摘要
土地利用分类结果对国土空间的管理至关重要。为提高土地利用分类结果的准确性,本文以博湖县为研究区,使用Sentinel-2A影像提取光谱特征,并结合雷达、光谱指数、土壤和地形特征构建6个面向对象的土地利用分类模型,使用简单非迭代聚类(SNIC)算法和随机森林(RF)算法对影像进行分割和分类,得出模型的分类精度以及特征重要性排序,最后使用分类回归树(CART)算法验证辅助数据集对提高分类精度的影响。结果表明:使用SNIC算法分割影像时,分别设置种子大小为17、紧凑度为0时,该研究区影像分割效果最好。基于RF分类算法,在只使用光谱信息进行分类时分类精度最低,加入雷达、光谱指数、土壤和地形特征中任何一个辅助数据集均可提高土地利用的分类精度,其中地形特征对提高分类精度的效果更显著,加入所有辅助数据集时分类精度达到最高,OA=92.34%,Kappa系数=0.91。使用CART算法进行分类有效性验证得出,基于RF算法的分类效果优于CART算法。基于遥感云平台的SNIC分割算法,融入辅助数据集进行面向对象分类,为提高土地利用分类精度提供参考。
Land use classification is critical to the management of land space.To improve the accuracy of land use classification,this study takes Bohu County as the research area,uses Sentinel-2A images to extract spectral features,and combines radar,spectral index,soil,and terrain features to construct six object-oriented land use classification models.We then use a simple non-iterative clustering algorithm and random forest algorithm to segment and classify the images and obtain the classification accuracy and feature importance ranking of the model.In the final step,we use the classification regression tree algorithm to verify the influence of the auxiliary dataset on the improvement of the classification accuracy.The results show that when using the SNIC algorithm to segment the images,with seed size 17 and compactness 0,the image segmentation effect in this study area is the best.The classification accuracy is the lowest when only spectral information is used,and adding any auxiliary dataset of radar,spectral index,soil,and terrain features can improve the classification accuracy of land use.Among those auxiliary datasets,the effect of terrain features on improving classification accuracy is more significant,and the classification accuracy reaches the highest when all auxiliary datasets are added,with OA=92.34%and Kappa coefficient=0.91.The classification validity is verified using the categorical regression tree algorithm,it shows that the classification effect based on the random forest algorithm is better than that of the categorical regression tree algorithm.The SNIC segmentation algorithm based on the remote sensing cloud platform is integrated into an auxiliary data set for object-oriented classification,which provide a reference for improving the accuracy of land use classification.
作者
李坤玉
王雪梅
李锐
李顿
LI Kunyu;WANG Xuemei;LI Rui;LI Dun(College of Geographic Science and Tourism,Xinjiang Normal University/Key Laboratory of Lake Environment and Resources in Arid Region of Xinjiang,Urumqi 830054,China)
出处
《中山大学学报（自然科学版）（中英文）》
CAS
CSCD
北大核心
2024年第1期34-44,共11页
Acta Scientiarum Naturalium Universitatis Sunyatseni
基金
新疆维吾尔自治区自然科学基金(2020D01A79)
国家自然科学基金(41561051)。
