Case Study: Development of the CNN Model Considering Teleconnection for Spatial Downscaling of Precipitation in a Climate Change Scenario

Abstract

Global climate models (GCMs) are used to analyze future climate change. However, the observed data of a specified region may differ significantly from the model since the GCM data are simulated on a global scale. To solve this problem, previous studies have used downscaling methods such as quantile mapping (QM) to correct bias in GCM precipitation. However, this method cannot be considered when certain variables affect the observation data. Therefore, the aim of this study is to propose a novel method that uses a convolution neural network (CNN) considering teleconnection. This new method considers how the global climate phenomena affect the precipitation data of a target area. In addition, various meteorological variables related to precipitation were used as explanatory variables for the CNN model. In this study, QM and the CNN models were applied to calibrate the spatial bias of GCM data for three precipitation stations in Korea (Incheon, Seoul, and Suwon), and the results were compared. According to the results, the QM method effectively corrected the range of precipitation, but the pattern of precipitation was the same at the three stations. Meanwhile, for the CNN model, the range and pattern of precipitation were corrected better than the QM method. The quantitative evaluation selected the optimal downscaling model, and the CNN model had the best performance (correlation coefficient (CC): 69% on average, root mean squared error (RMSE): 117 mm on average). Therefore, the new method suggested in this study is expected to have high utility in forecasting climate change. Finally, as a result of forecasting for future precipitation in 2100 via the CNN model, the average annual rainfall increased by 17% on average compared to the reference data.

Deep Learning for Climate Downscaling: Generating high-resolution gridded temperature projections over India from low-resolution CMIP6 data

“…Nourani et al (2019) also used a larger geographic area for predictors than the geographic extent of the predictand to maximise spatial learning and potentially boost downscaling performance. This idea was tested more thoroughly by Kim et al (2022) where they identified most relevant regions for plausible teleconnections (climatic influences from remote regions) and included them into their training dataset. More recent studies have tested convolutional neural networks (CNNs) that rely on spatial learning and can read climatic data in form of 2D images (Vandal et al, 2019;Liu et al, 2020;Baño-Medina et al, 2021;Damiani et al, 2024) for climate downscaling.…”

Introducing Variations of Predictors as Optional Predictors Offers the Potential to Improve the Downscaling Performance of Geographically Weighted Regression Model

“…Therefore, satellite products with a high temporal resolution are often required to be spatially downscaled (disaggregation of a coarse cell into many finer cells) for various environmental applications. In the past 10 years, numerous spatial downscaling studies have been conducted on many satellite-derived products, such as precipitation [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], soil moisture [15], [16], [17], [18], [19], [20], [21], [22], [23], land surface temperature [2], [24], [25], [26], [27], [28], [29], [30], night-time light [31], solar radiation [32], evapotranspiration [33], [34], [35], chlorophyll [36], and wind speed [37]. The primary goal of spatial downscaling research is to improve the downscaling performance of satellite-derived products which is generally performed from two main aspects [38]: the introduction of new auxiliary variables [5], [8], [39], [40] and the development of new downscaling models [6], [13], [22],…”

A Multi-Dimensional Deep Siamese Network for Land Cover Change Detection in Bi-Temporal Hyperspectral Imagery

“…So far, many algorithms have been developed for HCD [20]. It is widely reported that Deep Learning (DL) algorithms have the highest accuracies [21][22][23]. A DL approach can automate the learning of features on input data at several abstraction levels [24,25].…”