09:30 〜 09:45
[ACG39-03] Machine learning prediction of wildfire over the Republic of Sakha, Russia
キーワード：森林火災、機械学習、シベリア、ロシア、予測、北極
In this study, we carried out machine learning (ML) prediction for the nation-averaged monthly mean datasets over the Republic of Sakha in Russia during 2003 and 2019. We used the Aqua and Terra MODIS averaged fire pixel count (FPC; https://feer.gsfc.nasa.gov/) data for objective variables. For explanatory variables, we selected the Aqua MODIS Snow Cover Fraction (SCF; https://nsidc.org/data/MYD10CM) and many atmospheric and land variables from NASA’s MERRA-2 (Modern-Era Retrospective analysis for Research and Applications, Version 2: https://gmao.gsfc.nasa.gov/reanalysis/MERRA-2/) re-analysis data (total 170 variables). The FPC and SCF data were re-gridded to the MERRA-2 grid spacing in advance. All the data were then averaged over the Republic of Sakha with the nation mask data by Natural Earth, using SPRING (SPeroidal coordinates Regridding Interpolation table Generator: http://hydro.iis.u-tokyo.ac.jp/~akira/page/tmp/2020Feb_nations/). To assess the impact of removing small-scale fires on prediction performance, we prepared four fire datasets: all the data, and data only considering FPC equal to or greater than 5, 10, 20 counts per MERRA-2 grid, M2G, in which we treated the months with no corresponding fire data as zero fire. For the explanatory variables, we also made the time-lagged data from zero to six months for each variable to take pre-fire information into account, which was then all standardized (i.e., we used a total of 1190 explanatory variables for ML). We implemented ML prediction by three Linear Regression (LR) methods (ordinary LR; penalized LR: Lasso and Ridge) and three non-linear ensemble methods (Extra Trees, Random Forest, and Gradient Boosting). The data during July 2003 and December 2014 and during January 2015 and December 2019 were used for training and test, respectively.
Both the ordinary LR and Ridge methods did not predict the test data at all. Those methods likely induced overfitting to the training data because of too high R2 scores. Lasso mitigated the overfitting to the training data and improved the test data’s performance, but the scores were still not increased (R2 of 0.45-0.48). All the non-linear ensemble methods showed much better results than those of the linear forms. Significantly, Random Forest and Gradient Boosting methods outperformed the others. The highest R2 score was 0.75 by Gradient Boosting with the fire data, including FPC equal to or greater than ten counts/M2G. In this case, the permutation importance showed the top 5 variables dominant in the following order: “cloud area fraction for middle clouds” (lag 0), “transcom planetary boundary layer height” (lag 1), “lifting condensation level” (lag 0), “fractional area of wilting zone” (lag 0), and “in cloud optical thickness of low clouds” (lag 0). In the non-linear methods, the fire data, including all the fire data, showed worse performance than the other fire data, which was similar to what Yasunari et al. (in preparation) discussed over a large area of Siberia.
Both the ordinary LR and Ridge methods did not predict the test data at all. Those methods likely induced overfitting to the training data because of too high R2 scores. Lasso mitigated the overfitting to the training data and improved the test data’s performance, but the scores were still not increased (R2 of 0.45-0.48). All the non-linear ensemble methods showed much better results than those of the linear forms. Significantly, Random Forest and Gradient Boosting methods outperformed the others. The highest R2 score was 0.75 by Gradient Boosting with the fire data, including FPC equal to or greater than ten counts/M2G. In this case, the permutation importance showed the top 5 variables dominant in the following order: “cloud area fraction for middle clouds” (lag 0), “transcom planetary boundary layer height” (lag 1), “lifting condensation level” (lag 0), “fractional area of wilting zone” (lag 0), and “in cloud optical thickness of low clouds” (lag 0). In the non-linear methods, the fire data, including all the fire data, showed worse performance than the other fire data, which was similar to what Yasunari et al. (in preparation) discussed over a large area of Siberia.