Changes in land surface phenology reveal the spread of Xylella fastidiosa through olive orchards in Apulia, Southern Italy

Abstract Xylella fastidiosa is a plant pathogenic bacterium that has severely affected olive orchards across Apulia, Southern Italy over the past decade, causing progressive canopy loss and transforming the landscape from tree-dominated to herbaceous-dominated vegetation. This structural and compositional shift alters the seasonal pattern of vegetation greenness, or land surface phenology, by replacing deep-rooted, evergreen olive trees with shallow-rooted, seasonal grasses that respond differently to shorter- and longer-term water availability. Here, we propose a model-based, sensor-agnostic framework for landscape-scale monitoring of the damages caused by Xylella fastidiosa in Apulia. We use long-term MODIS NDVI time series (2002–2025) combined with climate variables to establish a pre-epidemic baseline NDVI response against which disease-driven deviations can be detected as model residuals. The model achieved a high coefficient of determination (R² = 0.85) during the calibration period (2002–2008) before the first confirmed outbreak. Applying the model across all olive orchards in Apulia produced annual, landscape-wide maps of orchard-scale canopy damage, revealing a northward progression of severe decline from the original outbreak site near Gallipoli. Remote sensing-based detection of damages lagged qPCR-based field detection of the disease by 2.8 ± 2.4 years, consistent with the time required for disease to progress to orchard-level structural change detectable at satellite scale. Validation against field-based disease severity observations confirmed that the detection threshold corresponds to a severity index of approximately 3 on a 0–4 scale. Cross-sensor comparison with Sentinel-2 demonstrates that the approach is transferable across platforms through seasonal data harmonisation, enabling near-real-time monitoring even where long-term baselines are unavailable. Use The manuscript data is formatted as an R project, and can be loaded as such in the RStudio environment. The project makes use of the {targets} framework for reproducible analysis. To run the full workflow, run: library(targets) targets::tar_make() Please make sure to install all required libraries beforehand or restore the {renv} lockfile: # restore with the environment with the most recent # copies of the libraries packages <- c( "lme4", "zoo", "tidyverse", "patchwork", "sf", "terra", "lwgeom", "units", "rnaturalearth", "rnaturalearthdata", "rnaturalearthhires", "CDSE", "appeears", "ecmwfr" ) install.packages(packages) # restore the evironment using the lockfile snapshot # pinned to library versions install.packages("renv") renv::restore() Gathering raw driver data Due to the size of the source data not all data is included in full. The data-raw directory includes the data which is distributed, but more importantly, download scripts to download the source data. The download scripts (with the download_ prefix) adhere to the provided directory structure and post-processing is described in the process_ scripts. Combining final driver data The raw data (in data-raw) is combined into one large driver dataset using the 01_compile_model_data.R script. This will generate the main dataset from which all other analysis will flow. Expanding the analysis therefore requires updating the raw data and running the script anew, while also adjusting the analysis scripts. This final data is stored in the data folder, together with other data which are used without much further pre-processing. Statistics The final model approach is described in the analysis directory. This directory provides all scripts to fit the model, and generate all figures for publication (which are stored in manuscript/figures). When reviewing code the analysis folder should be the only one to consider, if you are not interested in how the data was compiled.