🤖 AI

Synthetic Data Augmentation for Satellite-Based Analysis of Battle-Damaged Agricultural Fields in Ukraine

This paper demonstrates that augmenting scarce satellite imagery with synthetic data generated by Denoising Diffusion Probabilistic Models (DDPMs) significantly improves the accuracy and balanced performance of computer vision classifiers in detecting battle-damaged agricultural fields in Ukraine.

Original authors: Marta Sumyk, Oleksandr Kosovan, Iryna Voitsitska

Published 2026-08-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Marta Sumyk, Oleksandr Kosovan, Iryna Voitsitska

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the aftermath of conflict, the land itself often bears the deepest scars. For nations trying to rebuild, knowing exactly where the soil has been torn apart by shelling is a matter of survival, food security, and safety. Modern technology offers a way to see these wounds from above using satellites, but turning those images into useful maps requires a great deal of human effort. Computers need to be taught what a damaged field looks like, and that teaching usually demands thousands of labeled examples. In war zones, gathering such a large collection of images is slow, dangerous, and often impossible. When there are too few examples, the computer programs that analyze the sky become unreliable, often missing the very damage they are meant to find. To solve this, researchers have begun exploring a method where computers create their own practice materials, generating new images of damaged and undamaged fields to train the system, hoping these artificial examples can teach the machine as well as real photographs can.

A team of researchers at the Ukrainian Catholic University has put this idea to the test, focusing specifically on the agricultural fields of Ukraine. Their goal was to see if they could teach a computer to distinguish between a peaceful, green field and one scarred by bomb craters, even when they had very few real images to work with. They started with a small dataset of satellite photos from the Bakhmut region, where the ground is captured in high detail, showing individual craters. The data was heavily skewed, containing many more images of bombed fields than of undamaged ones, a common problem that confuses computer vision systems. To fix this, the researchers trained two different types of artificial intelligence models on the real images they had. One model, known as a generative adversarial network, and another, a diffusion model, learned the visual patterns of the fields. Instead of just copying the existing photos, these models learned to create entirely new, synthetic images of both bombed and undamaged fields, effectively filling in the gaps in their training library.

The researchers then used these newly created images to train a sophisticated image-recognition system called a Vision Transformer. They tested two different strategies: one where they generated just enough fake images to balance the number of damaged and undamaged examples, and another where they simply doubled the amount of data while keeping the original imbalance. The results were clear. The system trained with the balanced, synthetic data performed significantly better than the one trained only on real images. Most importantly, the system became much better at correctly identifying the undamaged fields, a category it had previously struggled to recognize. By using the synthetic images to balance the training set, the researchers improved the system's ability to correctly identify undamaged fields from 41 percent to 69 percent. The overall accuracy of the system rose from 84 percent to 88 percent, and a measure of how well the system handled both types of fields equally jumped from 67 percent to 81 percent.

The study also revealed that not all methods of creating fake data are equal. The diffusion model, which builds images by gradually removing noise from a random pattern, produced better results than the adversarial network. The images generated by the diffusion model were more diverse and closely matched the statistical patterns of the real satellite photos, whereas the other model tended to repeat similar textures. This suggests that for this specific task, the quality and variety of the synthetic data matter more than simply having a larger volume of images. The researchers found that simply adding more data without fixing the imbalance did not help as much as carefully adding the missing type of data. The best-performing system was the one that used the diffusion model to specifically generate more examples of the rare, undamaged fields, allowing the computer to learn a more complete picture of what a safe field looks like.

While the results are promising, the researchers are careful to note the limits of their work. The images came from a single region and a specific type of satellite, so it is not yet known if these methods will work perfectly in different parts of Ukraine or with different sensors. Furthermore, the system is designed to identify damage, not to determine if a field is safe to enter or if unexploded ordnance is present. Nevertheless, the study demonstrates that synthetic data can be a powerful tool for geospatial analysis in war-affected regions. By teaching computers with a balanced mix of real and generated examples, it is possible to build more reliable tools for assessing the damage to the land, providing a clearer view of the path toward recovery.

Technical Summary: Synthetic Data Augmentation for Satellite-Based Analysis of Battle-Damaged Agricultural Fields in Ukraine

Problem Statement
The monitoring of war-induced damage to agricultural land in Ukraine is critical for assessing food security, environmental stability, and post-war recovery. However, developing computer-vision systems for this task is hindered by a scarcity of labeled satellite imagery, particularly for damaged fields. Real-world data collection is constrained by security risks, cloud cover, restricted access to high-resolution imagery, and the labor-intensive nature of annotating craters. Furthermore, existing datasets are often small and imbalanced, with "bombed" patches significantly outnumbering "not-bombed" ones. This imbalance causes standard classifiers to perform well on the majority class while failing to recognize the underrepresented minority class, rendering overall accuracy metrics misleading.

Methodology
The authors propose a pipeline to evaluate whether class-conditional generative models can improve classification performance under these data-scarce and imbalanced conditions.

Dataset: The study utilizes a subset of 600 real satellite-image patches from the Bakhmut region (acquired by Planet SkySat satellites at ~0.5m resolution). The data is split into 470 training images (402 bombed, 68 not-bombed) and 130 test images (108 bombed, 22 not-bombed). The task is formulated as a binary classification: identifying whether a patch contains visible impact craters ("bombed") or not.
Generative Models: Two class-conditional architectures are trained exclusively on the real training set to generate synthetic samples for both classes:
- Class-Conditional GAN: A generator maps latent vectors and class labels to synthetic images, while a discriminator distinguishes real from fake.
- Class-Conditional DDPM: A Denoising Diffusion Probabilistic Model uses a U-Net to iteratively predict and remove noise from a latent image, conditioned on the class label.
Augmentation Strategies: The generated images are added to the real training set under two regimes:
- Balanced: Synthetic samples are added until the "not-bombed" class matches the size of the "bombed" class (correcting the imbalance).
- Proportionally Doubled: The number of samples in both classes is doubled, preserving the original imbalance.
Downstream Evaluation: A Vision Transformer (ViT) classifier is trained on these augmented datasets. Crucially, all evaluation is performed on a fixed, exclusively real test set to measure practical generalization.
Quality Metrics: Synthetic image quality is assessed via feature-space analysis (using a pretrained encoder), including Kernel Inception Distance (KID), class-conditional Mahalanobis distance, Gaussian Mixture Model (GMM) log-likelihood, precision, recall, diversity, and memorization checks (duplicate detection).

Key Contributions

Task Formulation: Defines shelling-related agricultural damage recognition as a binary satellite-image classification problem.
Model Comparison: Trains and compares class-conditional GANs and DDPMs specifically for generating bombed and not-bombed agricultural field imagery.
Augmentation Strategy Analysis: Compares targeted class balancing against proportional data expansion to determine which strategy better addresses data scarcity and imbalance.
Rigorous Evaluation: Evaluates all downstream models on an exclusively real test set, ensuring that performance gains are not artifacts of synthetic data overfitting.
Multi-Perspective Analysis: Analyzes synthetic data through distributional similarity, feature-space coverage, diversity, and memorization criteria, rather than relying solely on visual inspection.

Results
The study demonstrates that synthetic augmentation significantly improves the classification of the underrepresented "not-bombed" class without compromising performance on the majority class.

Performance Gains: The best configuration (Real + Balanced DDPM) improved:
- Accuracy: 84% $\to$ 88%
- Balanced Accuracy: 67% $\to$ 81%
- Macro F1: 65% $\to$ 78%
- Recall for "not-bombed" class: 41% $\to$ 69%
Model Comparison: DDPM-based augmentation consistently outperformed GAN-based augmentation in both balanced and doubled regimes. While the paper notes that the specific feature-space metrics (GMM likelihood, Mahalanobis distance, precision, and recall) reported in the draft are provisional estimates intended to complete the current draft and must be replaced by direct measurements before final submission, the observed downstream classification trends suggest that DDPM samples align more closely with the real-data distribution and offer broader coverage compared to GANs, which showed signs of lower diversity.
Strategy Comparison: Targeted class balancing yielded superior results compared to proportionally doubling the data, indicating that correcting the class imbalance is more effective than simply increasing dataset volume while preserving the skew.

Significance and Claims
The paper claims that synthetic satellite imagery holds significant potential for geospatial applications in war-affected regions where labeled data is scarce. The primary significance lies in demonstrating that targeted synthetic augmentation can effectively mitigate the performance degradation caused by class imbalance in remote sensing tasks.

The authors modestly note that the system is intended to support damage analysis and prioritization, not to determine field safety or detect unexploded ordnance. They acknowledge limitations regarding the small size and geographic specificity of the dataset (Bakhmut region only), which limits conclusions about generalization to other soil types, seasons, or sensors. However, the results suggest that generative models, particularly DDPMs, can serve as a viable tool for enhancing training data in conflict zones, provided the generated samples are rigorously validated against real-world performance metrics.

Technical Summary: Synthetic Data Augmentation for Satellite-Based Analysis of Battle-Damaged Agricultural Fields in Ukraine

More like this