[A-CG43] Earth & Environmental Sciences and Artificial Intelligence/Machine Learning

[ACG43-08] Estimation of cloud and meteorological state from satellite image by deep neural network

★Invited Papers

*Hironobu Iwabuchi¹, Xinyue Wang¹, Hana Kato¹, Takaya Yamashita¹ (1.Graduate School of Science, Tohoku University)

Keywords:deep learning, satellite earth observation, cloud

Machine learning has been applied to satellite remote sensing of atmosphere, ocean, and land. Recent machine learning techniques show high ability to represent a relationship between satellite measurement signals and an atmosphere/surface state utilizing rich information contained in high-dimensional, spatial/temporal/spectral measurement data. This well meets both a requirement of processing big data available from recent satellite observation systems and a demand for high accuracy of estimation. Popular deep neural networks are good at analysis of satellite measurement image, effectively using spatial/temporal features contained in the image or a sequence of images. If we use time series of satellite images, information about temporal variation of meteorological state can be well obtained. This is a big difference from previous satellite data analysis because traditional satellite remote sensing uses data at an image pixel at a time step. The authors have applied deep neural networks to cloud retrieval from GCOM-C and Himawari-8 satellites. With a multispectral image as input, a deep neural network estimates a spatial distribution of cloud properties fusing multi-scale, spatial and spectroscopic features. We show two different training approaches. One makes a training dataset from physics-based (realistic) simulations, and the other takes independent observations that are assumed to be correct. Prior of cloud and radiative transfer imprinted in multispectral images should be learned by machine in the training process. High accuracy of estimation is shown for cloud properties from visible or thermal infrared channels. Infrared-only method is particularly attractive because it can be used uniformly irrespective to sunlight. We show not only cloud properties but also meteorological states such as wind and humidity can be estimated from a snapshot or time series of Himawari-8 infrared image. An advantage of deep learning is the usage of spatial and temporal features for estimation/prediction of atmospheric properties, which is a new direction for the community.