Ji Luo 

Climate change and development of data-oriented methods are appealing for new climate classification schemes. Based on the most widely used Köppen-Geiger scheme, this article proposes a neural network based climate classification method from a data science perspective. In conventional schemes, empirically handcrafted rules are used to divide climate data into climate types, resulting in certain defects. In the proposed method, a machine learning mechanism is employed to do the task. Specifically, the method first trains a convolutional neural network to fit climate data to land cover conditions, then extracts features from the trained network and finally uses a self-organizing map to cluster land pixels on the extracted features. The method is applied to cluster global land represented by 66,501 pixels (each covers 0.5 latitude degree × 0.5 longitude degree) using 2020 land cover data and 1991-2020 climate normals, and a 4 × 3 × 2 hexagonal self-organizing map clusters the land pixels into twenty-four climate types. By Kappa statistics, the obtained scheme shows good agreement with the Köppen-Geiger and Köppen-Trewartha schemes. In addition, our scheme addresses some issues of the Köppen schemes, suggests new climate types such as As (severe dry-wet season) and Fw (arctic desert), and identifies the highland group H without input of elevation. The proposed method is expected as an intelligent tool to monitor changes in the global climate pattern and to discover new climate types of interest that possibly emerge in the future. It may also be valuable for bio-ecology communities.