Estimating ground-level PM2.5 using micro-satellite images by a convolutional neural network and random forest approach

Published: 23 July 2020| Version 7 | DOI: 10.17632/n3ywbm3y2t.7
Contributor:

Description

Version 7 updates: Updated Image_Downloader again. Promise this will be the last recent update. Revised the part of image downloading from url and saving to directory. Sometimes, internal server error from Planet side can occur during the image downloading and saving process, which will cause the program to stop. Now this issue has been taken care of in this update. Some further remarks: To obtain the most possible matched images for a given query criteria, I recommend activating all your wanted images first, then wait for ~24 hours. This is the empirical time that roughly all the requested images can finish being activated. I am, however, strongly against downloading images right after activating images using this code. In such a case, you will lose a lot of valid images, as some of them will not get activated right away. Any questions can be directed to tongshu.zheng@duke.edu --- # based on the PlanetLabs' terms of use (https://www.planet.com/terms-of-use/) and Weather Underground’s terms of use (https://www.wunderground.com/company/legal), we are not allowed to publish/distribute the images data from PlanetLabs or meteorology data from Weather Underground used in this study. However, we have provided a python script (“Image_Downloader.ipynb”) that can download the current study’s images from PlanetLabs and we have provided links (“Links_to_Download_Meteorology_Data.docx”) to direct readers to the corresponding Weather Underground webpages where they can download the meteorology data used in our study.# This dataset includes the following items: 1) raw 1 h PM2.5 mass concentration measurements of the 35 regulatory AQM stations and the US Embassy station in Beijing from January 1, 2017 to July 20, 2019; 2) raw 1 h PM2.5 mass concentration measurements of the 10 regulatory AQM stations in Shanghai from January 1, 2017 to July 20, 2019; 3) "Links_to_Download_Meteorology_Data.docx" links to direct readers to the corresponding Weather Underground webpages where they can download the meteorology data used in our study; 4) "Image_Downloader.ipynb" python code to download the Beijing and Shanghai stations' images used in our study from PlanetLabs; 5) "PM_Meteorology_Image_process_filter_match.ipynb" python code to process the raw PM2.5 data and the raw meteorology data; to process and filter the raw images data; to match the processed PM2.5 data and meteorology data with the filtered and processed images data; store the matched records for model training and evaluation purposes; 6) "Model_Training_and_Evaluation_updated.ipynb" python code to build, train, and evaluate the CNN-RF models used in our study; 7) "670m_model.hdf5" the CNN model trained on the Beijing training dataset at a spatial resolution of 670 * 670 m; 8) "500m_model.hdf5" the CNN model trained on the Beijing training dataset at a spatial resolution of 500 * 500 m; 9) "200m_model.hdf5" the CNN model trained on the Beijing training dataset at a spatial resolution of 200 * 200 m.

Files

Institutions

Duke University

Categories

Computer Vision, Remote Sensing, Air Pollution Modeling, Deep Learning

Licence