基于SNIC-CNN-SVM模型的京津风沙源二期工程区土地利用/土地覆盖遥感识别研究

稀疏植被覆盖（草地、沙地、戈壁）演变能够直接表征区域生态环境和人类活动的动态影响变化。但由于大尺度稀疏植被区一般都具有地理跨度大，景观结构复杂多样，破碎化程度高，现有地表覆盖分类产品针对性不足等问题，使得该区域内林草沙的遥感提取难度较大，精度普遍偏低，直接制约生态效应评价模型的应用效果。因此，以典型大尺度稀疏植被区--京津风沙源治理二期工程区为研究区，研建了SNIC-CNN-SVM （SCS）模型，实现了大尺度稀疏植被区林草沙典型要素的信息自动提取和主要土地利用/土地覆盖类型识别。研究结果表明：1）引入惩罚性机制优化后的SNIC分割算法，有效提升了稀疏植被区与沙地区的边界区分度，有助于分类精度的提升；2）基于改进SNIC-CNN-SVM模型方案的研究区总体分类精度达89.41%，较优化前提高了11.17%，特别是乔、灌、草、沙地和戈壁的分类识别精度显著提升，表明该优化方案在以研究区为代表的稀疏植被区域分类中具有较好的应用效果和推广价值；3）分类结果显示，2020年工程区草地面积最大，占到了一半以上（51.52%），沙地占比11.96%，稀疏植被覆盖（草地、沙地、戈壁）区域占比68.68%，表明工程区处在林地-稀疏植被-沙地的过渡地带，生态环境保护压力与防沙治沙形势依然严峻；4）近20年来，乔灌草等植被增加面积约占工程区20.64%，主要由沙化土地转化，沙化土地减少面积约占工程区的4.58%，表明研究区植被状况不断改善，实施的各项生态工程作用显著，能够更有效地服务于多维度生态系统服务功能评价。该研究以期能够为京津风沙源二期工程区的生态系统演变规律研究及生态工程评价等工作提供重要科学支撑。

Remote sensing recognition of land use/land cover in the Beijing-Tianjin sandstorm source region based on SNIC-CNN-SVM model

The evolution of sparse vegetation cover (grassland, sandy land, Gobi) can directly characterize the dynamic changes of the regional ecological environment and human activities. However, large-scale sparse vegetation areas generally have large geographic spans, complex and diverse landscape structures, high fragmentation, and insufficient pertinence of existing land cover classification products. These problems make it difficult to extract the forest, grass, and sand in this area by remote sensing, and the accuracy is generally low, which directly restrict the application results of the ecological effect evaluation model. Therefore, this study took the typical large-scale sparse vegetation area, the Beijing-Tianjin sandstorm source region (BTSSR) as the study area, and the SNIC-CNN-SVM (SCS) model is proposed. The forest, grass, sand element information is automatically extracted and the land use/land cover types are identified in the study area. The results show that:1) the SNIC segmentation algorithm was optimized by the penalty mechanism and can effectively increase the boundary discrimination between the sparse vegetation area and the sandy area, which is helpful to improve the classification accuracy; 2) the overall classification accuracy of the optimized model reached 89.41%, which is 11.17% higher than that before optimization, especially the classification and recognition accuracy of trees, shrubs, grasses, sandy land and Gobi is significantly improved, indicating that the optimization model has good application effect and promotion value in the classification of sparse vegetation areas represented by the study area; 3) in 2020, the grassland area in the study area was the largest, accounting for more than half (51.52%), sandy land accounted for 11.96%, and sparse vegetation coverage (grassland, sandy land, Gobi) accounted for 68.68%, indicating that the project area is in the transition zone of woodland-sparse vegetation-sandy land, and the pressure on ecological environment protection and the situation of desertification prevention and control are still severe; 4) in the past 20 years, the increased area of vegetation such as arbor, shrub and grass accounted for about 20.64% of the project, which is mainly derived from desertified land. The reduced area of desertified land accounted for about 4.58% of the project area, indicating that the vegetation conditions in the study area are continuously improving, and the implemented ecological projects have a significant effect. The method in this paper can be more effectively used for multi-dimensional ecosystem service function evaluation. This study is expected to provide important scientific support for the research on the evolution law of the ecosystem and the evaluation of ecological engineering in the BTSSR.