Exploring machine learning techniques to retrieve sea surface temperatures from passive microwave measurements

Abstract

Two machine learning (ML) models are investigated for retrieving sea surface temperature (SST) from passive microwave (PMW) satellite observations from the Advanced Microwave Scanning Radiometer – Earth Observing System (AMSR-E) and auxiliary data, such as ERA5 reanalysis data. The first model is the Extreme Gradient Boosting (XBG) model and the second is a multilayer perceptron neural network (NN). The performance of the two ML algorithms is compared to that of an existing state-of-the-art regression (RE) retrieval algorithm. The performance of the three algorithms is assessed using independent in situ SSTs from drifting buoys. Overall, the three models have similar biases; 0.01, 0.01 and −0.02 K for the XGB, NN and RE, respectively. The XGB model performs best with respect to standard deviation; 0.36 K. While the NN model performs slightly better than the RE model with respect to standard deviation, 0.50 and 0.55 K, respectively, the RE model is found to be more sensitive to changes in the in situ SST. Moreover, the XGB model is the least sensitive with an overall sensitivity of 0.78, compared to 0.90 for the RE model and 0.88 for the NN model. The good performance of the two 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 algorithms for the retrieval of SST from PMW satellite observations.