Bocchino et al. (2025) Crop flood damage assessment integrating Sentinel-2 imagery and in situ data: the 2023 Emilia-Romagna case

Identification

• Journal: Remote Sensing Applications Society and Environment
• Year: 2025
• Date: 2025-12-31
• Authors: Filippo Bocchino, Valeria Belloni, Roberta Ravanelli, Camillo Zaccarini, Mattia Crespi, Roderik Lindenbergh
• DOI: 10.1016/j.rsase.2025.101852

Research Groups

• Geodesy and Geomatics Division, Department of Civil, Constructional and Environmental Engineering (DICEA), Sapienza University of Rome, Rome, Italy
• Sapienza School for Advanced Studies, Sapienza University of Rome, Rome, Italy
• Risk Management Department, Institute of Services for Agricultural and Food Market (ISMEA), Rome, Italy
• Department of Geoscience and Remote Sensing, Delft University of Technology, Delft, The Netherlands
• Geomatics Unit, Department of Geography, University of Liège, Liège, Belgium

Short Summary

This study proposes a data-driven machine learning framework to quantitatively assess crop flood damage by integrating Sentinel-2 imagery and in situ field data. Applied to the May 2023 Emilia-Romagna flood, the Random Forest model achieved an overall accuracy of 0.74 in classifying agricultural fields into three damage categories, providing a reliable tool for post-event support.

Objective

• To develop a data-driven approach for assessing crop flood damage using a machine learning classification framework applied to features derived from Earth Observation (EO) data.
• To train and test the model on field-level in situ damage data collected by agronomists, enabling a direct, quantitative assessment of flood impacts linked to real-world observations.
• To classify agricultural fields into three flood damage categories (no damage, medium damage, and high damage) by integrating spatial and temporal variations in Sentinel-2-derived spectral indices, topographic information, and flood extent maps.

Study Configuration

• Spatial Scale: Emilia-Romagna region, Italy, focusing on 412 agricultural fields. Input data resolutions: Sentinel-2 (10–20 m), NASADEM (30 m), Copernicus Emergency Management Service (CEMS) flood products (2.5–16 m).
• Temporal Scale: Analysis of the May 2023 flood event. Sentinel-2 imagery collected between April and June 2023. NASADEM data from February 2000. CEMS flood products from May 2023. In situ ISMEA surveys conducted at harvest dates (per product).

Methodology and Data

• Models used: Random Forest (RF) classifier. Other machine learning models compared include Gradient Boosting (GB), AdaBoost (AdaB), k-Nearest Neighbors (kNN), C-Support Vector Classification (SVC), Nu-Support Vector Classification (NuSVC), Decision Tree Classifier (DT), Gaussian Naive Bayes (GauNB), and Logistic Regression (LR).
• Data sources:
  • Remote Sensing:
    • Harmonized Sentinel-2 MSI: MultiSpectral Instrument, Level-2A imagery (via Google Earth Engine).
    • NASADEM: NASA 30m Digital Elevation Model (via Google Earth Engine).
    • Copernicus Emergency Management Service (CEMS) flood extent maps (EMSR659, EMSR664 for May 2023 floods).
  • In situ data:
    • Field-level crop damage assessments for 412 agricultural fields provided by the Institute of Services for Agricultural and Food Market (ISMEA). Damage was defined as (Potential Production - True Production) / Potential Production × 100.
  • Processing Platform: Google Earth Engine (GEE) for data pre-processing and Google Colaboratory for machine learning pipeline implementation.

Main Results

• The Random Forest (RF) model achieved the highest mean F1 score of 0.74 ± 0.04 across 100 independent tests, outperforming other tested machine learning models.
• The model demonstrated robust and stable classification performance with an overall accuracy of 0.74 ± 0.04, precision of 0.75 ± 0.04, and recall of 0.74 ± 0.04.
• High-damage fields were accurately identified, characterized by greater flood exposure (mean flooded area of 62.67%), lower elevations (mean of 2.42 m), and significant declines in vegetation indices (e.g., mean ΔNDVI April-May decrease of 0.33).
• The model struggled to distinguish between no-damage and medium-damage fields, particularly for permanent crops, where damage beneath the canopy or occluded flooded areas posed challenges. Misclassification between medium-damage and no-damage fields primarily affected permanent crops (99.53% and 65.13% respectively).
• The most significant features influencing classification were flooded area, ΔMNDWI April-May, crop class, ΔNDVI April-May, ΔMNDWI May-June, and elevation.

Contributions

• Provides a direct, quantitative assessment of flood-induced agricultural damage in terms of yield losses, moving beyond traditional flood extent mapping.
• Utilizes a unique in situ crop damage dataset for training and testing a Random Forest model, enabling a data-driven evaluation directly linked to real-world observations.
• Integrates a diverse set of Earth Observation-derived features, including vegetation indices (NDVI, LAI, MNDWI), flood extent maps, and field elevation.
• Offers a decision-support tool based on freely available EO data, designed to aid post-event compensation and decision-making in flood-prone agricultural regions.

Funding

• Grant for young researchers AR123188B3C1EB82 (Sapienza University of Rome, Italy)
• Grant for international mobility (Sapienza University of Rome, Italy)
• Doctoral Program fellowship within the National PhD in Earth Observation (Institute of Services for Agricultural and Food Market (ISMEA), Italy)

Citation

@article{Bocchino2025Crop,
  author = {Bocchino, Filippo and Belloni, Valeria and Ravanelli, Roberta and Zaccarini, Camillo and Crespi, Mattia and Lindenbergh, Roderik},
  title = {Crop flood damage assessment integrating Sentinel-2 imagery and in situ data: the 2023 Emilia-Romagna case},
  journal = {Remote Sensing Applications Society and Environment},
  year = {2025},
  doi = {10.1016/j.rsase.2025.101852},
  url = {https://doi.org/10.1016/j.rsase.2025.101852}
}