Overview

This repo introduces two methods to estimate the air quality from satellite imagery.

• estimation of AOD values based on 6S radiative transfer model, or
• directly PM2.5/PM2.10 estimation from TOA reflectance

There are currently two case study, one based on Landsat-8 satellite data and the other based on Zhuhai-1 hyperspectral data.

Directly estimate PM2.5/PM10 from TOA reflectance using Zhuhai-1 data

Using the pretrained model to estimate PMx concentration

To use the pretrained model in demo_predict.py, you may change the arguments directly in the script or use the inference below (plz change inference to True before usage):

python demo_predict.py --im 'G:/orbita/data_32band/HAM1_20181006215942_0013_L1_MSS_CCD1.tif' --model model_pmnet_v0.1.h5 --m 10 --ang1 0.4 --ang2 50.1 --lat 28.8 --lng 115.6

Here are the parameters availble for inference, you may find them in the header pdf of Zhuhai-1 raw data

--im        Zhuhai-1 hyperspectral image location
--model     pretrained model location
--m         month
--ang1      off-nadir angle, 侧摆角
--ang2      solar zenith angle, 太阳高度角
--lat       centered latitude of the hyperspectral image
--lng       centered longitude of the hyperspectral image

The input image is Zhuhai-1 TOA reflectance hyperspectral image in 32 bands.

To convert the raw Zhuhai-1 hyperspectral data to TOA reflectance, please refer to:

https://github.com/sjliu68/Orbita-Hyperspectral-Imagery-to-Top-of-Atmosphere-Reflectance

Training

For this method, we need the following data to build the model

• ground truth or site-based PM2.5/PM10 data (as label)
• a large amount of satellite imagery representing TOA reflectance (as variables)

After collection of the data,

• run extract_points.py to create a excel table, containing the value of PM2.5 of each record and its corresponding satellite image completed
• train.py to train a model todo
• demo_predict.py to estimate PM2.5 based on the input image based on the pretrained model completed

Based on 6S model to estimate AOD

Note: this part is in experiment

• First, one needs to create a lookup table (LUT) from the 6S model. This Python version of 6S emulator is strongly recommended: https://github.com/samsammurphy/6S_emulator
• After creating the lookup table, run estimate_aod.py based on your Landsat-8 data