A Multi-Metric NDVI-Derived Dataset of Vegetation Dynamics and Land Surface Phenology in Southern and Central Europe (1982–2022)

Abstract

This Data Descriptor presents a value-added suite of derived vegetation phenology products for Southern and Central Europe (10° W–28° E; 35° N–50° N), generated from the PKU GIMMS NDVI v1.2 archive (1982–2022) through a standardized processing workflow including quality screening, temporal compositing, phenological metric extraction, eco-phenological regionalization, and variability analysis. The final collection includes 82 GeoTIFF raster layers and one CSV file, structured into long-term monthly NDVI climatologies, decadal NDVI composites, eco-phenological regionalization, Land Surface Phenology (LSP) metrics with inter-decadal shift layers, and a Phenology Variability Index (PVI). In addition, cluster-level Mann–Kendall and Theil–Sen trend statistics are provided in tabular form. All products are distributed in WGS84 (EPSG:4326) at 0.0833° spatial resolution. The dataset provides ready-to-use vegetation phenology indicators, variability metrics, and spatially consistent climatological products, supporting applications in ecosystem monitoring, climate impact assessment, biodiversity studies, and large-scale environmental analysis across European bioclimatic regions.

Dataset: https://doi.org/10.5281/zenodo.19049310

Dataset License: CC-BY 4.0 (Creative Commons Attribution 4.0 International)

1. Summary

This Data Descriptor presents an integrated suite of ready-to-use vegetation phenology products for Southern and Central Europe (10° W–28° E; 35° N–50° N) covering the period 1982–2022. The dataset complements the methodological study by Samela et al. [1], where the processing workflow and ecological interpretation of the derived products are presented in detail, while this Data Descriptor focuses on the structure, content, technical characteristics, and reuse potential of the openly available dataset deposited in Zenodo [2].

It consists of newly generated geospatial products derived from the PKU GIMMS NDVI v1.2 archive (Peking University Global Inventory Monitoring and Modelling Studies Normalized Difference Vegetation Index, version 1.2) [3] through a standardized processing workflow including quality-control filtering, spatial clipping, temporal aggregation, monthly compositing using the Maximum Value Composite (MVC) method, phenological metric extraction, eco-phenological regionalization, and variability analysis. The resulting products provide a spatially and temporally consistent representation of long-term vegetation dynamics across European ecosystems.

The archive contains 82 GeoTIFF raster files and one CSV table, organized into six main folders as described below: (i) long-term monthly NDVI climatologies; (ii) decadal monthly NDVI climatologies for four multi-year periods; (iii) an eco-phenological regionalization of Southern and Central Europe derived from K-means clustering of long-term NDVI seasonal profiles; (iv) Land Surface Phenology (LSP) metrics, including Start, Peak, End, and Length of Season, together with their inter-decadal shift maps; (v) a Phenology Variability Index (PVI) describing the temporal stability of phenological timing; and (vi) cluster-level Mann–Kendall and Theil–Sen trend statistics provided as complementary analytical outputs. All raster products share the same spatial extent, coordinate reference system (WGS84, EPSG:4326), spatial resolution (0.0833°), and GeoTIFF format, ensuring full interoperability across the dataset.

The dataset supports a wide range of applications including vegetation monitoring, phenological analysis, ecosystem assessment, biodiversity and climate-change studies, regional environmental modelling, and spatial stratification of ecological processes.

By publicly releasing these processed geospatial products, this work enables researchers to directly exploit long-term vegetation phenology information without repeating the complete pre-processing, quality-control, compositing, clustering, and phenological metric extraction workflow required to generate these products from the original PKU GIMMS NDVI archive [3]. Among the released products, the Phenology Variability Index (PVI), first introduced by Samela et al. (2026) [1], provides a standardized measure of interannual phenological stability that complements conventional Land Surface Phenology metrics.

2. Data Description

The dataset is provided as a compressed archive named European_NDVI-derived_Phenology_Dataset_1982_2022.zip. The collection consists of 82 GeoTIFF raster files and one CSV table. All spatial raster products share the same coordinate reference system (WGS84, EPSG:4326) and spatial resolution (0.0833°) and can be directly opened in standard GIS software (e.g., QGIS, ArcGIS) or processed using geospatial libraries in Python or MATLAB. The dataset is fully spatially aligned and consistent across all components. A README.md file is included at the root level of the archive, providing a comprehensive guide to the dataset structure, file nomenclature, and technical specifications. The dataset structure is summarized in

Table 1. Summary of the dataset components and technical specifications.

Dataset Component	Format	Spatial Resolution	CRS	Temporal Coverage	Description
01_Long_term_Monthly_Mean_NDVI_1982_2022	GeoTIFF	0.0833° (~9 km)	EPSG:4326	1982–2022	12 long-term monthly NDVI composites.
02_Decadal_Monthly_Mean_NDVI	GeoTIFF	0.0833° (~9 km)	EPSG:4326	Four decades (1982–2022)	48 monthly NDVI composites divided into four decades (1982–1992; 1993–2002; 2003–2012; 2013–2022).
03_Phenological_Clusters_Map	GeoTIFF	0.0833° (~9 km)	EPSG:4326	1982–2022	Categorical map (K = 5) of phenologically coherent regions identified via K-means.
04_Land_Surface_Phenology_Metrics_and_Shifts	GeoTIFF	0.0833° (~9 km)	EPSG:4326	Four decades (1982–2022)	16 decadal metrics (SOS, POS, EOS, LOS) and 4 inter-decadal shift maps (Delta).
05_Phenology_Variability_Index_PVI	GeoTIFF	0.0833° (~9 km)	EPSG:4326	1983–2022	Pixel-wise index [0–1) describing the interannual variability of phenological timing.
06_Mann_Kendall_Theil_Sen_Trend_Analysis	CSV	N/A	N/A	1982–2022	Cluster-level Mann–Kendall and Theil–Sen statistics for monthly NDVI trends.

and organized into six main folders to ensure consistency with the documentation and ease of navigation. Each folder is hereafter referred to as a Component and numbered accordingly (Component 01 through Component 06).

• 01_Long_term_Monthly_Mean_NDVI_1982_2022/
  This component contains 12 GeoTIFF files representing long-term monthly mean NDVI climatologies computed over the entire 1982–2022 period.
  Contents: 12 GeoTIFFs (one per month).
  Values: NDVI (dimensionless).
  Naming convention: Long_term_NDVI_monthlymean_1982_2022_MM.tif, where MM indicates the month from 01 to 12.
• 02_Decadal_Monthly_Mean_NDVI/
  Contents: 48 GeoTIFFs organized in four decadal sub-folders:
  NDVI_monthlymean_1982_1992/,
  NDVI_monthlymean_1993_2002/,
  NDVI_monthlymean_2003_2012/,
  and NDVI_monthlymean_2013_2022/.
  Each period includes 12 monthly composites derived from temporally averaged NDVI values.
  Values: NDVI (dimensionless).
  Naming convention: NDVI_monthlymean_YYYY_YYYY_MM.tif, where YYYY_YYYY indicates the multi-year period and MM indicates the month.
• 03_Phenological_Clusters_Map/
  A single GeoTIFF raster mapping the five eco-phenological regions characterized by similar seasonal NDVI dynamics derived from K-means clustering of long-term NDVI seasonal profiles.
  Contents: 1 GeoTIFF
  Name: NDVI_phenological_clusters_k5_Europe.tif
  Values: Categorical integers from 1 to 5, representing eco-phenological regions.
• 04_Land_Surface_Phenology_Metrics_and_Shifts/

  -

  4.1_LSP_Metrics:

  This sub-component includes 16 GeoTIFF layers representing phenological metrics computed for four decadal periods. Metrics include: Start of Season (SOS), Peak of Season (POS), End of Season (EOS), Length of Season (LOS).
  Content: 16 GeoTIFFs (SOS, POS, EOS, LOS for the four decadal periods).
  Naming convention: GIMMS_NDVI_<METRIC>_<THRESHOLD>_<YYYY_YYYY>_Europe.tif.
  SOS, POS, and EOS values represent month of the year from 1 to 12.
  LOS values represent the number of months.
  Thresholds: 20% for SOS; 30% for EOS/LOS. No threshold suffix is included for POS, as it corresponds to the annual maximum NDVI.

  -

  4.2_Phenological_Shifts:

  Content: 4 GeoTIFF files showing inter-decadal differences between the first decade, 1982–1992, and the last decade, 2013–2022. These layers quantify long-term shifts in phenological timing.
  Values: signed monthly differences (−6 to +5 months)
  Naming convention: GIMMS_NDVI_delta_<METRIC>_<THRESHOLD>_2013_2022_minus_1982_1992_Europe.tif.
  Values represent the signed monthly difference (ranging from −6 to +5 months), computed using a cyclic difference approach to correctly account for transitions across the calendar-year boundary.
• 05_Phenology_Variability_Index_PVI/
  The Phenology Variability Index (PVI) is derived from variability in SOS, POS, and EOS across ten four-year intervals.
  Contents: 1 GeoTIFF
  Name: NDVI_Phenology_Variability_Index_PVI_Europe.tif
  Values: Dimensionless index in the range [0–1). Values ≥ 0.5 indicate higher phenological variability. A representative spatial distribution of the PVI across the study area is shown in Figure 1.
• 06_Mann_Kendall_Theil_Sen_Trend_Analysis/
  Contents: 1 CSV file following RFC 4180 standard.
  Name: MK_TheilSen_results.csv
  Columns: Cluster (C1–C5), Month (Jan–Dec), SenSlope (NDVI units/year), p_value (significance), N (observations).
  Rows: The table contains a header row of column names followed by 60 data rows, corresponding to all combinations of the five eco-phenological clusters (C1–C5) and the twelve calendar months (January–December).
  This table provides a complete statistical characterization of monthly NDVI trends aggregated at cluster level.
• Root Files:
  README.md: Comprehensive guide to the dataset structure, file content, nomenclature, and technical specifications.

3. Methods

The dataset was generated using the PKU GIMMS NDVI v1.2 (1982–2022) [3] as the primary input data source. The original PKU GIMMS NDVI v1.2 is freely available under a Creative Commons Attribution 4.0 International (CC-BY 4.0) licence via Zenodo (https://doi.org/10.5281/zenodo.8253971). All processing steps were implemented in MATLAB (R2023b) following a reproducible geospatial workflow designed to ensure spatial, temporal, and methodological consistency across the full study period. The processing pipeline includes: (i) spatial subsetting; (ii) quality control filtering; (iii) temporal compositing; (iv) multi-scale aggregation; (v) phenological metric extraction; (vi) spatial clustering; and (vii) statistical trend analysis.

3.1. Data Acquisition and Pre-Processing

The source data is provided in 16-bit unsigned integer format with a scale factor of 0.001. The pre-processing workflow was implemented through the following steps:

• Spatial Clipping and Geo-referencing: Global rasters were clipped to the Southern and Central European study area (10° W–28° E; 35° N–50° N). The study area was limited to 35–50° N, where data completeness and NDVI reliability are more suitable for long-term phenological analyses, while reducing the influence of persistent snow cover and prolonged winter data gaps at higher latitudes. Spatial referencing information was extracted from the source files to ensure that all derived products maintain the original spatial alignment and the WGS84 (EPSG:4326) geographic coordinate system.
• Bit-level Quality Control (QC): A multi-criteria filtering protocol was applied by deconstructing the three-digit quality flag provided in the auxiliary band of the PKU GIMMS data. For each pixel, the flag value was decomposed into its three functional components: the consolidation method, the GIMMS-specific quality, and the MODIS-derived quality. Pixels were retained only if all three components simultaneously met the criteria for “good quality” or “estimated data” as defined by the original data providers. In all other cases (such as detected snow, persistent cloud cover, or marginal data) the pixel was assigned a null value (NaN).
• Temporal Aggregation (MVC): Following the filtering stage, the semi-monthly NDVI maps were aggregated into monthly composites using the Maximum Value Composite (MVC) method. For each pixel, the monthly value was determined by selecting the highest NDVI value from the two available semi-monthly observations. This procedure effectively minimizes residual atmospheric noise and cloud contamination not captured by the initial quality flags.

3.2. Long-Term and Decadal Means

The monthly aggregated series were averaged across the entire 41-year period using a masked-mean function to generate 12 multi-annual monthly composites (Component 01). Similarly, the filtered monthly series were averaged across four distinct periods (1982–1992, 1993–2002, 2003–2012, 2013–2022) to establish the decadal baselines (Component 02) required for the subsequent phenological shift analysis.

3.3. Eco-Phenological Clustering (K-Means)

To group pixels with similar vegetation dynamics, an unsupervised K-means clustering algorithm was applied using Euclidean distance:

• Feature Selection: The input features were the 12-dimensional long-term monthly mean NDVI values (1982–2022). This allows the algorithm to distinguish regions based on their characteristic seasonal profiles;
• Optimization of K: A range of K (3 to 10) was evaluated using the CVIK Toolbox [4], which allowed for the computation of 19 Cluster Validity Indices (CVIs) to assess the optimal partition. The final selection of K = 5 was based on a majority consensus among these indices;
• Data Masking: Pixels with missing values (NaN) in any of the 12 monthly layers (e.g., due to persistent snow cover or coastal artifacts) were excluded from the clustering process.

3.4. Extraction of Land Surface Phenology (LSP) Metrics and Shifts

LSP metrics were extracted from decadal NDVI composites using a pixel-by-pixel “moving year” approach to accommodate regions where the growing season crosses the calendar year boundary. The algorithm identifies the annual minimum NDVI and evaluates the following 12 months using a cyclic logic to correctly identify phenophases regardless of their onset month. For each pixel and decade, NDVI values were normalized to the [0, 1] interval.

Metric Definitions:

• Start of Season (SOS): First month exceeding 20% of the annual NDVI range;
• Peak of Season (POS): Month of maximum NDVI;
• End of Season (EOS): First month after POS where NDVI drops below 30% of the range;
• Length of Season (LOS): Number of months between SOS and EOS.

The selection of these dynamic thresholds (20% for SOS and 30% for EOS) follows established protocols in remote sensing phenology [5,6,7]. These levels provide a robust balance between noise suppression and sensitivity to vegetation transitions, falling within the typical range (15–50%) used in large-scale GIMMS-based datasets and meta-analyses [8,9,10].

To quantify the shift in timing, Delta maps were generated by comparing the first and last decades. Shifts were calculated using a circular time difference logic to ensure that transitions across the calendar year boundary (e.g., from December to January) are correctly represented as a ±1 month shift rather than a ±11 month jump. These results are stored in the phenological shift sub-folder of Component 04 (4.2_Phenological_Shifts/).

3.5. Phenology Variability Index (PVI)

The PVI is a recently introduced metric proposed by Samela et al. [1] to complement traditional trend analysis by quantifying temporal variability in phenological timing.

The index is calculated from the aggregated variance of the three phenophases:

$$PVI={\displaystyle \frac{{\sigma }_{tot}}{1+{\sigma }_{tot}}}$$

where ${\sigma }_{tot}^2={\sigma }_{SOS}^2+{\sigma }_{POS}^2+{\sigma }_{EOS}^2$.

This maps ${\sigma }_{tot}$ to a [0, 1) range. A PVI = 0.5 represents a total standard deviation of approximately one month, used as an interpretative threshold for distinguishing relatively stable from more variable phenological behaviour.

The PVI was computed from 10 consecutive, non-overlapping four-year intervals spanning the period 1983–2022 (40 years total). The year 1982 was excluded from this specific index to allow for a perfectly balanced division into sub-intervals of four years each. For each interval, monthly mean NDVI values were computed to derive local SOS, POS, and EOS metrics. For each pixel, a stack of 10 values per metric was analyzed. Only pixels with at least 7 valid intervals out of 10 were retained to ensure statistical robustness; others were assigned as NaN.

3.6. Statistical Trend Analysis

The non-parametric Mann–Kendall test was applied in combination with the Theil–Sen slope estimator to monthly NDVI time series (1982–2022) for each of the five eco-phenological clusters identified through K-means clustering. The Mann–Kendall test was used to assess the presence and significance of monotonic trends (p < 0.05), while the Theil–Sen estimator was used to quantify the magnitude of change, expressed in NDVI units per year. All analyses were performed on non-normalized NDVI values. The resulting slope estimates, p-values, and sample sizes aggregated at cluster level are provided in Component 06.

3.7. Data Quality, Noise, and Limitations

Despite rigorous quality control and compositing procedures, residual noise may persist due to atmospheric contamination, snow cover effects, and sensor limitations inherent to AVHRR-based NDVI products. The use of monthly composites reduces high-frequency variability but limits temporal precision in phenological transition detection. Consequently, phenological metrics should be interpreted as monthly-scale approximations rather than exact transition dates.

3.8. Ethics Statement

The dataset does not involve human subjects, animal experiments, or personally identifiable data. All input data are openly available remote sensing products used in accordance with their original licensing and citation requirements. Therefore, ethical approval is not applicable.

4. User Notes

The dataset is designed to support multi-scale analyses of vegetation dynamics, phenology, and ecosystem variability across Southern and Central Europe. Given the modular structure of the product suite, users are advised to select specific components depending on the intended application rather than using all layers simultaneously.

For analyses of long-term seasonal behaviour, the long-term monthly NDVI climatologies (Component 01) provide a baseline representation of vegetation seasonality. These layers are suitable for characterizing broad bioclimatic gradients, comparing vegetation regimes, or serving as input features for clustering and classification tasks.

The decadal monthly NDVI composites (Component 02) are intended for inter-period comparisons and are particularly useful for detecting changes in seasonal greenness patterns across the four defined temporal windows. Users should ensure consistency in temporal aggregation when comparing these products with external climate or land cover datasets.

The eco-phenological cluster map (Component 03) provides a spatial stratification of Southern and Central Europe into phenologically homogeneous regions. This layer can be used as a static mask for regional analyses, stratified sampling, or to reduce spatial heterogeneity in modelling exercises. Users should be aware that cluster labels are categorical and non-ordinal.

The Land Surface Phenology (LSP) metrics and shift layers (Component 04) should be interpreted as monthly-scale approximations of phenological transitions. Due to the temporal resolution of the input data, SOS, POS, and EOS represent aggregated monthly states rather than exact day-of-year estimates. For applications requiring fine temporal precision, these layers should be used in combination with higher-frequency datasets.

The Phenology Variability Index (Component 05) provides a relative measure of interannual variability in phenological timing. Values close to zero indicate stable phenological behaviour, whereas higher values reflect greater temporal variability in seasonal transitions. Users should interpret this index as a comparative metric rather than an absolute uncertainty estimate.

The trend statistics (Component 06) are provided at cluster level and should be interpreted as aggregated summaries of long-term NDVI behaviour rather than pixel-level trend estimates. These results are intended for regional-scale interpretation and should not be downscaled to local inference.

All datasets are provided in WGS84 (EPSG:4326) at 0.0833° spatial resolution. Users combining these products with external datasets should ensure appropriate spatial and temporal harmonization.