Polygonization of discontinuous raster classes from machine-learning predictive ecosystem mapping (PEM)

Biogeoclimatic Ecosystem Classification (BEC) has been applied extensively in characterizing forested ecosystems in British Columbia. With a lack of qualified vectorization method used for BEC data transformation, the main goal of this research is to polygonize discontinuous BEC raster classes into vector map with better overall effectiveness and efficiency especially regarding the linear areas. The original data input for analysis is a machine-learning BEC zone raster map of Deception Study Area located in middle BC near Telkwa, with a resolution of 5m*5m. A comprehensive comparison between vectorization algorithms in GIS applications was conducted, including different filtering, simplifying and smoothing algorithms. Since we have the original predicted BEC raster map as the performance measurement, accuracy was directly measured as the percentage of correctly classified pixels when rasterizing the polygons. The evaluation criteria include visual effect, number of polygons, linear patches accuracy processing time. We found an appropriate vectorization routine to polygonize the classification raster maps. The polygonal map using Scenario D has overall satisfactory effectiveness and efficiency with a 46% linear patch accuracy and 62,014 polygons. The method also provides good approximations of the areas with moderate processing time. This is partly because we allow vertices to be located anywhere and not just exactly on the boundary of the original raster zones. We can promote this polygonization method in future predicted ecosystem mapping (PEM) product with similar linear and discontinuous areas. Priority of several key BEC zone classification with importance level regarding to the ecosystem condition related to endangered species can be further explored and added to the algorithms to better polygonize those areas in future studies.

生物地理气候生态系统分类（Biogeoclimatic Ecosystem Classification, BEC）已被广泛应用于不列颠哥伦比亚省森林生态系统的特征刻画工作中。由于当前缺乏适用于BEC数据转换的合格矢量化方法，本研究的核心目标是将非连续的BEC栅格类别多边形化为矢量地图，以期在整体有效性与效率上获得提升，尤其针对线性区域而言。本次分析的原始输入数据为位于不列颠哥伦比亚省中部特尔克瓦（Telkwa）附近的德塞普申（Deception）研究区的机器学习BEC分区栅格地图，空间分辨率为5米×5米。研究针对地理信息系统（Geographic Information System, GIS）应用中的多种矢量化算法开展了全面对比，涵盖不同的滤波、简化与平滑算法。由于以原始预测得到的BEC栅格地图作为性能评估基准，模型精度直接通过将多边形栅格化后分类正确的像素占比来衡量。本次评估的指标包括视觉效果、多边形数量、线性斑块精度与处理时长。研究最终找到了一套适配的矢量化流程，可用于将分类栅格地图多边形化。采用方案D（Scenario D）生成的多边形地图在整体有效性与效率上表现良好，其线性斑块精度达46%，多边形总数为62014个。该方法在中等处理时长下，也能对各类区域实现较好的近似还原。这一效果的部分原因在于，研究允许顶点可置于任意位置，而非仅限定于原始栅格分区的边界之上。后续可将该多边形化方法推广应用于具备相似线性与非连续区域特征的预测生态系统制图（Predicted Ecosystem Mapping, PEM）产品中。未来研究中，可进一步探索与濒危物种相关的生态系统状况挂钩的关键BEC分区分类优先级，并将其纳入算法体系，以实现对这类区域更精准的多边形化处理。