9:00 AM - 10:30 AM
[ACG34-P01] Predicting dominant terrestrial biomes at a global scale: Assessments of machine learning algorithms, climate variables indexing, and extreme climate
Keywords:Potential Natural Vegetation, Machine Learning, Extreme Climate
Modeling biome distribution is a classic subject, and many methods have been proposed. In this research area, climate data was typically summarized as small numbers of climate indices, such as the coldest month mean temperature, and constructed models driven by these indices.
However, with the availability of machine learning algorithms in recent years, it is no longer essential to summarize environmental data into climatic indices. Besides, increasing the number of variables in models entails costs such as a decrease in model adaptability and an increase in the demands for computational power, so it would also be important to employ only a small number of variables that are selected in a parsimonious way. On the other hand, the intensity of incidents such as droughts and low temperatures that occur once every ten years (extreme climate) is a major factor limiting the biome boundaries.
Here, I compared four machine learning algorithms for accuracy to model relationships between the global biome distributions and climate characteristics. I also assessed the influence of converting monthly precipitation and average air temperature (24 variables) into 16 climatic indices on model accuracy. Finally, the influence of the inclusion of extreme climate indices on modeling accuracy was also assessed.
The Random Forest (RF) algorithm gives the best performance. In the RF applications: (1) simply employing 24 monthly climatic variables gives the best reconstruction performance (83.7%), indexing of climate variables significantly reduced the accuracies of reconstruction by 0.7%, and (3) adding extreme climate increases the accuracy significantly by +0.8% only when indexed climatic variables are used. These results suggest that monthly averaged climate tightly correlates with climate extremes at the global scale, and indexing it may eliminate the correlated information.
Reference
Sato, H., Ise, T. (2022) Predicting Global Terrestrial Biomes with Convolutional Neural Network. Geoscientific Model Development, 15, 3121-3132
However, with the availability of machine learning algorithms in recent years, it is no longer essential to summarize environmental data into climatic indices. Besides, increasing the number of variables in models entails costs such as a decrease in model adaptability and an increase in the demands for computational power, so it would also be important to employ only a small number of variables that are selected in a parsimonious way. On the other hand, the intensity of incidents such as droughts and low temperatures that occur once every ten years (extreme climate) is a major factor limiting the biome boundaries.
Here, I compared four machine learning algorithms for accuracy to model relationships between the global biome distributions and climate characteristics. I also assessed the influence of converting monthly precipitation and average air temperature (24 variables) into 16 climatic indices on model accuracy. Finally, the influence of the inclusion of extreme climate indices on modeling accuracy was also assessed.
The Random Forest (RF) algorithm gives the best performance. In the RF applications: (1) simply employing 24 monthly climatic variables gives the best reconstruction performance (83.7%), indexing of climate variables significantly reduced the accuracies of reconstruction by 0.7%, and (3) adding extreme climate increases the accuracy significantly by +0.8% only when indexed climatic variables are used. These results suggest that monthly averaged climate tightly correlates with climate extremes at the global scale, and indexing it may eliminate the correlated information.
Reference
Sato, H., Ise, T. (2022) Predicting Global Terrestrial Biomes with Convolutional Neural Network. Geoscientific Model Development, 15, 3121-3132