|Paper title||Exploring machine learning techniques to retrieve sea surface temperatures from passive microwave measurements with a focus on the Arctic|
|Form of presentation||Poster|
Accurate sea surface temperature (SST) observations are crucial for climate monitoring, understanding of air-sea interactions as well as in weather and sea ice forecasts through assimilation into ocean and atmospheric models. In general, two types of retrieval algorithms have been used to retrieve SST from passive microwave satellite observations: statistical algorithms and physical algorithms based on the inversion of a radiative transfer model (RTM). The physical algorithms are constrained by the accuracy of the RTM and the representativeness of the of the observation and prior error covariances. They can be used to identify measurement errors but require ad-hoc corrections of the geophysical retrievals to take these into account. Statistical algorithms may account for some of the measurement errors through the coefficient derivation process, but the retrievals are limited to the established relationships between the input variables. Machine learning (ML) algorithms may supplement or improve the existing retrieval algorithms through their higher flexibility and ability to recognize complex patterns in data.
In this study, several types of ML algorithms have been trained and tested on the global ESA SST CCI multi-sensor matchup dataset, with a focus on their performances in the Arctic region. The machine learning algorithms include two multilayer perceptron neural networks (NNs) and different types of ensemble algorithms e.g. a random forest algorithm and two boosting algorithms: least-squares boosting and the Extreme Gradient Boosting (XGB). The algorithms have been evaluated for their capability to retrieve SST from passive microwave (PMW) satellite observed brightness temperatures from the Advanced Microwave Scanning Radiometer – Earth Observing System (AMSR-E). To validate the algorithms independent SST observations from drifting buoys have been used. The performance of the ML algorithms has been compared and evaluated against the performance of an existing state of the art regression (RE) algorithm with a focus on the Arctic. In general, the ML algorithms show good global performances with decreasing performances towards higher latitudes. The XGB algorithm performs best in terms of bias and standard deviation followed by the NNs and the RE algorithm. The boosting algorithms and the NNs are able to reduce the bias in the Arctic compared to the other ML algorithms. For each of the ML algorithms, the sensitivity (i.e. the change in retrieved SST per unit change in the true SST) has been estimated for each matchup by using simulated brightness temperatures from the Wentz/DMI forward model. In general, the sensitivities are lower in the Arctic compared to the global averages. The highest sensitivities are found using the neural networks, and the lowest using the XGB algorithm, which underlines the importance of including sensitivity estimates when evaluating retrieval performances.
The good performance of the ML algorithms compared to the state of the art RE algorithm in this initial study demonstrates that there is a large potential in the use of ML techniques to retrieve SST from PMW observations. The ML methodology, where the algorithms select the important features based on the information in the training data, work well in complex problems where not all physical and/or instrumental effects are well determined. A suitable ML application could e.g. be in a commissioning phase of new satellites (e.g. for the Copernicus Imaging Microwave Radiometer (CIMR) developed by ESA).