Empirical Evaluation of Small Object Detection Techniques on Satellite Imagery

Global: Empirical Evaluation of Small Object Detection Techniques on Satellite Imagery

Study Overview

Researchers Xiaohui Yuan, Aniv Chakravarty, Lichuan Gu, Zhenchun Wei, Elinor Lichtenberg, and Tian Chen submitted an empirical study on February 5, 2025, later revised on December 30, 2025, that assesses methods for detecting small objects in satellite images. The work focuses on evaluating four state‑of‑the‑art detection algorithms using publicly available high‑resolution remote‑sensing data, with the goal of identifying performance strengths and technical challenges.

Context and Importance

Detecting diminutive targets such as individual vehicles or agricultural structures from space‑borne platforms is critical for applications ranging from urban traffic analysis to precision farming. Small‑object detection remains difficult because of limited pixel footprints, background clutter, and variability in imaging conditions.

Methodological Approach

The authors conducted a literature review to select four leading detection models that have demonstrated strong results on similar tasks. Each model was re‑implemented and trained on the chosen satellite datasets, allowing a side‑by‑side comparison under consistent experimental settings.

Experimental Scenarios

Two application scenarios were examined: (1) detection of cars in densely built urban environments and (2) identification of bee boxes on agricultural lands. These scenarios were chosen to represent both densely populated and sparsely structured settings, testing the algorithms’ adaptability.

Data Sources

All experiments relied on publicly released, high‑resolution satellite image collections. The datasets provide sufficient spatial detail to resolve objects measuring only a few meters across, enabling a realistic assessment of detection capabilities.

Key Findings

Results indicate that while all four methods achieve reasonable recall on the car‑detection task, performance diverges markedly on the bee‑box scenario, where object size and background homogeneity pose additional challenges. The study highlights the trade‑off between model complexity and inference speed, and underscores the need for specialized training strategies to mitigate class imbalance.

Future Directions

The authors suggest that further research should explore multi‑scale feature aggregation and domain‑adaptation techniques to improve robustness across varied terrain types. Enhancing annotation quality and expanding benchmark datasets are also recommended to foster continued progress in the field.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.