Synthetic Data Augmentation for Satellite-Based Analysis of Battle-Damaged Agricultural Fields in Ukraine
Using class-conditional GAN and DDPM to augment satellite images, boosting classification accuracy to 88%.
Key Findings
Methodology
This study employs class-conditional GAN and DDPM trained on real satellite images to generate labeled bombed and non-bombed farmland samples. These synthetic images are combined with real data under balanced and proportional augmentation strategies. A Vision Transformer classifier is trained on each configuration and evaluated exclusively on real test data. GAN utilizes latent space mapping with conditional embeddings, while DDPM employs iterative noise prediction with a learned denoising network. Both models aim to improve minority class recognition by addressing data imbalance and enhancing diversity. The evaluation includes distributional metrics, feature-space distances, and downstream classification performance to assess synthetic data quality and utility.
Key Results
- Balanced DDPM augmentation improved accuracy from 84% to 88%, balanced accuracy from 67% to 81%, macro F1 from 65% to 78%, and minority class recall from 41% to 69%. These results demonstrate significant gains in recognizing underrepresented classes, especially in imbalanced datasets.
- DDPM outperformed GAN in distributional similarity, with lower FID and KID scores, indicating better alignment with real data distribution. Feature-space precision and recall metrics confirmed higher diversity and coverage for DDPM samples.
- Targeted class balancing proved more effective than proportional expansion, confirming that addressing class imbalance directly yields better downstream performance. Synthetic augmentation notably enhanced model robustness and generalization.
Significance
This work addresses the critical challenge of limited labeled data in satellite-based war damage assessment, especially in conflict zones where data collection is risky and costly. By leveraging advanced generative models, it demonstrates a practical pathway to improve classification accuracy, particularly for scarce classes. The approach enhances the capability of remote sensing systems to support post-conflict recovery, environmental monitoring, and food security analysis. Its methodology can be extended to other domains with data scarcity issues, offering a scalable solution for geospatial intelligence. The integration of synthetic data with transformer-based classifiers marks a significant step toward more resilient and adaptable remote sensing analytics.
Technical Contribution
The paper provides a comprehensive comparison between class-conditional GAN and DDPM in the context of satellite image augmentation, introducing a dual-strategy framework for balanced and proportional data expansion. It develops a multi-metric evaluation pipeline including distributional distances (KID, Mahalanobis), feature diversity, and downstream classification metrics, establishing a rigorous standard for synthetic data assessment. The work also innovates in applying diffusion models to remote sensing, demonstrating their superior stability and diversity over GANs in small-data regimes. These contributions advance the theoretical understanding and practical deployment of generative models in geospatial applications.
Novelty
This is the first systematic comparison of class-conditional GAN and DDPM for satellite imagery augmentation in a war damage context, specifically targeting Ukrainian agricultural fields. The proposed balanced augmentation strategy directly addresses class imbalance, a common challenge in remote sensing. The work introduces a comprehensive evaluation framework combining distributional metrics and classification performance, filling a gap in the literature. Its application of diffusion models to satellite damage detection is novel, showing that iterative denoising yields more diverse and realistic samples than traditional GANs, thus significantly improving minority class recognition.
Limitations
- The dataset is limited to a specific region and sensor type, restricting the generalizability of the findings. Larger, multi-region datasets are needed to validate robustness across different environments.
- Current binary classification oversimplifies damage severity, ignoring crater size, count, or damage extent. Future work should incorporate object detection and segmentation for finer analysis.
- Training diffusion models is computationally intensive, and inference speed remains a challenge for real-time applications. Optimization of model efficiency is necessary for operational deployment.
Future Work
Future research will expand datasets across multiple regions and seasons, integrating multispectral and radar data to improve model robustness. Developing multi-class and instance-level detection models will enable detailed damage assessment. Additionally, efforts to reduce computational costs and accelerate inference will facilitate real-time monitoring. Exploring semi-supervised and transfer learning approaches can further mitigate data scarcity issues, broadening the applicability of these methods in diverse geospatial contexts.
AI Executive Summary
In conflict zones like Ukraine, monitoring agricultural land damage is vital for food security and environmental stability. Traditional remote sensing approaches face significant hurdles due to limited labeled data, especially for damaged fields obscured by clouds or security restrictions. This study introduces an innovative solution by leveraging advanced generative models—class-conditional GANs and DDPMs—to synthetically augment satellite imagery. These models are trained solely on real images, then used to produce additional labeled samples of bombed and unbombed farmland, which are integrated into the training dataset.
The core idea is to address the data imbalance and scarcity issues that hinder accurate classification. The synthetic images, generated under balanced and proportional augmentation regimes, are evaluated through multiple metrics including distributional similarity (KID, Mahalanobis distance), diversity, and downstream classification performance using a Vision Transformer. Results show that the balanced DDPM augmentation notably improves accuracy from 84% to 88%, with a significant boost in recognizing the minority class, increasing recall from 41% to 69%. These improvements demonstrate that high-quality synthetic data can effectively enhance model robustness in data-scarce, high-stakes scenarios.
The research advances the field by systematically comparing GAN and DDPM in remote sensing, highlighting the superior stability and diversity of diffusion models. Its comprehensive evaluation framework ensures that synthetic data quality aligns with practical utility, setting a new standard for future work. While the dataset's limited scope poses challenges, the methodology paves the way for broader applications, including disaster assessment, environmental monitoring, and post-conflict reconstruction. Future efforts will focus on expanding datasets, refining detection granularity, and optimizing computational efficiency, ultimately enabling real-time, scalable geospatial intelligence solutions.
Deep Dive
Abstract
Monitoring war-induced damage to agricultural land in Ukraine is important for understanding threats to food security, environmental stability, and post-war recovery. However, the development of computer-vision systems for satellite-based damage analysis is limited by the scarcity of labeled imagery, especially for damaged agricultural fields. This work investigates synthetic data augmentation as a method for improving classification under limited and imbalanced training data. We train class-conditional Generative Adversarial Network (GAN) and Denoising Diffusion Probabilistic Model (DDPM) architectures on real satellite images and use them to generate additional bombed and not-bombed agricultural-field samples. The generated images are used only for training augmentation, while all downstream evaluation is performed on an exclusively real test set. A Vision Transformer classifier is trained under multiple real and synthetic data configurations to measure the practical utility of each generative approach. The best configuration, based on balanced DDPM augmentation, improves accuracy from 84\% to 88\%, balanced accuracy from 67\% to 81\%, macro F1 from 65\% to 78\%, and recall for the underrepresented not-bombed class from 41\% to 69\%. These results demonstrate the potential of synthetic satellite imagery for data-scarce geospatial applications in war-affected regions.