Abstract Particulate matter (PM) of size 2.5 and 10 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\upmu$</EquationSource> <EquationSource Format="MATHML"><math> <mi mathvariant="normal">μ</mi> </math></EquationSource> </InlineEquation>m poses significant health risks as it penetrates deep into the lungs. This paper presents a novel AI model to estimate concentrations of PM from night-time images. Low ambient illumination at night causes the captured images to be granular, and they are additionally affected by the presence of bright, spurious light sources, making feature extraction challenging. It is essential to monitor PM levels at night, as they are often reported to be higher than during the day due to cooler ground temperatures. Developing a single model to tackle both day and night-time images is likely to result in a performance that is far from optimal. Hence, the focus of this article is on the constraints and issues specific to night-time images and on the construction of a model exclusively for PM estimation from such images. Two separate modules are developed: one for capturing the depth information and the other for the spectral reflectance. The fusion of depth information of objects from the camera and the spectral reflectance caused by particulate matter forms the basis of our unique approach. Furthermore, we introduce a robust and diverse dataset featuring images captured across different seasons with a wide range of structural variety. In the absence of existing approaches for PM determination from night-time images, we have compared our results with the performance of existing approaches as applied to our night-time image dataset. This research marks the first successful attempt to assess air quality, which varies over a wide range, using night-time images, offering a promising new avenue for pollution monitoring. Research Highlights <UnorderedList Mark="Bullet"> <ItemContent> Major Indian cities face a severe problem of poor air quality specifically during winter, requiring an affordable solution for the residents to be able to estimate the level of air pollution. The study emphasizes on an affordable vision based solution to gauge the concentration of particulate matters. </ItemContent> <ItemContent> This is the first research of its kind to measure the particulate concentration during nighttime only using images. </ItemContent> <ItemContent> A diverse dataset of 60,000 images collected using a PTZ camera covering different scenes with structural objects at different levels at an interval of 15 minutes. The reference particulate matter (PM) data were obtained from the Air Quality Monitoring Station at Rabindra Bharati University, Kolkata, located approximately 2.5 km from various locations captured by the camera. </ItemContent> <ItemContent> A novel data-driven model based on Convolution Neural Network (CNN) takes into consideration the scattering of light with the change in the concentration of particulates exhibiting daily and seasonal variations and the structures at varying depths from the point of image acquisition. </ItemContent> <ItemContent> Results of the CNN model affirm the model’s efficacy in capturing complex, non-linear features responsible for the fluctuation of the particulate concentration. </ItemContent> </UnorderedList>

Dark-AIRNET: A deep learning network for image-based estimation of air particulate concentration

期刊论文
发表日期：

2026年1月29日

查看全文

Harsh Bhandari,
Sarbani Palit

摘要

Abstract

Particulate matter (PM) of size 2.5 and 10 $\upmu$ $μ$ m poses significant health risks as it penetrates deep into the lungs. This paper presents a novel AI model to estimate concentrations of PM from night-time images. Low ambient illumination at night causes the captured images to be granular, and they are additionally affected by the presence of bright, spurious light sources, making feature extraction challenging. It is essential to monitor PM levels at night, as they are often reported to be higher than during the day due to cooler ground temperatures. Developing a single model to tackle both day and night-time images is likely to result in a performance that is far from optimal. Hence, the focus of this article is on the constraints and issues specific to night-time images and on the construction of a model exclusively for PM estimation from such images. Two separate modules are developed: one for capturing the depth information and the other for the spectral reflectance. The fusion of depth information of objects from the camera and the spectral reflectance caused by particulate matter forms the basis of our unique approach. Furthermore, we introduce a robust and diverse dataset featuring images captured across different seasons with a wide range of structural variety. In the absence of existing approaches for PM determination from night-time images, we have compared our results with the performance of existing approaches as applied to our night-time image dataset. This research marks the first successful attempt to assess air quality, which varies over a wide range, using night-time images, offering a promising new avenue for pollution monitoring.

Research Highlights

Major Indian cities face a severe problem of poor air quality specifically during winter, requiring an affordable solution for the residents to be able to estimate the level of air pollution. The study emphasizes on an affordable vision based solution to gauge the concentration of particulate matters.

This is the first research of its kind to measure the particulate concentration during nighttime only using images.

A diverse dataset of 60,000 images collected using a PTZ camera covering different scenes with structural objects at different levels at an interval of 15 minutes. The reference particulate matter (PM) data were obtained from the Air Quality Monitoring Station at Rabindra Bharati University, Kolkata, located approximately 2.5 km from various locations captured by the camera.

A novel data-driven model based on Convolution Neural Network (CNN) takes into consideration the scattering of light with the change in the concentration of particulates exhibiting daily and seasonal variations and the structures at varying depths from the point of image acquisition.

Results of the CNN model affirm the model’s efficacy in capturing complex, non-linear features responsible for the fluctuation of the particulate concentration.