Prediction of temperature anomaly in Indian Ocean based on autoregressive long short-term memory neural network.

Surface temperature is one of the first ocean variables investigated. Ocean temperature is a key indicator of global climate change. The anomalies in ocean temperature have caused significant deterioration of marine systems. Existing works on surface temperature anomaly considered a suite of other remote sensing measurements such as wave height, salinity and models such as support vector machine, self-organizing maps and convolutional neural networks. Neural networks are used for predicting the surface temperature. This paper proposes use of long short-term memory, a recurrent neural network method to the estimate sea surface temperature anomalies based on previous year's sea surface temperature anomalies. NOAA OI SST V2 dataset with 40 years of data is used in the experimentation. Auto-regression is used during data preprocessing. The basic LSTM method with 3 blocks of size 2 is enhanced to 50 neurons. This proposed LSTM model has been assessed for performance on time series data, yearly wise and for the entire dataset and found that the model has been able to predict the anomalies with a reasonably good precision. The model produced error of 0.036 indicating that the model is feasible for predicting the temperature anomaly and mean absolute error of 0.14 on the testing data. [ABSTRACT FROM AUTHOR]