When Bigger is Worse: A Practitioner's Guide to Model Selection Under Data Scarcity

Kwame Mbobda-Kuate; Gabriel Kasmi

cs.CVMar 2, 2026

When Bigger is Worse: A Practitioner's Guide to Model Selection Under Data Scarcity

Authors: Kwame Mbobda-Kuate, Gabriel Kasmi

Organizations: ENSAE Paris, Palaiseau, 91120, France. · Centre O.I.E., Mines Paris – PSL University, Sophia-Antipolis, 06904, France.

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 photovoltaic (PV) detection in Madagascar, yielding 180 training runs across 60 configurations. Optimizing for model efficiency (mAP $_{50}$ per unit of model size), we find a consistent efficiency inversion: YOLO11N achieves the highest efficiency ( $22\times$ higher than YOLO11X) with no accuracy penalty: it reaches the second highest absolute mAP $_{50}$ (0.459), outperforming all models except YOLO11S by a margin smaller than run-to-run variance, directly contradicting the scaling prior. Resolution is the dominant resource allocation lever: moving from 416 px to 1280 px at 10 % of the data matches the efficiency gain of collecting the full dataset at low resolution. These findings are robust to the deployment objective: small high-resolution configurations are Pareto-dominant across all 60 experimental setups in the joint accuracy-throughput space, leaving no tradeoff to resolve. In data-scarce EO, the scaling prior does not just fail: it inverts.

Figures & tables

Variant	Parameters	Size (MB)
YOLO11N	2.6M	5.1
YOLO11S	9.5M	18.1
YOLO11M	20.1M	38.4
YOLO11L	25.4M	48.5
YOLO11X	57.0M	108.6

Table 1: YOLO11 model variants evaluated in this study

Figure 1: Cumulative distribution of bounding box areas (normalized by image area) across the training set. The median normalized area is 0.6% of the image area, and 64% of boxes fall below the 1% threshold commonly used to define small objects in detection benchmarks ( Lin et al., 2017 ). This confirms that rooftop PV detection in drone imagery is structurally a small object task, justifying the central role of input resolution in our efficiency analysis

Figure 2: Small models at high resolution dominate across all deployment objectives. Detection performance (mAP 50 ) versus inference speed (FPS) for all 180 experimental runs (60 configurations

\times

3 seeds) on the OpenStat Madagascar rooftop PV dataset ( OpenStat Madagascar, 2025 ) . Marker shape encodes input resolution (circle: 416 px, square: 640 px, triangle: 1280 px) and marker size encodes the training dataset fraction (10% to 100%). The dashed line delineates the Pareto frontier. YOLO11N and YOLO11S at 1280 px lie at the apex, simultaneously maximising accuracy and throughput with no accuracy–speed tradeoff to resolve

Figure 3: Efficiency ranking of the evaluated models (mAP per 10 MB). YOLO11N achieves

22\times

higher efficiency than YOLO11X while also reaching the second highest absolute mAP 50 — ruling out any accuracy–efficiency tradeoff

Figure 4: Data efficiency curves for each YOLO11 variant averaged across input resolutions. Gains are front-loaded across all variants, with the largest marginal improvements occurring between 10% and 25% of the training data. YOLO11N saturates rapidly beyond 25% (

+

0.007 mAP 50 from 50% to 100%), while YOLO11L remains data-hungry throughout (

+

0.025 from 50% to 100%)

Figure 5: mAP 50 vs. overparameterization ratio

\rho=\text{params}/N_{\text{train}}

. A negative log-linear trend confirms that higher overparameterization predicts lower performance, consistent with the bias-variance tradeoff. Each point is one configuration (model

\times

fraction

\times

resolution)

Resolution	10%	25%	50%	100%
416 px	1.00	1.06	1.08	1.10
640 px	1.06	1.11	1.12	1.17
1280 px	1.08	1.15	1.14	1.17

Table 2: Average relative efficiency normalized by baseline (416 px, 10%). Best per resolution in bold . Along both axes, gains are modest and comparable: moving from 10% to 100% of the data at 416 px yields

+

10%, while moving from 416 px to 1280 px at 10% data yields

+

Figure 6: Effect of input resolution on visual information for PV detection. Same scene at 416 px (top), 640 px (middle), 1280 px (bottom). Right column: zoomed crop at fixed physical size. At 416 px the panel texture is aliased and geometrically ambiguous; at 1280 px cell structure and edges are clearly resolved

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Geographic distribution of the 8,977 drone images retained for training, colored by split assignment (train / val / test). Sampling points follow major transportation corridors and population centers, with Antananarivo as the primary cluster. The random split preserves the spatial coverage of the full dataset across all three subsets

Panel of Figure 8 — Figure 8: Representative high-resolution drone images from the OpenStat Madagascar dataset. Images illustrate the variety of rooftop types, panel orientations, and background clutter encountered across acquisition sites

Figure 9: Validation set images with ground-truth bounding box annotations. Dense clusters of small boxes are characteristic of rooftop installations with multiple panels, and represent the dominant challenge in this detection task

Hyperparameter	Value
Batch size	{2, 4, 8, 16, 32, 64}
Number of epochs	50
Optimizer	Adam
Random seed	{693763239, 3810243382, 4221376603}
Initial learning rate	0.001
Final learning rate	0.001

Appendix

Table 3: Hyperparemeters used during the training. The batch size is modulated to avoid out of memory errors depending on the resolution and the number of parameters of the model in question. This variability is compensated by gradient accumulation, which always simulated a (nominal) batch size of 64

Augmentation	Scale	Probability
Hue variation	0.015	—
Saturation variation	0.7	—
Brightness variation	0.4	—
Translation	0.1	—
Rescaling	0.5	—
Left-to-right flip	—	0.5

Appendix

Table 4: Data augmentations performed during training. The mosaic is disabled on the last 10 epochs (41 - 50)

Model	Min Eff.	Max Eff.	Avg Eff.	Best Config	mAP 50
YOLO11N	0.729	0.907	0.841	100% / 1280px	0.463
YOLO11S	0.217	0.259	0.241	100% / 1280px	0.469
YOLO11M	0.101	0.119	0.112	100% / 1280px	0.458
YOLO11L	0.073	0.094	0.084	100% / 640px	0.455
YOLO11X	0.031	0.042	0.038	100% / 640px	0.452

Appendix

Table 5: Efficiency statistics per model variant. Efficiency is defined as mAP 50 / ModelSize (MB). Best Config reports the (dataset fraction, resolution) achieving the highest efficiency, along with the corresponding mAP 50

Model	Eff. (mAP 50 )	Eff. (mAP ${}_{50\text{-}95}$ )	Eff. (Precision)	Eff. (Recall)	Eff. (F 1 )
YOLO11N	0.841	0.606	1.183	1.022	1.096
YOLO11S	0.241	0.175	0.332	0.293	0.311
YOLO11M	0.112	0.081	0.155	0.138	0.146
YOLO11L	0.084	0.060	0.118	0.103	0.109
YOLO11X	0.038	0.027	0.053	0.047	0.050

Appendix

Table 6: Efficiency under alternative performance metrics. Each column reports the average efficiency (metric / model size in MB

\times

10) across all training configurations for each model variant. YOLO11N consistently achieves the highest efficiency regardless of the metric used, confirming that the efficiency inversion reported in the main text is robust to the choice of performance measure

Fraction	Resolution	Best Model	Efficiency	mAP 50
10%	416px	YOLO11N	0.763	0.389
25%	416px	YOLO11N	0.807	0.411
50%	416px	YOLO11N	0.831	0.424
100%	416px	YOLO11N	0.828	0.422
10%	640px	YOLO11N	0.808	0.412
25%	640px	YOLO11N	0.843	0.430

Appendix

Table 7: Best model per iso-configuration. Only configurations where all five YOLO11 variants were evaluated are included. YOLO11N dominates across all tested configurations

Model	mAP 50	mAP ${}_{50\text{-}95}$	Precision	Recall	F 1
YOLO11N	0.0022	0.0011	0.0041	0.0422	0.0075
YOLO11S	0.0014	0.0007	0.0027	0.0287	0.0049
YOLO11M	0.0012	0.0005	0.0024	0.0179	0.0042
YOLO11L	0.0010	0.0004	0.0018	0.0096	0.0030
YOLO11X	0.0013	0.0006	0.0024	0.0116	0.0040

Appendix

Table 8: Zero-shot evaluation of YOLO11 variants on the OpenStat Madagascar test set, averaged across available resolutions. Models use COCO-pretrained weights without fine-tuning. F 1 is computed from precision and recall. All values are near-zero (

\leq 0.004

), confirming the absence of domain transfer from COCO pretraining to satellite PV detection

	Resolution	YOLO11N	YOLO11S	YOLO11M	YOLO11L	YOLO11X
mAP 50	416 px	0.409	0.421	0.419	0.399	0.413
	640 px	0.433	0.441	0.436	0.413	0.429
	1280 px	0.444	0.448	0.436	0.410	0.390
Efficiency	416 px	0.80	0.23	0.11	0.08	0.04
	640 px	0.85	0.24	0.11	0.09	0.04
	1280 px	0.87	0.25	0.11	0.08	0.04

Appendix

Table 9: Resolution–model size interaction. Mean mAP 50 (upper block) and mean efficiency (mAP 50 / 10 MB, lower block) per (model, resolution) cell, averaged across dataset fractions. Bold denotes the best value per column. Grey cells indicate configurations excluded due to GPU memory constraints (OOM). The efficiency gap between YOLO11N and YOLO11S remains stable across resolutions (ratio

\approx 3.5\times

at all three levels), indicating that higher resolution does not compensate for the overparameterisation penalty of larger models in terms of efficiency. In absolute mAP 50 , the gap narrows as the resolution increases

Model Selection

Model selection chooses models, features, or configurations that balance predictive fit, generalization, complexity, and deployment constraints.

See all 96 papers

TESSERA v2: Scaling Pixel-wise Earth Foundation Models
Jul 4, 2026Zhengpeng Feng, Sadiq Jaffer, Ira Shokar +10Geospatial Foundation Models AI Scaling Laws
University of Cambridge · 2NVIDIA · 3dClimate Labs
Foundations of the Scaling Law
Oct 20, 2026Tianyu Zhao, Zhixiang RenScaling Laws Test-Time Scaling
LingQiong AI Lab, Shanghai Smart Logic Technology
Efficient Neural Network Model Selection for Few-Class Application Datasets
Jun 18, 2026Bryan Bo Cao, Abhinav Sharma, Lawrence O'Gorman +2Model Selection Classification
Department of Computer Science Stony Brook University Stony Brook, NY 11790, USA · Nokia Bell Labs Murray Hill, NJ 07974, USA

Jul 4, 2026J/K move · Enter open · S save

TESSERA v2: Scaling Pixel-wise Earth Foundation Models

Zhengpeng Feng, Sadiq Jaffer, Ira Shokar, Jovana Knezevic+9

University of Cambridge · 2NVIDIA · 3dClimate Labs

Geospatial Foundation Models AI Scaling Laws

Open paper AskPDF arXiv

Pixel-wise Earth-observation (EO) foundation models are now achieving state-of-the-art performance via generated spatial embeddings. However, how these models scale and how best to spend a pretraining budget remain poorly understood. We present the largest controlled scaling study for EO to date: 395 training runs on 1,024 GH200 superchips within a fixed pixel-wise Barlow Twins family, each evaluated on 15 downstream tasks. We find that pretraining loss barely predicts downstream performance (|Pearson r| < 0.2), so selecting models by loss wastes a large share of the compute. We also find that, as the training budget grows, the encoder and the data should grow together while the projector stays fixed, which gives a simple rule for allocating compute. Using this rule, we train a family of pixel-wise models (0.5B and 1B, with a 2B model in training) and distill them into compact students for embeddings-as-data deployment. The 21-million-parameter distilled TESSERA v2-1B-M in aggregate outperforms all open and proprietary models tested, some of which are orders of magnitude larger. These students produce Matryoshka representations that are inexpensive to serve: a 16-dimensional prefix keeps 92% of the full 128-dimensional performance at 1/8 of the storage. Upon completion of training we plan to release v2 global embeddings covering 2017-2025. Together, these results give a concrete, empirically grounded recipe for scaling pixel-wise EO foundation models: train large encoders, select by downstream performance, and distil into flexible student models. All code will be released at https://github.com/ucam-eo/tessera.

Abstract

Model Selection

Explore similar work

TESSERA v2: Scaling Pixel-wise Earth Foundation Models

Foundations of the Scaling Law

Efficient Neural Network Model Selection for Few-Class Application Datasets