Evaluating the Performance of Convolutional Neural Network (CNN) Model for Temperature Simulation-A Case Study in Fujian Province

Climate change has profound potential impacts on the environment, economy, and human health. Predicting future temperature change can assist decision-makers in understanding the overall patterns of climate change and related risks of extreme weather occurrences. In this study, a convolutional neural network (CNN) model is developed for establishing the quantitative relationship between atmospheric circulation factors and temperature (including maximum, minimum, and average temperature). To evaluate the performance of CNN model, three indicators are introduced, which are root mean squared error (RMSE), coefficient of determination (R2), and percentage bias (PBs). Then, CNN is applied to Fujian Province where 27 stations are selected. Results show that CNN model can achieve robust performance at each station. For maximum temperature, the R2 values of 4 stations are above 0.9, and the 23 stations are above 0.85. For the minimum temperature, the R2 values of 20 stations are above 0.9. For average temperature, the R2 values of the 27 stations are all above 0.9. From the perspective of spatial distribution, the model fits better for maximum temperature in the central and western inland areas of Fujian Province than in the coastal areas in the east. In contrast, the model fits better for minimum temperature is in the coastal areas in the east, and for average temperature is in the southeastern region. Finding will be helpful for predicting future temperature change in Fujian Province, thus facilitating to mitigate the negative impacts of extreme weather events.