| Since the late 19th century,almost all regions of the world have experienced a general warming trend.Most of the observed global warming is mainly caused by carbon emissions generated by human activities such as deforestation and burning fossil fuels,which will pose a great threat to human survival.Agriculture in many developing countries relies on rainwater irrigation,with a weak industrial base and underdeveloped infrastructure,making them highly vulnerable to the impact of climate change.Especially in many developing countries,the economic strength is relatively backward,with only sparse meteorological observation stations,it is far from enough to use only the observation data obtained from sparse and irregular meteorological stations to produce high-resolution climate maps to assess the impact of climate change and mitigate climate disasters in developing countries.Currently,the mainstream geostatistical downscaling technology is spatial interpolation or multiple linear regression models to extract linear relationships between large-scale and local climate variables,and then generate high-resolution climate maps from sparse climate observations.Because global climate change is a nonlinear process controlled by complex physical principles,these linear downscaling techniques cannot achieve the expected accuracy.The latest machine learning technologies(such as SRCNN and SRGAN)can extract nonlinear relationships,but they are only suitable for downscaling low resolution lattice data,and cannot use relationships with other climate variables to improve the performance of downscaling.Also,these techniques cannot utilize data of different resolutions as model input.In this study,we developed a hybrid network downscaling method that can well couple high-resolution terrain data with sparse climate observation data,and then generate highresolution climate maps.Sub Saharan Africa has been deeply affected by climate change,which not only gradually increases air temperature,but also leads to changes in precipitation intensity,and extreme events such as drought and floods are also gradually increasing.Ethiopia and Bangladesh arc considered to be one of the countries most severely affected by climate change in the world,and these two countries arc relatively representative,so this study focuses on Ethiopia and Bangladesh.To test the performance of our model,we generated high-resolution climate maps using sparse observational data from 21 meteorological stations in Ethiopia and 34 meteorological stations in Bangladesh.The accuracy of high-resolution climate maps generated using our composite network is significantly superior to those generated using multiple linear regression models or classic network models.In the context of spatial changes in air temperature and precipitation,developing countries urgently need climate big data to cope with meteorological disasters such as floods and droughts.The hybrid network proposed in this study can actively help developing countries achieve sustainable development goals. |
