Assessment of war-induced agricultural land use changes in Ukraine using machine learning applied to Sentinel satellite data
Оценка изменений землепользования сельскохозяйственных угодий в Украине, вызванных войной, с использованием машинного обучения и данных спутников Sentinel
2025-05-08
SCID: 54.1/99ksj8px
Discuss with AI
Agricultural land use changeMachine learning classificationRandom ForestSentinel-1 and Sentinel-2Transfer learning
Figures from the paper
Abstract (AI)
The ongoing war in Ukraine has significantly disrupted agricultural land use, leading to reduced cropland areas, increased land abandonment, and heightened uncertainty in food production. This study presents a multi-year assessment of war-induced agricultural land use changes in Ukraine using machine learning-based classification applied to Sentinel-1 and Sentinel-2 satellite imagery. By leveraging cloud computing platforms, including Google Earth Engine (GEE) and the Copernicus Data Space Ecosystem (CDSE), we develop high-resolution KPI-Ukraine (Igor Sikorsky Kyiv Polytechnic Institute (KPI) in Ukraine) land use maps spanning from 2016 to 2024. The study integrates Random Forest and Multi-Layer Perceptron classification techniques to improve accuracy, addressing spectral ambiguities and classification noise. Additionally, a novel transfer learning approach enables reliable classification in conflict-affected areas with limited ground-truth data. We achieved high classification accuracy across the 14 major crop types in Ukraine and abandoned land, validated through F1-scores exceeding 90 % for most classes. The fusion of the results generated on the GEE and CDSE platforms enhanced the classification accuracy for minor classes. Our analysis reveals significant reductions in cultivated land in 2022–2024, particularly in conflict zones, where agricultural activity has been heavily disrupted. Overall, Ukraine’s arable land area shrunk by 10 % nationwide. The consistently high accuracy of our classification methodology across the nine-year study period demonstrates its robustness and suitability for long-term monitoring of agricultural dynamics in conflict-affected regions and provides a valuable tool for guiding post-war recovery efforts. Our findings underscore the importance of leveraging satellite data for timely and accurate land use monitoring, supporting policymakers in addressing food security challenges and promoting sustainable agricultural practices. This framework also holds potential for broader applications in monitoring land use changes in conflict zones and regions undergoing rapid environmental shifts.
Key Findings
1
Classification achieved F1-scores above 90% for most of 14 major crop classes and abandoned land; combining Google Earth Engine and CDSE outputs improved minor-class accuracy.
2
Consistently high accuracy across nine years demonstrates the methodology’s suitability for long-term monitoring and post-war agricultural recovery planning.
3
Cultivated land declined substantially during 2022–2024, especially in conflict zones, and Ukraine’s total arable land area decreased by 10% nationwide.
4
Random Forest, Multi-Layer Perceptron, and transfer learning enabled reliable crop and abandoned-land classification in conflict-affected areas with limited ground-truth data.
5
The study produced high-resolution agricultural land-use maps for Ukraine from 2016–2024 using Sentinel-1 and Sentinel-2 imagery with machine-learning classification.
Research Object
Agricultural land use in Ukraine, including cropland, major crop types, and abandoned land, during the 2016–2024 war period
Research Subject
War-induced spatiotemporal changes in agricultural land use, including cultivated-area reduction, land abandonment, and classification dynamics across conflict-affected regions
Publication Details
Publication Date
2025-05-08
Journal
Publisher
ISSN
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest