TinyNina: A Resource-Efficient Edge-AI Framework for Sustainable Air Quality Monitoring via Intra-Image Satellite Super-Resolution

Prasanjit Dey; Zachary Yahn; Bianca Schoen-Phelan; Soumyabrata Dev

cs.LGApr 6, 2026

TinyNina: A Resource-Efficient Edge-AI Framework for Sustainable Air Quality Monitoring via Intra-Image Satellite Super-Resolution

Authors: Prasanjit Dey, Zachary Yahn, Bianca Schoen-Phelan, Soumyabrata Dev

Organizations: ADAPT Research Centre, Dublin, Ireland · School of Computer Science, Technological University Dublin, Ireland · School of Computer Science, University College Dublin, Ireland · School of Computer Science and Statistics, Trinity College Dublin

Abstract

Nitrogen dioxide (NO $_2$ ) is a primary atmospheric pollutant and a significant contributor to respiratory morbidity and urban climate-related challenges. While satellite platforms like Sentinel-2 provide global coverage, their native spatial resolution often limits the precision required, fine-grained NO $_2$ assessment. To address this, we propose TinyNina, a resource-efficient Edge-AI framework specifically engineered for sustainable environmental monitoring. TinyNina implements a novel intra-image learning paradigm that leverages the multi-spectral hierarchy of Sentinel-2 as internal training labels, effectively eliminating the dependency on costly and often unavailable external high-resolution reference datasets. The framework incorporates wavelength-specific attention gates and depthwise separable convolutions to preserve pollutant-sensitive spectral features while maintaining an ultra-lightweight footprint of only 51K parameters. Experimental results, validated against 3,276 matched satellite-ground station pairs, demonstrate that TinyNina achieves a state-of-the-art Mean Absolute Error (MAE) of 7.4 $μ$ g/m $^3$ . This performance represents a 95% reduction in computational overhead and 47 $\times$ faster inference compared to high-capacity models such as EDSR and RCAN. By prioritizing task-specific utility and architectural efficiency, TinyNina provides a scalable, low-latency solution for real-time air quality monitoring in smart city infrastructures.

Figures & tables

Fig. 1: Map illustrating the locations of air pollution monitoring stations that provide ground-truth data for NO 2 pollutant levels.

Attribute	Description
Geographic Coverage	27 EPA monitoring stations across the U.S. West Coast
Spatial Diversity	Urban, suburban, and rural environments
Temporal Coverage	2018–2020
Seasonal Variability	Winter accumulation and summer photochemical events
Satellite Data	Sentinel-2 Level-2A multispectral imagery
Ground Truth	EPA NO 2 monitoring measurements

TABLE I: Characteristics of the evaluation dataset used in this study.

Fig. 2: Detailed information on the twelve Sentinel-2 spectral bands for a specific location. The rows are color-coded to differentiate spatial resolutions: yellow highlights the 10m bands, red represents the 20m bands, and blue indicates the 60m bands, providing a clear overview of the wavelength and resolution for each band.

Fig. 3: End-to-end architecture for NO 2 prediction integrating (1) Sentinel-2 data preprocessing, (2) TinyNina spectral super-resolution, and (3) ResNet50-based concentration estimation. The system maintains temporal synchronization between satellite acquisitions and ground station measurements while preserving spectral-spatial features critical for accurate pollution mapping.

Fig. 4: TinyNina’s super-resolution architecture: (a) Spectral attention gates weight bands by NO 2 sensitivity, (b) Depthwise separable convolutions reduce parameters while extracting spatial-spectral features, and (c) Residual upsampling with PixelShuffle generates high-resolution outputs.

Model	Params	Key Characteristics
EDSR	40.7M	32 residual blocks with 256 channels; deep convolutional processing
RCAN	15.4M	Residual-in-residual structure with channel attention
NinaB1	1.02M	Hybrid attention-convolution with 64 feature channels
TinyNina	51K	Spectral-optimized with depthwise separable convolutions

TABLE II: Summary of super-resolution model architectures evaluated in this study, highlighting the progression from high-capacity baselines (EDSR, RCAN) to lightweight designs (NinaB1, TinyNina).

Hyperparameter	Super-Resolution ( $f_{\theta}^{(s)}$ )	NO 2 Prediction ( $f_{\phi}^{(s)}$ )
Optimizer	Adam	Adam
Learning rate	$5\times 10^{-5}$ – $1\times 10^{-3}$	$5\times 10^{-5}$ – $1\times 10^{-3}$
LR scheduler	Step decay ( $\times 0.5$ /10 epochs)	Step decay ( $\times 0.5$ /10 epochs)
Batch size	1	1
Number of epochs	200	70
Loss function	$\mathcal{L}_{\text{naive}}$ / $\mathcal{L}_{\text{channel}}$	MSE

TABLE III: Training hyperparameters used for TinyNina super-resolution and NO 2 prediction models.

Fig. 5: Training convergence comparison of super-resolution models across Naive SR and Channel SR tasks. The proposed TinyNina model demonstrates faster and more stable convergence compared to baseline architectures (EDSR, RCAN, and NinaB1). In the Channel SR setting, TinyNina achieves optimal performance within 50 epochs, while requiring substantially fewer training iterations than EDSR, which requires approximately 200 epochs to converge despite having 800 $\times$ more parameters.

Fig. 6: Qualitative comparison between the reference image, the native low-resolution input, and the corresponding TinyNina super-resolved output. The zoomed-in regions highlight that TinyNina restores finer spatial structures and local intensity variations that are degraded in the low-resolution input while preserving the overall scene layout.

Fig. 7: Convergence behavior of NO 2 prediction models trained on super-resolved inputs from different architectures. Models trained on TinyNina-enhanced images learn much faster, typically reaching their best performance 40 to 50 epochs earlier than those using outputs from traditional models. TinyNina also achieves higher accuracy, with a lower final validation MAE of 7.4 $\mu g/m^{3}$ compared to 8.2 $\mu g/m^{3}$ for EDSR.

Model	MSE ( $\mu g/m^{3}$ )	MAE ( $\mu g/m^{3}$ )
EDSR (Naive SR)	112	8.2
RCAN (Naive SR)	98	7.8
TinyNina (Channel SR)	97	7.4

TABLE IV: Comparative performance of super-resolution models on NO 2 prediction. TinyNina (Channel SR) achieves the lowest mean squared error (MSE = 97

\mu g/m^{3}

2 ) and mean absolute error (MAE = 7.4

\mu g/m^{3}

), outperforming state-of-the-art naive super-resolution models EDSR and RCAN.

Fig. 8: Geographic distribution of NO 2 prediction errors across monitoring sites. TinyNina consistently delivers low and stable MAE, particularly in urban areas, where it maintains a standard deviation below 2.1 $\mu g/m^{3}$ .

Variant	Attention	MSE	MAE
TinyNina (without attention)	$\times$	102	7.9
TinyNina (proposed)	$\checkmark$	97	7.4

TABLE V: Ablation study evaluating the contribution of spectral attention gates in TinyNina.

Model	Params	Ext. Data	NO 2 -Opt.	Real-Time
EDSR [ 5 ]	40.7M	\textcolor red✗	\textcolor red✗	\textcolor red✗
RCAN [ 25 ]	15.4M	\textcolor red✗	\textcolor red✗	\textcolor red✗
NinaB1 [ 6 ]	1.02M	\textcolor green✓	\textcolor red✗	\textcolor orange★
FeNet [ 24 ]	158K	\textcolor red✗	\textcolor red✗	\textcolor orange★
Omni-SR [ 23 ]	792K	\textcolor red✗	\textcolor red✗	\textcolor red✗
TinyNina (Ours)	51K	\textcolor green✓	\textcolor green✓	\textcolor green✓

TABLE VI: Comparison of TinyNina with state-of-the-art super-resolution models, evaluated on four critical criteria: parameter efficiency (Params), use of external training data (Ext. Data), NO 2 -specific optimization (NO 2 -Opt.), and real-time inference capability (Real-Time). \textcolor green✓ = fully supported, \textcolor orange★ = partially supported, \textcolor red✗ = not supported. TinyNina is the only model achieving all objectives simultaneously.

Fig. 9: Inference time comparison of super-resolution models (TinyNina, NinaB1, RCAN, EDSR) for processing 500 satellite images (200 $\times$ 200 pixels, $\sim$ 1.2 $\times$ 1.2 km) on an Intel Core i7 CPU. Faster inference reduces computational energy consumption per prediction, enabling more sustainable large-scale satellite monitoring systems.

Component	Tiny	EDSR	RCAN	Nina	FeNet	Omni
Parameters	51K	40.7M	15.4M	1.02M	158K	792K
Attention Mechanism	\textcolor green✓	\textcolor red✗	\textcolor green✓	\textcolor green✓	\textcolor green✓	\textcolor green✓
Spectral Optimization	\textcolor green✓	\textcolor red✗	\textcolor red✗	\textcolor red✗	\textcolor red✗	\textcolor red✗
Depthwise Convolution	\textcolor green✓	\textcolor red✗	\textcolor red✗	\textcolor red✗	\textcolor red✗	\textcolor red✗

TABLE VII: Architectural comparison of super-resolution models, emphasizing TinyNina’s novel design choices. TinyNina achieves radical efficiency (51K parameters) while uniquely incorporating spectral optimization and depthwise convolutions-features that are absent in all baseline models. \textcolor green✓ = supported; \textcolor red✗ = unsupported.

Fig. 10: Practical deployment pipeline integrating TinyNina with existing environmental monitoring infrastructure. Sentinel-2 satellite observations undergo preprocessing before spectral enhancement using TinyNina. The enhanced imagery is then used for NO 2 prediction through a trained ResNet-based regression model. The resulting predictions are integrated with ground monitoring stations and deployed through edge/cloud inference systems to support large-scale environmental monitoring and decision-making platforms.

Device / Model	Specifications	Latency	Throughput	Model Size
		(ms/tile)	(tiles/s)	(MB)
Intel Core i7 CPU	8 cores, 3.2 GHz, 16 GB RAM	$\sim$ 90	$\sim$ 11	0.2
Jetson Nano	128 CUDA cores, 4 GB RAM	$\sim$ 200–250*	$\sim$ 4–5	0.2
Jetson Xavier NX	384 CUDA cores, 48 Tensor cores	$\sim$ 90–100*	$\sim$ 10–12	0.2
EDSR (baseline)	40.7M parameters	$\sim$ 4200	$\sim$ 0.24	$\sim$ 163
RCAN (baseline)	15.4M parameters	$\sim$ 2520	$\sim$ 0.40	$\sim$ 62

TABLE VIII: Edge-device deployment feasibility and approximate inference performance for TinyNina compared with baseline models.

Metric	Value	Notes
Inference time per tile	$\sim$ 90 ms	Intel Core i7 CPU benchmark
CPU power consumption	$\sim$ 65 W	Typical desktop CPU TDP
Energy per inference	$\sim$ 5.85 J	Estimated from time × power
Energy for 1M tiles	$\sim$ 1.6 kWh	Large-scale monitoring scenario

TABLE IX: Estimated energy consumption for TinyNina inference.

Autoregressive Image Super-Resolution

See all 10 papers

PollutionNet: A Vision Transformer Framework for Climatological Assessment of NO $_2$ and SO $_2$ Using Satellite-Ground Data Fusion
Mar 31, 2026Prasanjit Dey, Soumyabrata Dev, Bianca Schoen-PhelanAir Quality Prediction Multimodal Sensor Fusion
School of Computer Science, Technological University Dublin, Ireland. · ADAPT Research Ireland Centre, Ireland. · School of Computer Science, University College Dublin, Ireland.
Tracking Urban Atmospheric Pollutants using Sentinel-5P Satellite Data
May 21, 2026Alice Gomez-Cantos, Henry O. Velesaca
Facultad de Ciencias Naturales y Matemáticas, Escuela Superior Politécnica del Litoral, ESPOL, Campus Gustavo Galindo, Km. 30.5 Vía Perimetral, Guayaquil, 090902, Ecuador · Software Engineering Department, Research Center for Information and Communication Technologies (CITIC-UGR), University of Granada, 18071, Granada, Spain
TinyML Enhances CubeSat Mission Capabilities
Mar 20, 2026Luigi Capogrosso, Michele MagnoOn-Device Inference Energy-Efficient ML
Interdisciplinary Transformation University of Austria · ETH Zurich

Mar 31, 2026J/K move · Enter open · S save

PollutionNet: A Vision Transformer Framework for Climatological Assessment of NO $_2$ and SO $_2$ Using Satellite-Ground Data Fusion

Prasanjit Dey, Soumyabrata Dev, Bianca Schoen-Phelan

School of Computer Science, Technological University Dublin, Ireland. · ADAPT Research Ireland Centre, Ireland. · School of Computer Science, University College Dublin, Ireland.

Air Quality Prediction Multimodal Sensor Fusion

Open paper AskPDF arXiv

Accurate assessment of atmospheric nitrogen dioxide (NO

_2

) and sulfur dioxide (SO

_2

) is essential for understanding climate-air quality interactions, supporting environmental policy, and protecting public health. Traditional monitoring approaches face limitations: satellite observations provide broad spatial coverage but suffer from data gaps, while ground-based sensors offer high temporal resolution but limited spatial extent. To address these challenges, we propose PollutionNet, a Vision Transformer-based framework that integrates Sentinel-5P TROPOMI vertical column density (VCD) data with ground-level observations. By leveraging self-attention mechanisms, PollutionNet captures complex spatiotemporal dependencies that are often missed by conventional CNN and RNN models. Applied to Ireland (2020-2021), our case study demonstrates that PollutionNet achieves state-of-the-art performance (RMSE: 6.89

μ

g/m

^3

for NO

_2

, 4.49

μ

g/m

^3

for SO

_2

), reducing prediction errors by up to 14% compared to baseline models. Beyond accuracy gains, PollutionNet provides a scalable and data-efficient tool for applied climatology, enabling robust pollution assessments in regions with sparse monitoring networks. These results highlight the potential of advanced machine learning approaches to enhance climate-related air quality research, inform environmental management, and support sustainable policy decisions.

Abstract

Autoregressive Image Super-Resolution

Explore similar work

PollutionNet: A Vision Transformer Framework for Climatological Assessment of NO2_22​ and SO2_22​ Using Satellite-Ground Data Fusion

Tracking Urban Atmospheric Pollutants using Sentinel-5P Satellite Data

TinyML Enhances CubeSat Mission Capabilities

PollutionNet: A Vision Transformer Framework for Climatological Assessment of NO $_2$ and SO $_2$ Using Satellite-Ground Data Fusion