Remote sensing of cloud by deep neural network: Training strategies

Abstract:

Machine learning has been applied in many problems of remote sensing of atmosphere, ocean, and land. Recent machine learning techniques show very high ability to represent a relationship between observation 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 observation systems and a demand for high accuracy of estimation. The authors have applied deep neural networks (DNNs) to cloud retrieval from 1) a ground-mounted camera, and 2) GCOM-C and 3) Himawari-8 satellites. With a multispectral image as input, DNN estimates a spatial distribution of cloud properties fusing multi-scale, spatial and spectroscopic features. We need a set of target cloud data and corresponding observation data to train the DNN. Spatial distribution of cloud properties such as cloud extinction coefficient and cloud particle size should be given. This kind of dataset can be from physics-based (realistic) simulations or independent observations that are assumed to be correct. Prior of cloud and radiative transfer should be learned by machine in the training process. Remote sensing requires appropriate consideration of noises induced by sensor and postprocessing of digital data and uncertainty in radiometric calibration. If noise properties are known, simulated noises can be embedded in the training data. Taking calibration uncertainty of known size into account, a random bias can be superimposed on each training data sample. This makes the DNN robust to calibration uncertainty. Being different from traditional retrieval using data at an image pixel or a time step, DNN can extract spatial/temporal features, which enables overcoming the calibration uncertainty. With more uncertainty in measurement, DNN should more heavily rely on spatial features, being more insensitive to absolute calibration. As DNN has very high representation ability, DNN retrievals can be statistically accurate. However, many questions have been raised on reliability, robustness, and ability to explain observation signals. Further studies are needed to diagnose and quantify uncertainties in retrieval, in addition to independent validation experiments.