Regional climate fluctuation analysis using convolutional neural networks

Regional climate classification aids the investigation of the causes of changes in natural vegetation distribution and allows the selection of appropriate crops under climate fluctuations. In this study, the Japanese climate was classified using a simple convolutional network (CNN) into nine regional areas based on meteorological factors (channels). One dataset of each channel was processed by an arrangement into two dimensions of 12 months and 10 years. Combinations of five channels were used by the CNN to search for the best combination for climate classification. A combination of four channels, excluding snow depth data, showed the best test accuracy. Regional climate change was analyzed by comparing the different patterns between the latest and former decades. The climate in most regions tended to shift towards the north. However, the number of regions that shifted towards north decreased in the most recent decade compared with those in previous decades, indicating that Japanese climate is generally oriented southward. The simple convolutional network based on the processed two-dimensional data from the meteorological time-series dataset enabled recent climate change evaluation and predicted regional climate change, which could help decision makers for choosing crops and formulating disaster management strategies in the near future.