講演情報
[U16-P05]モンゴルにおける機械学習を用いた樹冠被覆の高精度マッピング
*堀 聡悟1、石川 守2 (1.北海道大学大学院環境科学院、2.北海道大学大学院地球環境科学研究院)
キーワード：
モンゴル、機械学習、樹冠被覆、GIS、タイガ
Within the boundary zones between taiga and steppe, taiga–steppe ecotone (TSE), the tree coverages are highly varied in space. The TSE definition (e.g., percentages of tree area), therefore, still remains ambiguous. TSEs often overlap with the land modified by anthropogenic activities (i.e., cropland), which are other factors complicating TSE delineation. Require for quantifying TSE in order to predict TSE shift and alteration by climate change, especially for the regions where human impacts are negligible. Mongolia, located under semi-arid climatic condition having less cropland, is ideal research area for understanding interaction between TSE and climate change. This talk introduces quantification of the TSE as high-accuracy tree-canopy cover map of Ulaanbaatar and its surroundings, northeastern Mongolia. The map will be used for evaluating the environmental factors that control spatiality of tree canopy cover.
We quantified the TSE as areal coverage of tree canopy (%TC) by calibrating the Global Forest Cover Change (GFCC) tree-cover product (2015, 30×30 m), a relatively high-resolution, openly available global dataset that provides %TC, to our research area using machine learning as support vector regression (SVR) and isotonic regression (IR). The area was subdivided into four adjacent subregions that depend on dominant land cover and climatic gradients: NE (north-eastern; taiga-dominated), NW (north-western; transitional), SE (south-eastern; transitional), and SW (south-western; steppe/bare-dominated). Training and test datasets were obtained by visual interpretation of Google Earth imagery (GE reference points, 324 points) and onsite observation in SE and SW subregions (onsite reference points, 472 points). TSEs were considered to be either 5–20%TC (TR1), or 1–4% adjacent to pixels with 5% or higher (TR2). They were extracted from the original GFCC dataset (GFCC-TC), the SVR-corrected dataset (SVR-TC), and the IR-corrected dataset (IR-TC).
The study-area-wide SVR-TC (94 GE reference points from NE and NW) correlated well with the GE reference points (R² = 0.79), yet it did not agree with the onsite reference points (R² < 0). The subregion-specific SVR-TC and IR-TC (all GE reference points within each subregion) also showed generally good correlation with the GE reference points (R² = 0.66–0.72). To test whether the GFCC bias varies across subregions, the bias of GFCC against the GE reference points was calculated and evaluated among subregions using the Kruskal–Wallis test and Dunn’s post hoc test. Significant differences were detected (H = 16.64, df = 3, p < 0.001), particularly between the NE and NW subregions (p < 0.005) and between the NE and SW subregions (p < 0.01). In SVR-TC, the area of TR1 was 6.7–37.1% larger than that of GFCC-TC, mainly due to transitions from the low-%TC class TR2 and from treeless areas. In IR-TC, the area of TR1 was 4.8% smaller than that of GFCC-TC in the NW subregion, whereas it was 4.7–10.3% larger in the other subregions.
The differences observed among these subregions suggest that factors influencing tree canopy cover (or the optical observations used by GFCC) may vary regionally. In subregions where TR1/TR2 accounted for a large proportion in GFCC-TC, the increase rate of TR1 area varied markedly depending on the correction function. This implies that, in TSE-rich areas, small differences in correction amplified transitions from low %TC (0–5%) to TR1. It further suggests that high variability in datasets within the low-%TC range can distort the estimated spatial distribution of the TSE.
As a next step, exploration of environmental factors that control tree canopy cover will be conducted using the corrected canopy cover datasets and machine-learning models. To evaluate the importance of environmental factors more accurately, downscaling of meteorological variables will be performed as a preprocessing step before modeling.
We quantified the TSE as areal coverage of tree canopy (%TC) by calibrating the Global Forest Cover Change (GFCC) tree-cover product (2015, 30×30 m), a relatively high-resolution, openly available global dataset that provides %TC, to our research area using machine learning as support vector regression (SVR) and isotonic regression (IR). The area was subdivided into four adjacent subregions that depend on dominant land cover and climatic gradients: NE (north-eastern; taiga-dominated), NW (north-western; transitional), SE (south-eastern; transitional), and SW (south-western; steppe/bare-dominated). Training and test datasets were obtained by visual interpretation of Google Earth imagery (GE reference points, 324 points) and onsite observation in SE and SW subregions (onsite reference points, 472 points). TSEs were considered to be either 5–20%TC (TR1), or 1–4% adjacent to pixels with 5% or higher (TR2). They were extracted from the original GFCC dataset (GFCC-TC), the SVR-corrected dataset (SVR-TC), and the IR-corrected dataset (IR-TC).
The study-area-wide SVR-TC (94 GE reference points from NE and NW) correlated well with the GE reference points (R² = 0.79), yet it did not agree with the onsite reference points (R² < 0). The subregion-specific SVR-TC and IR-TC (all GE reference points within each subregion) also showed generally good correlation with the GE reference points (R² = 0.66–0.72). To test whether the GFCC bias varies across subregions, the bias of GFCC against the GE reference points was calculated and evaluated among subregions using the Kruskal–Wallis test and Dunn’s post hoc test. Significant differences were detected (H = 16.64, df = 3, p < 0.001), particularly between the NE and NW subregions (p < 0.005) and between the NE and SW subregions (p < 0.01). In SVR-TC, the area of TR1 was 6.7–37.1% larger than that of GFCC-TC, mainly due to transitions from the low-%TC class TR2 and from treeless areas. In IR-TC, the area of TR1 was 4.8% smaller than that of GFCC-TC in the NW subregion, whereas it was 4.7–10.3% larger in the other subregions.
The differences observed among these subregions suggest that factors influencing tree canopy cover (or the optical observations used by GFCC) may vary regionally. In subregions where TR1/TR2 accounted for a large proportion in GFCC-TC, the increase rate of TR1 area varied markedly depending on the correction function. This implies that, in TSE-rich areas, small differences in correction amplified transitions from low %TC (0–5%) to TR1. It further suggests that high variability in datasets within the low-%TC range can distort the estimated spatial distribution of the TSE.
As a next step, exploration of environmental factors that control tree canopy cover will be conducted using the corrected canopy cover datasets and machine-learning models. To evaluate the importance of environmental factors more accurately, downscaling of meteorological variables will be performed as a preprocessing step before modeling.