基于决策树分类的橡胶林地遥感识别

摘要

中老缅交界地区是橡胶林地的主要种植区，利用遥感手段快速动态监测橡胶林地的时空变化，对于橡胶合理种植、生态环境保护以及边境安全保障具有重要的科学价值和实践意义。研究基于 Landsat 数据和MODIS-NDVI数据，采用决策树分类的方法提取中老缅交界地区的橡胶林地。研究发现：1）1月上旬至3月下旬为提取橡胶林地的主要时间窗口；根据橡胶林不同树龄所表现的光谱差异，按照橡胶幼林（<10 a）和橡胶成林（≥10 a）提取橡胶林地；橡胶成林、高植被覆盖度的旱地、有林地容易发生误分，橡胶幼林、茶园、灌木林地和草地容易发生混淆。2）基于原始光谱特征、归一化指数、K-T变换指数以及纹理特征分别构建橡胶幼林和橡胶成林决策树分类模型；2010年橡胶成林分类总精度超过90%，橡胶幼林分类总精度超过75%；对同一地区1980、1990、2000年3个时相的决策树分类发现，橡胶幼林和橡胶成林决策树分类模型简单有效，结合晚期的橡胶成林来验证提取早期的橡胶幼林可以达到更高的分类精度。3）1980-2010年间，中老缅交界地区橡胶林地由7.05万hm2增至60.14万hm2，橡胶林地扩张趋势显著；橡胶幼林扩张速度明显快于橡胶成林，特别是近10 a来；西双版纳橡胶种植面积在中老缅交界地区占主导地位，橡胶林地不断向老挝、缅甸边境地区扩张。%The border region of China, Laos and Myanmar (BRCLM) has attracted much international attention due to the special geo-economic and geo-political characteristics, as well as being the hinterland of the world’s famous “Golden Triangle,” and the optimal rubber planting areas for Chinese investment. Monitoring the spatial-temporal pattern of the rubber plantations is significant for regional land resource development, eco-environment protection, and maintaining border security. Based on Landsat remote sensing image data and MODIS-NDVI data, rubber plantations were extracted by the decision tree classification method in BRCLM using spectral features and texture characteristics. The results showed that: (1) On account of spectral differences between rubber forests at different growth stages, we were able to extract rubber plantations according to young rubber forest (<10 a) and mature rubber forest (≥10 a) respectively. The optimum temporal window to discriminate rubber plantations was from early January to late March, which is especially appropriate for mature rubber forest. Mature rubber forest, dry land with high vegetation cover, and forest land were prone to misclassification. Meanwhile, young rubber forest, tea plantation, shrubland and grassland were confused with each type in spectral characteristics according to the index of NDVI. (2) Based on the original spectral characteristics, normalized indices, K-T transform indices, and texture features, we established young rubber forest and mature rubber forest decision tree classification models respectively. The overall accuracy of the mature rubber forest went beyond 90%, and the young rubber forest beyond 75%, which meant that the decision tree method was better for mature rubber forest extraction. The rubber plantation distribution maps were obtained using the established decision tree models in 1980, 1990, and 2000 with high classification accuracy, which indicated that the models were simple and efficient for extracting rubber plantations in the tropical areas. This is an effective method for perennial vegetation extraction and classification accuracy verification. (3) From 1980 to 2010, the size of rubber plantations in BRCLM increased nearly nine times, from 705 km2 to 6 014 km2, and the expansion rate of the young rubber forest was faster than that of the mature rubber forest. National differences of rubber plantations in BRCLM were significant, and the cross-border planting developed quickly in the recent 10 years. Rubber plantations in BRCLM will definitely expand across the borders of China to the territories of Laos and Myanmar.