基於土地利用與道路影像轉移學習之時空汙染濃度估計

Abstract

空氣污染暴露對人體健康有害。固定測站監測提供了高品質的空氣污染的測量結果，但在估計旅行者在道路上污染暴露的空間變異性時受到限制。此外，它可能導致街道層級的空間變化表徵不佳。因此，移動監測已被廣泛用於收集及時的時空空氣污染測量值。 土地利用回歸(LUR)是預測空氣污染物濃度的典型模型。 LUR模型通常利用來自固定監測站點的固定測量值。另一方面，許多研究開始使用移動監測。影像數據和LUR可以組成包含固定和移動測量的混合模型。圖像分割技術常用於街景圖像數據處理。然而，大多數深度學習方法都需要標記良好的數據來進行模型訓練。此外，處理數據需要大量的人工工作。 本研究提出了一種用於未標記的車載攝影機影像的遷移學習方法和一種利用deeplabV2模型進行圖像分割的圖像特徵提取方法。我們沿著台中台灣大道進行了移動監測，以開發一個混合模型來預測二氧化碳(CO2)、氮氧化物(NOx)、黑碳(BC)和粒子數(PN)濃度。我們模型的5折交叉驗證R2分別為CO2、NOx、BC和PN的0.79、0.88、0.61和0.63。當標記良好的數據難以獲取時，這種遷移學習方法可能會有所幫助。此外，這項工作使混合模型能夠適應不同的車載攝像頭場景，並可用於估計道路污染暴露。

Air pollution exposure is harmful to human health. Stationary monitoring of air pollution provides high-quality measurements while limited when estimating spatial variability of on-road exposure of travelers. In addition, it may lead to poor characterization of spatial variation at the street level. Therefore, mobile monitoring has been widely adopted for collecting real-time air pollution measurements. The Land Use Regression (LUR) is a typical model for predicting air pollutant concentrations. LUR models usually utilize stationary measurements from fixed monitoring sites. On the other hand, numerous studies took measures from mobile monitoring sources. Image data and the LUR can form a hybrid model with stationary and mobile measurements. The image segmentation technique is often used for street view image data processing. However, most deep learning methods require well-labeled data for model training. In addition, it needs a lot of human work to process data. This study presents a transfer learning approach for unlabeled onboard camera images and an image feature extraction approach utilizing image segmentation with the deeplabV2 model. We conducted mobile monitoring along Taiwan Avenue, Taichung, to develop a hybrid model to predict carbon dioxide (CO2), nitrogen oxides (NOx), black carbon (BC), and particle number (PN) concentration. The 5-fold cross-validation R2 for our model was 0.79, 0.88, 0.61, and 0.63 for the CO2, NOx, BC, and PN, respectively. This transfer learning method may be helpful when well-labeled data is difficult to acquire. Furthermore, this work enabled the hybrid model to adapt to different onboard camera scenarios and can be applied to estimate the on-road pollution exposure.