Ezzaher et al. (2026) NDVI-UNet: A novel approach for improved vegetation segmentation using Sentinel-2 images

Identification

• Journal: Remote Sensing Applications Society and Environment
• Year: 2026
• Date: 2026-01-01
• Authors: Fatima Ezahrae Ezzaher, Nizar Ben Achhab, Hafssa Naciri, Naoufal Raissouni
• DOI: 10.1016/j.rsase.2026.101905

Research Groups

• Mathematics and Intelligent Systems, Abdelmalek Essˆaadi University, Tangier, Morocco
• Remote Sensing, Systems and Telecommunications, Abdelmalek Essˆaadi University, T´etouan, Morocco

Short Summary

This study introduces NDVI-UNet, a novel deep learning approach for improved vegetation semantic segmentation using Sentinel-2 imagery, which effectively mitigates common misclassification issues like the "blue roof problem" and reduces the need for extensive manual annotation. The method, combining NDVI with a UNet model, achieved superior accuracy compared to traditional methods and other automatic labeling techniques across diverse Mediterranean climates and seasons.

Objective

• To develop a novel approach for vegetation semantic segmentation that mitigates misclassifications from Vegetation Indices (VIs), specifically the "blue roof issue" in NDVI, and reduces the laborious manual annotation required for training deep learning models, by merging the strengths of both methods.

Study Configuration

• Spatial Scale: Mediterranean climates (regional scale, analyzed using Sentinel-2 images).
• Temporal Scale: Multi-seasonal (across four seasons).

Methodology and Data

• Models used: UNet, LinkNet, FPN (Deep Learning models) with ResNet34 and ResNet50 backbones. Input configurations included RGB, RGB-NIR, and RGB-NDVI.
• Data sources: Sixteen Sentinel-2 satellite images. Vegetation masks generated using NDVI. Comparison baselines included Maximum Entropy and ESA WorldCover map.

Main Results

• The best-performing model was UNet with ResNet50 and RGB-NDVI inputs, achieving an Intersection over Union (IoU) of 94.82 %, an F1-score of 97.28 %, and an Accuracy of 98.21 %.
• The proposed method outperformed two other automatic labeling techniques (Maximum Entropy and ESA WorldCover map) with an IoU of 85.87 %, an F1-score of 93.05 %, and an Accuracy of 91.51 %.
• The approach effectively mitigated the "blue roof issue" in vegetation segmentation.

Contributions

• Presents a novel deep learning approach (NDVI-UNet) that effectively merges Vegetation Indices with advanced segmentation models.
• Significantly improves vegetation segmentation accuracy by addressing common misclassification problems, such as the "blue roof issue" associated with NDVI.
• Reduces the dependence on extensive and laborious manual annotation for training segmentation models.
• Demonstrates superior performance compared to existing automatic labeling techniques.

Funding

• Not specified in the provided text.

Citation

@article{Ezzaher2026NDVIUNet,
  author = {Ezzaher, Fatima Ezahrae and Achhab, Nizar Ben and Naciri, Hafssa and Raissouni, Naoufal},
  title = {NDVI-UNet: A novel approach for improved vegetation segmentation using Sentinel-2 images},
  journal = {Remote Sensing Applications Society and Environment},
  year = {2026},
  doi = {10.1016/j.rsase.2026.101905},
  url = {https://doi.org/10.1016/j.rsase.2026.101905}
}