Abstract
Scaling laws assume larger models trained on more data consistently outperform smaller ones—an assumption that drives model selection in computer vision but remains untested in resource-constrained Earth observation (EO). We conduct a systematic efficiency analysis across three scaling dimensions: model size, dataset size, and input resolution, on rooftop PV detection in Madagascar. Optimizing for model efficiency (mAP$_{50}$ per unit of model size), we find a consistent efficiency inversion: YOLO11N achieves both the highest efficiency ($24\times$ higher than YOLO11X) and the highest absolute mAP$_{50}$ (0.617). Resolution is the dominant resource allocation lever ($+$120% efficiency gain), while additional data yields negligible returns at low resolution. These findings are robust to the deployment objective: small high-resolution configurations are Pareto-dominant across all 44 setups in the joint accuracy-throughput space, leaving no tradeoff to resolve. In data-scarce EO, bigger is not just unnecessary: it can be worse.
Introduction
In the world of deep learning, the prevailing wisdom is often "scale or fail." Scaling laws suggest that larger models trained on more data will inevitably push the needle of performance. However, this mantra is increasingly challenged in Earth Observation (EO), where data is scarce and compute resources are often limited. This study investigates whether standard scaling laws hold true for detecting small objects, such as rooftop solar panels, in drone imagery over Madagascar.
Most computer vision research occurs in data-rich environments like COCO or ImageNet. In contrast, EO datasets are structurally data-scarce due to the high cost and domain expertise required for annotation. Furthermore, these models are often intended for deployment on edge devices where compute power is a luxury. This research addresses a critical question for practitioners: Given a fixed resource budget, how should it be allocated between model complexity, dataset size, and image resolution?
Methodology
The researchers conducted a systematic efficiency analysis using the YOLO11 family (Nano to Extra-Large), exploring three scaling dimensions:
- Model Size: Five variants ranging from 2.6M to 57M parameters.
- Dataset Fraction: Training on 10%, 25%, 50%, and 100% of the available data.
- Input Resolution: 416px, 640px, and 1280px.
The study utilized the OpenStat Madagascar rooftop PV dataset, a crowdsourced collection of drone and satellite imagery. The task is particularly challenging as it involves "small objects"; 64% of the annotated bounding boxes occupy less than 1% of the total image area. Efficiency was measured as $10 \times \frac{mAP_{50}}{ModelSize (MB)}$, evaluating performance per unit of storage.
The Efficiency Inversion
In a reversal of standard scaling laws, the smallest model, YOLO11N, achieved the highest absolute accuracy (mAP). It was not merely efficient for its size; it was the top performer overall. The Nano model proved to be 24× more efficient than the Extra-Large YOLO11X variant.
Large models suffered from a high "overparameterization ratio," tending to memorize noise rather than generalize features in this data-scarce regime. Consequently, YOLO11X required the full dataset just to approach the performance that YOLO11N achieved with only 10% of the data.
Resolution and the Pareto Frontier
Increasing input resolution emerged as the most effective way to boost performance. Moving from 416px to 1280px provided a +120% efficiency gain. At low resolutions, adding more training data yielded almost no improvement because the objects were too small to be resolved by the model.
Crucially, the study found no tradeoff between accuracy and speed. Small, high-resolution configurations were Pareto-dominant, being simultaneously the most accurate and the fastest at inference. YOLO11N at 1280px resolution achieved the highest absolute $mAP_{50}$ of 0.617.
Conclusion
For the EO community, the implications are clear: when working with limited data and small objects, massive models are often counterproductive. Prioritizing higher resolution imagery and utilizing "Nano" model variants provides superior results in both accuracy and inference speed. In resource-constrained small object detection, smaller is not just leaner—it is smarter.
Summary updated