Cloud and Aerosol Detection using Spectral, Spatial and Temporal Information from Passive Satellite Instruments
Abstract
Detection of atmospheric constituents such as cloud and aerosols with satellite observations is often a critical initial step in many remote sensing algorithms. Many traditional algorithms were developed based on underlying physics and hand tuned thresholds. Weakness of these methods is that it is challenging and time-consuming to develop algorithms across multiple instruments and locations. To overcome this weakness, we designed and trained a couple of Machine Learning (ML) based models for cloud and aerosol detection. Specifically, a Random Forest (RF) model, a Convolutional based Encoder-Decoder Neural Network, and a Recurrent Neural Network (RNN) are designed and trained with spectral, spectral/spatial, and spectral/spatial/temporal input, respectively. The first two models that require spectral and spatial input are designed for passive spectrometers such as VIIRS and MODIS; while the third model is designed for geostationary instruments, such as GOES-16/17 ABI. A hybrid training database is generated based on years of collocated satellite-satellite (e.g., CALIPSO/VIIRS and CALIPSO/ABI) and satellite-ground station (e.g., ABI/AERONET) data, and manually picked events with labels. In this presentation, we will introduce of the training database and compare the three different models.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.A52D..03W