PaperlayerView exact revision on arXiv

Source-linked AI summary

Realtime Profiling of Fine-Grained Air Quality Index Distribution using UAV Sensing

Yuzhe Yang, Zijie Zheng, Kaigui Bian, Lingyang Song, Zhu Han

arXiv:1711.02821v1eess.SP

TL;DR

Air-quality monitoring needs finer spatial and faster temporal resolution than many existing static or mobile approaches provide. The paper develops ARMS, a UAV-based system using GPM-NN to model particle dispersion and an adaptive monitoring algorithm to select measurements. Experiments in 2D and 3D scenarios report higher AQI-map prediction accuracy with GPM-NN and reduced power consumption through adaptive monitoring.

  • Problem

    Existing AQI monitoring can be spatially coarse-grained or unable to respond in realtime to pollution changes, while fine-grained maps are needed around buildings and campuses.

  • Method

    ARMS combines UAV sensing, the GPM-NN model integrating physical dispersion with neural networks, and adaptive monitoring of selected locations.

  • Results

    GPM-NN achieves higher AQI-map prediction accuracy than existing models, while adaptive monitoring greatly reduces ARMS power consumption in 2D and 3D scenarios.

  • Takeaways & Limitations

    Selected-location monitoring with GPM-NN can generate realtime fine-grained AQI maps while balancing estimation accuracy and UAV power consumption.

Abstract

from arXiv · show

Given significant air pollution problems, air quality index (AQI) monitoring has recently received increasing attention. In this paper, we design a mobile AQI monitoring system boarded on unmanned-aerial-vehicles (UAVs), called ARMS, to efficiently build fine-grained AQI maps in realtime. Specifically, we first propose the Gaussian plume model on basis of the neural network (GPM-NN), to physically characterize the particle dispersion in the air. Based on GPM-NN, we propose a battery efficient and adaptive monitoring algorithm to monitor AQI at the selected locations and construct an accurate AQI map with the sensed data. The proposed adaptive monitoring algorithm is evaluated in two typical scenarios, a two-dimensional open space like a roadside park, and a three-dimensional space like a courtyard inside a building. Experimental results demonstrate that our system can provide higher prediction accuracy of AQI with GPM-NN than other existing models, while greatly reducing the power consumption with the adaptive monitoring algorithm.

I. INTRODUCTION

Existing AQI monitoring is often too sparse or coarse-grained for meter-level, realtime mapping, while fine-grained maps are useful for understanding pollution around buildings and campuses. The paper introduces ARMS, a UAV-based system combining GPM-NN with adaptive monitoring to construct realtime fine-grained AQI maps while reducing battery use.

  • Motivation: Static stations provide limited spatial samples, incur high costs, and may measure AQI only every two hours.The introduction cites Beijing’s 28 stations and typical separations of several tens of thousands of meters.
  • Motivation: Existing mobile sensing can be realtime and relatively accurate but remains spatially coarse-grained, despite AQI varying over meter-scale distances.The paper motivates monitoring 3D spaces around buildings or across campuses rather than only city-wide areas.
  • Limitations of Existing Approaches: Historical-feature prediction models may not respond in realtime to hourly changes in pollution concentration.The cited feature sets include public stations, meteorological data, taxi trajectories, road networks, and points of interest.
  • Limitations of Existing Approaches: Prior approaches may omit physical particle dispersion, ignore battery constraints, focus on 2D regions, or provide unsatisfactory fine-grained performance.The paper identifies these limitations across estimation, graph-based regression, and neural-network approaches.
  • Proposed System: ARMS uses UAV mobile sensing and GPM-NN to estimate unmeasured locations and generate realtime fine-grained AQI maps from selected measurements.The system is designed to capture meter-level AQI variance while conserving mobile-device battery life.
  • Proposed System: The adaptive monitoring algorithm guides UAV trajectories using GPM-NN, reducing battery consumption while maintaining high realtime AQI-map accuracy.The contribution statement describes optimized trajectory selection across different scenarios.

A. System Overview

ARMS combines a UAV-mounted air-quality sensor with spatial modeling and regression-based parameter selection to collect fine-grained AQI data in 2D and 3D scenarios. Measurements were compared with official data, while hypothesis testing identified wind and location as relevant AQI factors.

  • System Architecture: ARMS mounts an air-quality sensor on a UAV to record pollutant concentrations and calculate AQI at measured locations.The laser-based detector measures common AQI pollutants with a monitor error of no more than ±3%.
  • System Architecture: The UAV follows programmed trajectories and hovers for 10 seconds at measurement locations, with battery life limiting continuous operation to at most 15 minutes.GPS supplies the UAV’s realtime 3D position during measurements.
  • Dataset: ARMS collected more than 100 days of data in a roadside park and an office-building courtyard representing 2D and 3D monitoring scenarios.The dataset was gathered from February 11 to July 1, 2017.
  • Dataset: The dataset represents each sample with 3D coordinates and an AQI value, using 5 m spacing in the measured directions.The 2D scenario sets z = 0, while the 3D scenario varies height along z.
  • Model Parameter Selection: Wind and location were highly related to AQI distribution, whereas temperature and humidity were not.These findings came from least-squares regression and coefficient hypothesis tests.
  • Modeling: The model-development section combines physical particle dispersion with a neural network to represent nonlinear cases such as extreme weather.The stated approach embeds the dispersion model in the neural network to form the distribution model.

A. Physical Particle Dispersion Model

GPM-NN combines a revised Gaussian Plume Model with a neural network to model fine-grained AQI distributions. The physical component captures particle dispersion, while the neural component addresses nonlinear changes that the physical model alone cannot handle.

  • Physical dispersion model: The revised Gaussian Plume Model describes particle dispersion using physical variables including wind speed, source height, source length, particle density, and diffusion parameters.The model represents pollution mainly as a line source and incorporates wind-direction effects along the z-axis.
  • GPM-NN structure: GPM-NN embeds the revised Gaussian Plume Model within a neural network to combine physical dispersion modeling with nonlinear estimation.
  • GPM-NN structure: The model uses the dispersion output, a constant value, and neural-network features as inputs to represent physical influence alongside learned nonlinear effects.
  • Model output: GPM-NN estimates AQI through a model output t(x,u), while ε(x) represents measurement error modeled as Gaussian white noise.

2) Parameter Estimation:

The parameter-estimation procedure fits GPM-NN parameters from ARMS data by minimizing residual error. Least-squares estimation handles the linear coefficients, while Newton’s method solves for parameters without an analytical solution.

  • Parameter setup: GPM-NN has (K+3) parameters, including H and β1 through βK+2, estimated from ARMS-collected data.The nonlinear component uses 50 days of training data.
  • Least-squares estimation: Least-squares regression estimates the model parameters by minimizing the residual error function S.
  • Numerical optimization: Newton’s method solves the parameter equations when the residual-error system has no analytical solution.
  • Model completion: After estimating H, the corresponding dispersion output C(x,u) is determined for the model.
  • Least-squares estimation: The parameter solution uses the Moore-Penrose pseudo-inverse J† = (J^T J)^−1J^T for an over-determined linear system.The cited formulation is stated to have a unique minimum solution.

3) Performance Evaluation:

The evaluation measures AQI-map estimation accuracy using average estimation accuracy and compares GPM-NN with existing models. Training preprocessing supplies initial neural-network weights and biases before the model is finalized.

  • Model preparation: Training data preprocessing determines initial hidden-layer weights W and biases b before the model is fully specified.
  • Accuracy metric: Average estimation accuracy evaluates AQI-map performance by comparing estimated and measured AQI values across all scenario locations.The metric uses n locations, estimated values Ĉf(i), and measured values Cf(i).
  • Model comparison: The study compares the accuracy of AQI maps constructed with GPM-NN against existing models.

IV. ADAPTIVE AQI MONITORING ALGORITHM

ARMS reduces monitoring effort by alternating complete baseline construction with selective measurements at informative locations. It uses GPM-NN to update realtime AQI maps and retriggers complete monitoring when measurements indicate major environmental changes.

  • Algorithm design: The adaptive algorithm divides AQI monitoring into complete monitoring and selective monitoring to balance map accuracy with energy use.The approach relies on GPM-NN’s physical dispersion characteristics to measure only a few locations regularly.
  • Monitoring phases: Complete monitoring establishes a baseline AQI distribution, while selective monitoring measures a small set of observation points periodically.The algorithm describes complete monitoring between days and selective monitoring between hours.
  • Selective monitoring: Selective monitoring updates the realtime AQI map by combining current measurements with the previous map through GPM-NN.
  • Adaptive retriggering: ARMS retriggers complete monitoring when newly measured results differ substantially from the most recent measurements.The discrepancy is treated as an indication of severe environmental changes.
  • Algorithm outcome: The monitoring procedure is designed to reduce measurement effort while handling unpredictable spatial and temporal AQI variation.

A. Complete Monitoring

ARMS combines complete monitoring for baseline AQI characterization with selective monitoring that uses PDT-guided cube selection to update maps while reducing measurement overhead. It then plans UAV trajectories by balancing target-cube PDT against travel cost, with complete and selective monitoring triggered at different intervals.

  • Complete monitoring measures all 5m×5m×5m cubes to build a baseline AQI map, but its high dissipation requires long observation intervals.
  • Selective monitoring uses the previous AQI map and physical characteristics to measure only selected cubes while generating an updated map over the entire area.
  • PDT captures local parameter-change thresholds and identifies cubes with large changes or very small PDT values for measurement.
  • Measuring cubes in M can generate a realtime AQI map while greatly reducing measurement overhead.
  • Lower PDT increases measured cubes, accuracy, and battery consumption, whereas higher PDT reduces consumption but can reduce accuracy.
  • ARMS uses a greedy sub-optimal trajectory that maximizes the next cube’s PDT while minimizing travel cost, with total operation time O(n^2).
  • Complete monitoring is triggered between days, selective monitoring between hours, and large map deviations after severe changes can trigger baseline rebuilding.

V. APPLICATION SCENARIO I: PERFORMANCE ANALYSIS IN HORIZONTAL OPEN SPACE

In a horizontal open-space scenario, ARMS evaluates GPM-NN and its adaptive monitoring algorithm against alternative prediction and selection methods. GPM-NN achieves the highest accuracy, including nearly 80% accuracy at PDT = 0.75 while measured cubes are reduced.

  • The evaluation uses a roadside-park-like horizontal open space and examines GPM-NN, monitoring algorithms, and battery–accuracy tradeoffs.
  • GPM-NN is compared with linear interpolation and multi-variable linear regression for AQI estimation across PDT values.
  • GPM-NN achieves the highest accuracy in the horizontal open-space scenario, and accuracy decreases as PDT increases because fewer cubes are measured.
  • At PDT = 0.75, GPM-NN maintains almost 80% accuracy while the other models decline rapidly.
  • Adding the neural-network nonlinear component improves accuracy over the physical regression-only model, with 1000 hidden-layer neurons achieving the highest accuracy in this scenario.

2) Effects of Neuron Numbers:

The experiments examine neuron count, AQI level, trajectory-selection power use, and the battery–accuracy tradeoff in the 2D scenario. Results favor richer nonlinear models and show that adaptive monitoring can sharply reduce consumption while retaining substantial accuracy.

  • Effects of Neuron Numbers: Nonzero hidden-layer neurons outperform the regression-only model, and 1000 neurons achieve the highest estimation accuracy in this scenario.
  • AQI Values: GPM-NN performs best when AQI ≥200, remains highly accurate for 50 ≤AQI ≤200, and is relatively worse at low AQI while still outperforming other models.
  • Trajectory Selection: Sequential selection consumes the most normalized battery power, while the adaptive monitoring algorithm performs best against greedy selection for 0.1 ≤PDT ≤0.4.
  • Consumption and Accuracy: When PDT ≥0.4, total system consumption decreases by 90%, while average estimation error increases as PDT becomes larger.
  • Consumption and Accuracy: Choosing PDT = 0.41 achieves predicting accuracy over 80% while greatly reducing system battery consumption.

VI. APPLICATION SCENARIO II: PERFORMANCE ANALYSIS IN VERTICAL ENCLOSED SPACE

In the vertical enclosed 3D scenario, ARMS evaluates GPM-NN against other models and examines how model size and monitoring choices affect accuracy and battery use. GPM-NN maintains the highest accuracy, while neuron count and PDT influence performance.

  • Scenario: The experiment evaluates ARMS in a vertical enclosed space such as a courtyard inside a high-rise building.
  • Model accuracy: GPM-NN maintains the highest accuracy among three models across different PDTs in the vertical enclosed space scenario.At high PDT, its accuracy can decline sharply because fewer target cubes are selected.
  • Model accuracy: Compared with the 2D scenario, LI declines more rapidly as PDT increases, indicating heterogeneous AQI distribution in 3D.
  • Effects of neuron numbers: As PDT increases, models with fewer hidden-layer neurons perform worse than models with more neurons.The 3D results identify 500 neurons as achieving the highest estimation accuracy, unlike the 2D result.
  • Effects of neuron numbers: GPM-NN with combined linear and nonlinear components is more robust than a model using only the linear component.The optimal hidden-layer neuron count varies across scenarios.

3) Effects of Various AQI:

The 3D evaluation studies GPM-NN across AQI levels and compares monitoring strategies and their battery–accuracy tradeoff. GPM-NN performs best at moderate and high AQI, while adaptive monitoring can substantially reduce power consumption when PDT is selected appropriately.

  • Effects of various AQI: GPM-NN performs best when AQI is moderate or high, but performs relatively worse when AQI is low.The authors conclude that this accuracy profile suits ARMS operation.
  • Monitoring algorithms: Adaptive monitoring performs best among the compared monitoring methods in the 3D scenario.At low PDT, sequential selection consumes much more power than adaptive monitoring and the greedy algorithm.
  • Battery–accuracy tradeoff: As PDT increases, average error grows rapidly while power consumption drops substantially.
  • Battery–accuracy tradeoff: When ERR = 0.04, corresponding to average estimation accuracy of about 80%, PDT = 0.51 and power consumption can fall to 37%.Selecting PDT appropriately therefore reduces measuring effort while balancing accuracy and battery consumption.
  • Overall result: The conclusion reports that GPM-NN improves AQI-map accuracy over existing models while adaptive monitoring greatly saves UAV battery life.

APPENDIX A PROOF OF PROPOSITION 1

The appendix proves convexity properties for the optimization terms and establishes a unique minimum for the relevant equation under a bounded measurement-height assumption.

  • Convexity proof: The proof also states that ∂S/∂βj is convex for j ∈ [1, K+2].
  • Convexity proof: The second-order derivative with respect to H is analyzed by expressing each summation term as a quadratic function Qi(ti).
  • Assumption: The proof constrains H to [0, H0], where H0 is the upper bound for a fine-grained measurement, and selects an appropriate diffusion parameter σz.
  • Result: The derivative ∂S/∂H is convex, so equation (6) has a minimum with a unique corresponding value.
Loading 1711.02821v1…