Deep learning reveals severe shelterbelt degradation in eastern Ukraine due to the war
- 1National University of Life and Environmental Sciences of Ukraine
- 2Swedish University of Agricultural Sciences
Abstract
Russia’s full-scale invasion of Ukraine in February 2022 has caused extensive damage to forests, particularly in eastern and southern regions with sustained combat, where field shelterbelts have often been used as cover and subsequently heavily destroyed. This study presents a systematic region-scale assessment of pre-war shelterbelt area and war-related loss between 2021 and 2025 using 3 m PlanetScope imagery and deep learning. We evaluated three input datasets – monthly PlanetScope RGB mosaics, multiseasonal PlanetScope RGB stacks, and surface reflectance PlanetScope imagery – in combination with two semantic segmentation architectures: a convolutional U-Net with a ResNet-34 encoder and a transformer-based SegFormer with a MiT-B0 backbone. Across all input configurations, the U-Net consistently outperformed SegFormer, achieving the highest F1 scores. Monthly RGB mosaics delivered accuracy comparable to more complex multiseasonal and surface reflectance inputs in pixel-based evaluation and are a cost-efficient option for mapping narrow shelterbelts. The RGB-based model trained on July 2021 mosaics reached an F1 score of 0.71 and retained values of 0.58–0.67 for other months of the same leaf-on season. Temporal transfer to 2025 resulted in lower but still reasonable performance with F1 scores of 0.53–0.63. Error-adjusted mapping estimated the shelterbelt area in 2021 at 254±14.5 thousand ha, equivalent to a mean density of 2.6 ha/km2. By 2025, this area had been reduced by approximately 41 thousand ha, or 16% of the pre-war area. Losses were strongly concentrated along frontlines where local losses exceeded 1.5 ha/km2 compared to an average loss of 0.5 ha/km2 across the entire study area. The resulting maps complement the existing forest monitoring framework in Ukraine and provide a spatially explicit basis for prioritizing post-war remediation and assessing losses in protective functions and carbon accumulation.
Related
- Classification and quantification of low-visibility events using deep learning over eastern China
- Deciphering the impacts of meteorology on surface ozone variability in eastern China using explainable machine learning models
- Multi source Earth observation and machine learning framework for mapping land degradation vulnerability in salinity- and hydrothermal-stressed regions of the southern main Ethiopian Rift
- Machine learning reveals biocontrol agents shaping disease outcome in natural Arabidopsis populations
- Sensitivity Fidelity Reveals Physical Dependencies Beyond Forecast Accuracy in Deep-Learning Earth-System Models
BibTeX
@article{myroniuk2026deep,
title = {{Deep learning reveals severe shelterbelt degradation in eastern Ukraine due to the war}},
author = {Viktor Myroniuk and Yevhenii Khan and Oleksandr Sovakov and Maksym Matsala and Andrii Lashchenko and Anatolii Makarevych and Andriy Terentyev},
journal = {Environmental Research Letters},
year = {2026},
month = oct,
doi = {10.1088/1748-9326/aebaa1},
url = {https://iopscience.iop.org/article/10.1088/1748-9326/aebaa1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Deep learning reveals severe shelterbelt degradation in eastern Ukraine due to the war
AU - Viktor Myroniuk
AU - Yevhenii Khan
AU - Oleksandr Sovakov
AU - Maksym Matsala
AU - Andrii Lashchenko
AU - Anatolii Makarevych
AU - Andriy Terentyev
PY - 2026
DA - 2026/10/28
JO - Environmental Research Letters
DO - 10.1088/1748-9326/aebaa1
UR - https://iopscience.iop.org/article/10.1088/1748-9326/aebaa1
AB - Russia’s full-scale invasion of Ukraine in February 2022 has caused extensive damage to forests, particularly in eastern and southern regions with sustained combat, where field shelterbelts have often been used as cover and subsequently heavily destroyed. This study presents a systematic region-scale assessment of pre-war shelterbelt area and war-related loss between 2021 and 2025 using 3 m PlanetScope imagery and deep learning. We evaluated three input datasets – monthly PlanetScope RGB mosaics, multiseasonal PlanetScope RGB stacks, and surface reflectance PlanetScope imagery – in combination with two semantic segmentation architectures: a convolutional U-Net with a ResNet-34 encoder and a transformer-based SegFormer with a MiT-B0 backbone. Across all input configurations, the U-Net consistently outperformed SegFormer, achieving the highest F1 scores. Monthly RGB mosaics delivered accuracy comparable to more complex multiseasonal and surface reflectance inputs in pixel-based evaluation and are a cost-efficient option for mapping narrow shelterbelts. The RGB-based model trained on July 2021 mosaics reached an F1 score of 0.71 and retained values of 0.58–0.67 for other months of the same leaf-on season. Temporal transfer to 2025 resulted in lower but still reasonable performance with F1 scores of 0.53–0.63. Error-adjusted mapping estimated the shelterbelt area in 2021 at 254±14.5 thousand ha, equivalent to a mean density of 2.6 ha/km2. By 2025, this area had been reduced by approximately 41 thousand ha, or 16% of the pre-war area. Losses were strongly concentrated along frontlines where local losses exceeded 1.5 ha/km2 compared to an average loss of 0.5 ha/km2 across the entire study area. The resulting maps complement the existing forest monitoring framework in Ukraine and provide a spatially explicit basis for prioritizing post-war remediation and assessing losses in protective functions and carbon accumulation.
SN - 1748-9326
ER -