Spatiotemporal semantic network for ENSO forecasting over long time horizon

Abstract El Niño-Southern Oscillation (ENSO) has substantial influence on global climate variability and is responsible for extreme weather events such as drought and heavy rains, along with global ecosystem modifications. The successful prediction of ENSO is of considerable interest to reducing eco...
Ausführliche Beschreibung

Abstract El Niño-Southern Oscillation (ENSO) has substantial influence on global climate variability and is responsible for extreme weather events such as drought and heavy rains, along with global ecosystem modifications. The successful prediction of ENSO is of considerable interest to reducing economic and social adverse effects. Recently, deep learning models show great potential in this task. However, despite decades of efforts, predicting ENSO at lead times of more than one year remains a major challenge and most of these existing studies focus on the single channel of meteorological data, ignoring the spatial and temporal dependence of these factors. To capture the spatiotemporal information of meteorological factors as well as promote the skill of ENSO forecasting over longer time horizon, we propose an end-to-end deep learning model SpatioTemporal Semantic Network named STSNet, which consists of three main modules: (1) Geographic Semantic Enhancement Module (GSEM) distinguishes the geographic semantics of various latitudes and longitudes through a learnable adaptive weight matrix. (2) A novel SpatioTemporal Convolutional Module(STCM) is designed specially to extract the multidimensional features by alternating the execution of temporal and spatial convolution and temporal attention. (3) Multi-scale temporal information is combined and exploited in Three-stream Temporal Scale Module (3sTSM) to further enhance the performance. Integrating these modules gives a powerful feature extractor STSNet, which has multi-scale receptive fields across both spatial and temporal dimensions. In order to verify the effectiveness and progressiveness of our model, we execute experiments on Historical Climate Observation and Simulation Dataset. The results show that STSNet can simultaneously provide effective ENSO prediction for 16 months, with higher correlation and lower deviation compared with other deep learning models. Ausführliche Beschreibung