial defoliation and tissue injury. The differences between the ΔNDVI > 0.10 and >0.20 thresholds highlight the trade-off between sensitivity (capturing a larger, possibly noisier, affected area) and specificity (focusing on the most severely damaged patches).

Vegetation damage as measured by ΔNDVI within the HKE > 300 J/m² mask shows a clear decline over time after the hail event, with meaningful differences across land use classes (

Table 5. Mean ΔNDVI values (±SD) for different land use classes within the HKE > 300 J/m² area at 5, 8, and 15 days post-event, with results of Friedman and Wilcoxon tests.

Land Use Class	ΔNDVI (5 Days)	ΔNDVI (8 Days)	ΔNDVI (15 Days)	Friedman p-Value	Wilcoxon Post hoc (Significant Pairs)
Arable	0.11 ± 0.09	0.1 ± 0.09	0.07 ± 0.09	<0.001	5 d > 8 d; 8 d > 15 d
Complex agriculture	0.11 ± 0.07	0.1 ± 0.08	0.07 ± 0.08	<0.001	5 d > 8 d; 8 d > 15 d
Orchards	0.09 ± 0.05	0.05 ± 0.04	0.02 ± 0.05	<0.001	5 d > 8 d; 8 d > 15 d
Pastures	0.07 ± 0.06	0.05 ± 0.05	0.02 ± 0.05	<0.001	5 d > 8 d; 8 d > 15 d
Forests	0.08 ± 0.06	0.07 ± 0.06	0.05 ± 0.06	<0.001	5 d > 8 d; 8 d > 15 d

, Figure 17). Across all classes, the strongest damage signal was recorded at 5 days post-event, gradually decreasing at 8 and 15 days. This temporal attenuation is consistent with previous studies reporting that hail-induced vegetation stress is best detected in the immediate days following the event, with the spectral signal diminishing as plants either recover or stabilize [37].

For arable fields, mean ΔNDVI values were highest at 5 days, significantly greater than at 8 days, and 8 days greater than at 15 days (Friedman test p < 0.001; Wilcoxon: 5d > 8d; 8d > 15d). This indicates that arable crops are the most sensitive to hail impacts, consistent with results reported by Ha et al. (2022) [37] for field crops (canola, wheat, lentil), where ΔNDVI profiles showed marked declines in the first days following hail events. Complex agriculture and orchards also showed significant temporal decreases (p < 0.01), though of lower magnitude than arable lands. Pastures and forests exhibited smaller ΔNDVI changes, with weaker but still significant declines in some cases.

Boxplots (Figure 16) provide a detailed view of the variability and temporal dynamics of ΔNDVI values across different land use categories following the hail event. For arable land, the highest median ΔNDVI values were recorded at 5 days post-event, with numerous parcels exceeding ΔNDVI > 0.20, reflecting severe canopy loss and strong hail sensitivity. By 15 days, the median values decreased substantially, though several outliers remained, indicating localized persistence of damage in some plots. A similar pattern was observed for complex agriculture, where the overall ΔNDVI signal was slightly weaker but still showed a significant stepwise decline between 5, 8, and 15 days.

In orchards, ΔNDVI values displayed a narrower range and lower medians compared to arable land, suggesting that perennial crops, while affected, are less immediately responsive to canopy damage in spectral terms. Pastures exhibited moderate declines in NDVI, with broader variability and scattered outliers, possibly reflecting heterogeneous ground cover and partial recovery dynamics within short intervals. Finally, forests showed the lowest ΔNDVI values overall, with compact distributions around near-zero values. This pattern may reflect both the higher ecological resilience of tree canopies and the reduced detectability of damage due to structural complexity and dense vegetation layers [69].

The temporal decrease across all categories reinforces the observation that hail-induced vegetation stress is most visible in the immediate aftermath of the storm, with the NDVI signal progressively attenuating as vegetation begins to recover or as canopy reflectance stabilizes. These results align with previous work, showing that the detection of hail damage through multispectral indices is most reliable within the first week post-event [37,48].

4. Discussion

4.1. Radar as a Temporal Hail Indicator

Radar-derived parameters provide one of the earliest and most reliable tools for identifying hail-producing storms and outlining potential damage areas. Among these, hail kinetic energy (HKE) is a robust diagnostic for assessing the likelihood of crop-damaging hail. The 300 J/m² threshold adopted in this study aligns with previous applications in hail climatology and event-based analyses, where it marks storm cells with a high probability of agricultural damage [39,40,56].

Unlike optical satellite data, which are limited by acquisition timing and cloud cover, weather radar offers continuous temporal monitoring of storm evolution. In both case studies, periods of maximum reflectivity (Zmax > 60 dBZ) and high vertically integrated liquid water (VIL > 30–40 kg/m²) closely matched eyewitness reports of large hail. These associations are consistent with previously reported thresholds linking VIL and reflectivity to hail occurrence and size [5,39,40].

Vertical structure metrics such as the height of the 35 dBZ echo (H35) and the depth above the 45 dBZ level (dH45) further confirmed the presence of severe hail when exceeding critical values (H35 > 8–9 km; dH45 > 4 km) [39,40,58]. The simultaneous occurrence of these elevated radar signatures and strong HKE cores in both Rădăuți and Dolhasca indicates sustained hail formation throughout the storm lifecycle.

Therefore, radar serves not only as an indicator of hail likelihood but also as a temporal and spatial filter guiding post-event optical analysis. Radar-derived HKE effectively delineates the zones and timing of hail occurrence, enabling Sentinel-2 NDVI differencing to focus precisely on areas where vegetation impacts are most likely.

4.2. Sentinel-2 NDVI and the Critical Role of Timing

The results from both case studies highlight the critical role of timing in detecting hail-induced vegetation stress using Sentinel-2 imagery. NDVI differences were most pronounced in the immediate days following the hail events, with a progressive attenuation of the spectral signal over time. In Dolhasca, the largest affected area was detected 5 days post-event, while at 15 days the extent of disturbance had significantly decreased. This pattern suggests that canopy defoliation and tissue damage induced by hail are best captured shortly after impact, before recovery processes, senescence, or noise from other agricultural practices (e.g., management operations, phenological changes) diminish the detectability of the signal.

These findings are consistent with previous studies reporting that hail-induced reductions in NDVI are strongest when imagery is acquired within one week after the event [37,69]. By contrast, when the temporal gap increases, vegetation response tends to stabilize, either due to partial recovery or because plant canopies reach new reflectance equilibria [48]. The Rădăuți case further illustrates this dependency: only one usable post-event image was available at 6 days after the storm, and although strong reductions in NDVI were evident, the lack of additional temporal data limited the capacity to assess vegetation recovery dynamics.

The practical implication is that the temporal availability of Sentinel-2 imagery is as important as its spatial resolution. Even with the high spatial detail of 10 m pixels, delayed post-event acquisitions may underestimate the extent of hail damage. This underlines the importance of exploiting all cloud-free Sentinel-2 acquisitions in the days immediately following hailstorms, and, where available, integrating them with data from other missions such as Landsat-8/9 or commercial constellations to reduce the temporal gap [63,64].

Thus, Sentinel-2 is not only a powerful tool for spatially detailed mapping of hail impacts, but its effectiveness depends fundamentally on the timing of acquisitions relative to the storm. This study extends previous approaches by demonstrating, through two distinct case studies, that short-term monitoring windows (<7 days post-event) maximize the detectability of hail damage and improve the accuracy of damage estimation across different vegetation types.

4.3. Land Use Sensitivity and Vegetation Resilience

For arable fields, mean ΔNDVI values were highest at 5 days post-event, significantly greater than at 8 and 15 days (Friedman test p < 0.001; Wilcoxon: 5d > 8d; 8d > 15d). This rapid sensitivity agrees with Ha et al. (2022) [37], who found sharp NDVI declines in crops such as canola and wheat within the first week after hailstorms, followed by gradual recovery. Angearu et al. (2022) [48] similarly detected ~3143 ha of damaged cropland in the Bărăgan Plain using Sentinel-2 imagery (NDVI ≥ 0.05), with reflectivity >55 dBZ spatially matching NDVI losses—supporting the reliability of radar-based hail indicators.

Complex agriculture and orchards showed a comparable but attenuated NDVI decline, reflecting heterogeneous vegetation and mixed canopy structures. Pastures recorded smaller reductions, yet still displayed immediate damage followed by partial regrowth. The initial sharp decline (3–6 August 2020) suggests canopy defoliation, while partial recovery by mid-August reflects regrowth, consistent with rapid grassland recovery after disturbance [71,72].

Forests exhibited the weakest but detectable NDVI response. Pre-hail imagery (29 July 2020) indicated dense canopies, whereas post-hail reductions (3–6 August) point to leaf loss and branch breakage. The slight increase by 13 August suggests early recovery, though NDVI values remained below pre-event levels—typical of slower forest regeneration following extreme weather [37,70,73,74].

Overall, NDVI differencing revealed clear contrasts in hail sensitivity and resilience: arable fields were most responsive, orchards and pastures moderately affected, and forests showed buffered but longer-lasting impacts. This gradient reflects ecosystem structure and function, emphasizing the value of integrating land-use data with multispectral indices to evaluate hailstorm impacts across vegetation types.

4.4. Methodological Novelty, Limitations and Implications for Future Studies

This study demonstrates the potential of integrating radar-derived Hail Kinetic Energy (HKE) fields with Sentinel-2 NDVI differencing for robust hail damage assessment. The main contribution lies in applying radar-based thresholds (HKE > 300 J/m²) as a rapid temporal indicator of hail occurrence, allowing early delineation of potentially affected areas. Sentinel-2 imagery provided the spatial refinement needed to map damage at parcel scale and to evaluate the temporal attenuation of canopy stress. Integrating land-use data enabled comparison of sensitivity and resilience among vegetation types, thereby extending previous research on hailstorm impacts.

Despite these strengths, several limitations must be acknowledged. The temporal availability of Sentinel-2 imagery is constrained by orbital cycles and cloud cover. In the Rădăuți case, only one post-event image limited recovery analysis, while in Dolhasca, cloud-free acquisitions at 5, 8, and 15 days post-event improved temporal understanding but still missed the critical <3-day window [37]. The use of Level-1C imagery without atmospheric correction may introduce reflectance uncertainties, though the short temporal intervals and use of relative ΔNDVI values minimize their effect [75]. Land-use data, derived from Sentinel-2, covered broad categories (arable, orchards, pastures) without crop-specific or phenological information, which affects sensitivity estimates [10,76].

The reliance on radar HKE as a spatial mask also introduces scale-related uncertainty. While the 300 J/m² threshold is supported by prior studies [46,61,77,78], the 500 m radar resolution can overestimate affected areas. Nevertheless, the strong spatial correspondence between HKE contours, NDVI reductions, and eyewitness hail reports supports the method’s validity.

A key limitation remains the lack of direct ground validation. Although ESWD observations and remote-sensing consistency support our interpretation, field-based crop loss data would provide more rigorous confirmation. Future work should integrate in situ surveys, agricultural insurance records, or UAV imagery to validate affected areas and calibrate ΔNDVI thresholds by crop type and phenological stage.

Looking ahead, integrating Sentinel-2 with other high-frequency missions (PlanetScope, Landsat-8/9) could reduce temporal gaps and enhance recovery tracking. Sentinel-1 radar backscatter offers a valuable complement under cloudy conditions. Incorporating crop-specific and phenological data would improve vulnerability modeling, while machine learning frameworks combining radar, NDVI trajectories, and land-use data could support automated, operational hail damage assessment.

5. Conclusions

This study demonstrates the effectiveness of integrating weather radar and Sentinel-2 multispectral imagery to assess hail-induced vegetation damage in northeastern Romania. Radar-derived Hail Kinetic Energy (HKE) fields provided a reliable temporal and spatial indicator of hail occurrence, while Sentinel-2 NDVI differencing quantified vegetation stress at parcel level. The combined use of these complementary datasets proved powerful for delineating affected areas and analyzing the temporal dynamics of vegetation response.

The two case studies, Rădăuți (2016) and Dolhasca (2020), yielded two main insights. First, an HKE threshold of 300 J m⁻² effectively identifies hail swaths with a high likelihood of crop damage, aligning with previous international findings. Second, NDVI reductions were strongest within the first week post-event, highlighting the critical role of timing in post-storm satellite acquisitions. In Dolhasca, the decline in the affected area between 5, 8, and 15 days post-event reflects both crop sensitivity and the temporal attenuation of the spectral signal.

Analysis by land-use category revealed differentiated responses: arable fields and complex agriculture were most sensitive, orchards and pastures showed moderate impacts with partial recovery, while forests exhibited the weakest yet measurable declines, indicating structural resilience. These gradients emphasize the value of integrating land-use data with spectral indices to assess both damage extent and recovery dynamics.

Methodologically, the study advances a time-sensitive radar–optical framework for hail damage assessment. Limitations include the restricted temporal availability of cloud-free Sentinel-2 imagery, the absence of atmospheric correction, and the use of generalized land-use classes. Future research should incorporate higher-frequency sensors (e.g., PlanetScope, Landsat-8/9, Sentinel-1 SAR), crop-specific and phenological data, and field-based validation to refine vulnerability assessment and improve model calibration.

Overall, these results confirm that multi-sensor approaches enhance both the accuracy and timeliness of hail damage detection. The workflow presented here provides a scalable methodology for monitoring the impacts of severe convective storms on crops and ecosystems, applicable to hail-prone regions worldwide under a changing climate.