融合时序Sentinel数据多特征优选的南方丘陵区油茶种植区提取

油茶作为江西省经济林树种之一，也是江西省特色优势产业，准确获取其空间分布在产量估算、生产管理和政策制定等方面具有重要意义。本研究针对南方多云多雨气候导致光学影像不足，以及丘陵山区地形破碎问题，以江西省宜春市袁州区为研究区，采用时序Sentinel系列影像数据和SRTM DEM数据为数据源，构建和优选了光谱特征、植被-水体指数、红边指数、雷达特征、地形特征和纹理特征共125个特征变量，其中，纹理特征采用累计差法(Δf)对比15种不同尺度窗口，计算Sentinel-1和Sentinel-2影像最佳纹理特征。基于ReliefF特征优选算法和随机森林分类算法，设计了8种特征组合方案开展实验，探讨不同特征类型对油茶提取精度的影响。结果表明：利用累计差法计算出的Sentinel-1和Sentinel-2的最佳纹理特征窗口尺寸均为35×35，最佳纹理特征组合为均值(Mean)、方差(Variance)和对比度(Contrast)；在光谱特征、植被-水体指数的基础上加入不同特征对油茶进行分类，不同类型特征对油茶提取的有利程度由大到小依次为S2纹理特征、S1纹理特征、地形特征、雷达特征、红边指数，相比于单一光谱和指数特征，纹理特征的加入可大幅度提高分类精度。多特征协同分类结果优于单特征分类结果，基于特征优选的油茶提取精度最高；基于ReliefF算法特征优选后的方案精度最高，总体精度为88.29%，Kappa系数为0.81。本研究利用时序Sentinel系列遥感影像和DEM地形数据，构建了针对多云雨南方丘陵山区的大范围油茶遥感提取方法，可为中国南方丘陵区域油茶资源调查与监测提供参考。

Extraction of Camellia oleifera Planting Areas in Southern Hilly Area by Combining Multi-features of Time-series Sentinel Data

As one of the economic forest species in Jiangxi Province, Camellia oleifera is also a characteristic advantageous industry in Jiangxi Province, and it is of great significance to accurately obtain its spatial distribution in terms of yield estimation, production management and policy formulation. In response to the lack of optical images due to the cloudy and rainy climate in the south, as well as the problem of fragmented terrain in hilly and mountainous areas, Yuanzhou District, Yichun City, Jiangxi Province, was taken as the study area. Using time-series Sentinel satellite imagery and SRTM DEM data as data sources, a total of 125 feature variables were constructed and selected, including spectral features, vegetation-water indices, red edge indices, radar features, terrain features and texture features. Among them, the texture features were calculated by comparing 15 different scale windows by using the cumulative difference method to calculate the best texture features for Sentinel-1 and Sentinel-2 images. Based on ReliefF feature preference algorithm and random forest classification algorithm, eight feature combination schemes were designed to carry out experiments to explore the impact of different feature types on the extraction accuracy of Camellia oleifera. The results showed that the optimal texture feature window for both Sentinel-1 and Sentinel-2 calculated experimentally by using the cumulative difference method was 35×35, and the optimal texture feature combinations were mean, variance and contrast. Building upon spectral features and vegetation-water indices, the incorporation of different features for Camellia oleifera classification demonstrated varying degrees of effectiveness. The favorability ranking of different feature types for Camellia oleifera extraction from large to small was as follows: S2 texture features, S1 texture features, terrain features, radar features and red edge index. Compared with single-spectrum and index features, the inclusion of texture features significantly enhanced classification accuracy. The synergistic classification results of multiple features surpass those of single-feature classification, with the highest precision achieved through Camellia oleifera extraction based on feature selection. The ReliefF algorithm feature optimized scheme had the highest accuracy with overall accuracy of 88.29% and Kappa coefficient of 0.81. This study utilized time-series Sentinel satellite imagery and DEM terrain data to develop a large-scale remote sensing extraction method for Camellia oleifera in the cloudy and rainy southern hilly mountainous region. This method can serve as a reference for the investigation and monitoring of Camellia oleifera resources in the hilly areas of southern China.