PaperlayerView exact revision on arXiv

Source-linked AI summary

Federated Learning in the Sky: Aerial-Ground Air Quality Sensing Framework with UAV Swarms

Yi Liu, Jiangtian Nie, Xuandi Li, Syed Hassan Ahmed, Wei Yang Bryan Lim, Chunyan Miao

arXiv:2007.12004v1eess.SPcs.LG

TL;DR

Accurate, timely AQI monitoring is needed because air pollution threatens human health, while existing systems face coverage, energy, and privacy constraints. The paper proposes a federated aerial-ground framework using UAV Dense-MobileNet vision and ground GC-LSTM sensing; experiments report accurate, energy-efficient, privacy-preserving AQI sensing. Its stated future scope includes reducing federated communication overhead and compressing models for UAV deployment.

  • Problem

    AQI monitoring must be accurate, timely, scalable, energy-efficient, and privacy-preserving despite existing coverage, energy, and data-sharing constraints.

  • Method

    The framework combines UAV federated learning with lightweight Dense-MobileNet vision and ground graph-based GC-LSTM modeling for fine-grained 3D and spatio-temporal AQI inference.

  • Results

    The proposed framework achieves accurate aerial and ground AQI sensing while preserving raw-data privacy and reducing UAV energy consumption.

  • Takeaways & Limitations

    Federated aerial-ground sensing supports collaborative, fine-grained AQI monitoring without requiring institutions to share raw user data.

  • Takeaways & Limitations

    Future work must reduce communication overhead and compress deep-learning models because UAV power limits long-term monitoring and large models are difficult to deploy on mobile devices.

Abstract

from arXiv · show

Due to air quality significantly affects human health, it is becoming increasingly important to accurately and timely predict the Air Quality Index (AQI). To this end, this paper proposes a new federated learning-based aerial-ground air quality sensing framework for fine-grained 3D air quality monitoring and forecasting. Specifically, in the air, this framework leverages a light-weight Dense-MobileNet model to achieve energy-efficient end-to-end learning from haze features of haze images taken by Unmanned Aerial Vehicles (UAVs) for predicting AQI scale distribution. Furthermore, the Federated Learning Framework not only allows various organizations or institutions to collaboratively learn a well-trained global model to monitor AQI without compromising privacy, but also expands the scope of UAV swarms monitoring. For ground sensing systems, we propose a Graph Convolutional neural network-based Long Short-Term Memory (GC-LSTM) model to achieve accurate, real-time and future AQI inference. The GC-LSTM model utilizes the topological structure of the ground monitoring station to capture the spatio-temporal correlation of historical observation data, which helps the aerial-ground sensing system to achieve accurate AQI inference. Through extensive case studies on a real-world dataset, numerical results show that the proposed framework can achieve accurate and energy-efficient AQI sensing without compromising the privacy of raw data.

I. INTRODUCTION

AQI monitoring must become more accurate, timely, fine-grained, scalable, energy-efficient, and privacy-preserving. The paper proposes a federated aerial-ground framework combining UAV vision, Dense-MobileNet, and GC-LSTM ground sensing.

  • Air pollution threatens human health, making accurate and timely AQI monitoring important for advance protective measures and pollution control.
  • Existing sensor-based methods provide coarse 2D coverage or require costly, energy-intensive dense deployment and mobile travel.
  • Vision-based and crowdsourcing approaches can improve coverage but remain constrained by low-quality data and privacy policies.
  • Government AQI monitoring stations cover less than 5% of urban area, limiting real-time public access to air-quality inference.
  • GDPR restrictions on direct inter-agency data sharing create data islands, motivating accurate AQI monitoring that protects privacy.
  • The proposed framework combines UAV federated learning, Dense-MobileNet, and ground GC-LSTM sensing for fine-grained 3D AQI monitoring with lower energy use.

A. AQI Monitoring Methods

AQI monitoring spans sensor-based and vision-based approaches, each with coverage, cost, energy, data-quality, or privacy constraints. The paper combines federated UAV vision with graph-based ground sensing to infer AQI without sharing raw data.

  • Sensor-based Monitoring Methods: Sensor-based monitoring uses stationary or mobile devices, but station spacing limits precision while dense deployment and travel increase cost and energy consumption.
  • Vision-based Monitoring Methods: Vision-based monitoring supports fine-grained 3D estimation, yet crowdsourced data can suffer from low quality and restricted inter-agency exchange.
  • AQI Inference Models: Existing AQI inference models use either sensor data or vision data to predict real-time or future AQI distributions.
  • AQI Inference Models: The paper introduces Dense-MobileNet for low-energy image-based AQI scale inference and GC-LSTM for fine-grained ground AQI values.
  • Federated Learning Model: Federated learning trains models across distributed UAV datasets while preventing data leakage and enabling collaborative regional AQI monitoring.
  • Federated Learning Model: The FL procedure uploads locally trained gradients to a central server, which aggregates them into a global model for region-level AQI prediction.

B. Aerial Sensing

The aerial sensing design uses UAV imagery and federated Dense-MobileNet learning to infer region-level AQI distributions without raw-data sharing. UAV mobility and camera-only payloads support broader, longer-duration 3D monitoring.

  • UAVs capture haze images, train Dense-MobileNet locally, upload gradients, and receive a central-server global model for end-to-end region-level AQI prediction.
  • Ground WSN topology is represented with graph convolution and combined with LSTM for AQI inference and future air-quality prediction.
  • UAV swarms from different agencies collaboratively train models without sharing raw data, expanding sensing coverage across monitoring regions.
  • Different UAV heights and viewing angles provide 3D samples, while wireless communication supports federated training among swarms.
  • Dense-MobileNet learns haze features from UAV images to infer fine-grained AQI scales and addresses the limited image diversity of static-camera approaches.
  • The aerial sensing pipeline includes haze-image imaging, haze-feature extraction, Dense-MobileNet, and lightweight federated AQI inference.

A. Overview of Haze Image Imaging Principle

The framework extracts haze-related visual features from UAV images to characterize atmospheric pollution. These features connect image appearance, haze density, and particulate-matter-related changes to AQI inference.

  • Haze imaging model: The haze-imaging model describes atmospheric light attenuation using transmission, scene depth, and the medium’s extinction coefficient.The extinction coefficient is used to estimate particulate-matter concentration from wavelength attenuation.
  • Haze imaging model: The observed UAV pixel value combines attenuated scene radiance with airlight scattered by air molecules and particulate matter.The scene-radiance term and airlight term represent the two components reaching the UAV camera.
  • Haze feature extraction: Six haze-related features are extracted from binarized UAV images after grayscale conversion and Otsu thresholding.The feature set includes dark channel prior, depth map, blue channel, RMS contrast, image entropy, and image smoothness.
  • Haze feature extraction: Dark channel prior and depth map provide haze-thickness or haze-density cues from local color intensity and scene structure.The dark channel prior assumes outdoor images contain very low-intensity pixels in at least one color channel.
  • Haze feature extraction: RMS contrast measures pixel-intensity variation, while image entropy quantifies texture information that decreases as particulate matter blurs image details.Image smoothness is separately defined from the average gradient magnitude in the sky region.

C. Dense-MobileNet based Federated Learning for AQI Scale Inference

Dense-MobileNet learns AQI scale estimates from six haze features using a compact convolutional architecture. Its depthwise separable convolutions and dense feature transfer reduce computational demands while preserving end-to-end inference.

  • Model motivation: Dense-MobileNet uses six extracted haze features as inputs to a lightweight CNN for AQI scale inference.The model is selected for mobile-device suitability and learns the relationship between haze features and AQI scale.
  • Dense feature transfer: Dense feature transfer feeds outputs from all preceding layers into the next layer to improve feature learning with fewer parameters.The design is inspired by DenseNet while applying the mechanism within MobileNet.
  • Model architecture: The architecture comprises feature extraction, depthwise separable convolution, global average pooling, and fully connected layers.The feature extraction layer produces a 128 × 128 × 6 feature tensor from six haze-related image features.
  • Lightweight design: Depthwise separable convolution decomposes standard convolution into feature-map and channel operations, substantially reducing parameters and computation.The paper compares the resulting parameter and computation ratios with traditional standard convolution.
  • AQI scale inference: AQI scale inference is formulated as classification using WHO AQI intervals and combines aerial model outputs with ground perception results.The model performs end-to-end learning from haze images to AQI scale estimates.

D. Dense-MobileNet based Federated Learning Algorithm

The federated algorithm lets UAVs from different institutions train a shared Dense-MobileNet model without exchanging private sensing data. Local updates are aggregated centrally and redistributed for iterative global learning.

  • Federated learning workflow: The federated framework enables public and private institutions to collaboratively infer AQI scale while expanding UAV-swarm monitoring.The proposed algorithm is based on federated learning with Dense-MobileNet.
  • Federated learning workflow: The process begins with training that uses domain-specific public datasets without privacy concerns.This initializes the collaborative learning workflow before local UAV training and aggregation.
  • Local training: The central server distributes a global-model copy, and each UAV trains its local Dense-MobileNet copy on its own dataset.Each UAV computes a local loss over its local data during training.
  • Privacy-preserving exchange: UAVs upload model updates rather than private data, preserving the stated data-sharing boundary during collaboration.The updates are sent to the central server through the uplink.
  • Global aggregation: The central server aggregates uploaded parameters to form a new global model and distributes it back to the UAVs.This completes the iterative server-side aggregation and client-side redistribution cycle.

V. GROUND SENSING: AQI INFERENCE BY GARPH-BASED GROUD SENSOR MONITORING

The proposed GC-LSTM model combines graph-based spatial processing with LSTM-based temporal processing to infer real-time and future AQI values from ground-monitoring data.

  • GC-LSTM uses ground-monitoring topology and historical data to capture spatio-temporal dependencies for AQI inference.Ground stations are represented as graph nodes, while their relationships and historical observations provide spatial and temporal information.
  • The framework uses ground-sensing networks to provide accurate prior knowledge for aerial sensing inference.
  • The model architecture contains a spatio-temporal block, input layer, LSTM-based processing layer, and output layer.

1) Graph Construction:

The ground sensing system models monitoring stations as a graph, using station locations and spatial correlations to represent AQI observations as graph signals.

  • Graph Construction: Each monitoring station is represented as a graph node, with edges encoding relationships between monitoring stations.
  • Graph Construction: The input features of N monitoring stations at time step t are represented as graph signals in feature matrix X_t ∈ R^N×M.
  • Graph Construction: The spatial weight matrix A represents correlations between station pairs and is constructed from their spatial distances.Station distances are calculated from latitude and longitude.
  • Graph Construction: GCN propagation uses a normalized graph structure, convolution weights, and a nonlinear activation function to process graph features.

B. Temporal Dependency Modelling

The temporal modeling component combines graph convolution with LSTM sequence processing to capture spatial and temporal dependencies and predict future AQI values.

  • Temporal Dependency Modelling: Previous time-series models cannot capture the spatio-temporal dependencies in historical AQI parameters, motivating a sequence model integrated with GCN.
  • Temporal Dependency Modelling: LSTM selectively passes information through forget, input, and output gates while updating cell and hidden states.
  • Temporal Dependency Modelling: GC-LSTM predicts real-time and future AQI values from the topology of monitoring stations and historical spatio-temporal correlations.
  • Temporal Dependency Modelling: Graph convolution extracts spatial features, which are combined with graph signals as LSTM inputs before a fully connected layer produces AQI values.

VI. EXPERIMENTS

Experiments evaluate the aerial-ground framework using real-world ground observations and aerial haze images. The DMFL model achieves high prediction accuracy and robustness while preserving privacy, and GC-LSTM is compared with sequence-model baselines.

  • Evaluation Setup and Data Description: The evaluation uses static monitoring stations from several cities as graph nodes and six months of 2019 AQI observations, split into training and testing periods.The first five months form the training data and the sixth month forms the test data.
  • Evaluation Setup and Data Description: The aerial dataset contains 5,298 labeled haze images from Hangzhou, Guangzhou, Shenzhen, Harbin, and Daqing, divided 8:2 into training and test sets.
  • Ground Sensing Evaluation: GC-LSTM is evaluated against GRU, LSTM Net, and AQNet to assess fine-grained spatio-temporal AQI prediction.
  • Vision-based Aerial Sensing: DMFL achieves higher accuracy than 3D CNN, 2D CNN, and SVM, while remaining below Dense-MobileNet in the reported comparison.The reported DMFL accuracy is 97.13% lower than Dense-MobileNet in this experiment.
  • Vision-based Aerial Sensing: DMFL has RMSE results close to Dense-MobileNet and better than the other compared models.The comparison uses RMSE as an inference-robustness measure.

2) Sensor-based Ground Sensing:

The framework combines GC-LSTM ground sensing with federated aerial sensing to improve AQI inference while addressing energy use and privacy. Ground sensing captures spatio-temporal station relationships, while the aerial system uses lightweight models and federated collaboration.

  • Sensor-based Ground Sensing:: The ground system predicts AQI in real time and at horizons of 2, 4, and 6 hours.The comparison includes GRU, LSTM Net, and AQNet.
  • Sensor-based Ground Sensing:: GC-LSTM captures spatial correlations among monitoring stations and temporal dependence in air-quality data for AQI prediction.Its RMSE is reported as much smaller than those of other models.
  • Energy Efficiency:: At a monitoring range of 100 m3, DMFL energy consumption is about half that of 2D CNN.DMFL also requires the least energy for the training task among the compared models.
  • Privacy Preservation:: Organizations share model-update parameters rather than user data, preserving privacy without direct data sharing.The framework is designed to address privacy constraints while collaboratively monitoring AQI.
Loading 2007.12004v1…