| 摘要: |
| 土地盐渍化作为一种土壤灾害，严重制约着社会经济与农业的发展。对盐碱地进行实时监测，可为盐碱地的评价改良提供科学依据。由于盐碱地的信息复杂、提取精度不高，因此本文以高分六号（GF-6）卫星遥感影像为数据源，采用分形网络演化算法（fractal net evolution approach，FNEA）进行影像对象的多尺度分割，从面向对象的角度减少高分影像分类结果中的椒盐噪声问题，通过计算图像对象的局部方差和变化率来确定适宜的盐碱地分割尺度。利用基于特征选择的相关性算法（correlations-based feature selection，CFS）与Relief F算法分别对由光谱、纹理、形状、遥感指数构成的初始特征空间进行特征优选，精简特征子集，解决特征数量冗余问题，以此来优化随机森林对盐碱地提取精度。结果表明：CFS约简后的特征子集更小，精度更高，说明在盐碱地提取过程中，筛选特征数目能够减小冗余数据对提取精度的影响。CFS优化后的随机森林对盐碱地的提取效果较好，该方法总体分类精度达到83.7%。 |
| 关键词: GF-6 盐碱地 面向对象 特征选择 随机森林 |
| DOI：10.7515/JEE222035 |
| CSTR： |
| 分类号: |
| 基金项目:天津市自然科学基金项目（18JCYBJC90900）；国家自然科学基金项目（41971310） |
|
| Extraction of saline-alkali land based on multi-scale segmentation and feature optimization |
|
ZHU Li, GUO Qiaozhen, WU Zhengpeng, WU Huanhuan, HE Yunhai
|
|
1. School of Geology and Geomatics, Tianjin Cheng jian University, Tianjin 300384, China
2. Tianjin Institute of Surveying and Mapping Co., Ltd., Tianjin 300381, China
|
| Abstract: |
| Background, aim, and scope As a kind of soil disaster, soil salinization restricts the development of social economy and agriculture seriously. Real-time monitoring of saline-alkali land can provide scientific basis for evaluation and improvement of saline-alkali land. This paper mainly discussed the effect of object-oriented feature optimization algorithm applied to saline-alkali land. Taking Binhai New Area of Tianjin as the research area and the saline-alkali land as the research object, an optimization model was established to extract saline-alkali land. Materials and methods Taking GF-6 satellite remote sensing image as the data source, from the perspective of object-oriented, FNEA was used to segment the image object in multi-scales. Based on the statistics of local variances and change rates under different scales, the appropriate scale for saline-alkali land recognition was selected. On this basis, the initial feature space was constructed from four perspectives: spectral feature, texture feature, shape feature and remote sensing index feature. Two algorithms, CFS and Relief F algorithm, was used to optimize the initial feature space respectively. The obtained feature subset was used to optimize the extraction effect of random forest algorithm on salinized land, and the two optimization results were compared and discussed. Results (1) FNEA algorithm was used for multi-scale segmentation, and the appropriate segmentation scale of saline-alkali land in the study area was 123. (2) The initial feature space was optimized by CFS algorithm and Relief F, and the number of features was reduced to 40 and 17 respectively. (3) The overall classification accuracy of random forest extraction of saline alkali land was 76.3%, and that of random forest optimized by Relief F algorithm was 77.4%. The overall classification accuracy of CFS-optimized random forest in salt and alkali extraction was 83.7%. Discussion These results indicated that the CFS and Relief F, as two classical data filtering algorithm, for feature selection, can improve accuracy of random forest model in saline-alkali land, and can make the model improved to a certain extent. The overlap rate feature subsets optimized by the two algorithms was as high as 82%, indicating that the two kinds of algorithm of important characteristics were good search results. In addition, the features filtered out by CFS algorithm were almost twice as many as those filtered out by Relief F, but CFS optimized random forest had a better extraction effect on saline-alkali land, which indicated that the number of features was not positively correlated with the final classification accuracy. Conclusions Compared with Relief F algorithm, the random forest model optimized by CFS algorithm had a better recognition effect on saline-alkali land extraction, and the overall accuracy was 83.7%, which was 7.4% higher than that before optimization. CFS algorithm reduced the features to 17 and filtered 81.7% of the features, which solved the problem of data redundancy to a certain extent and improved the quality of data subset and the operating efficiency of stochastic forest algorithm. Recommendations and perspectives Firstly, The research ideas proposed in this paper can solve the problem of machine learning capability degradation caused by high-dimensional data redundancy, and can be applied to the optimization of other algorithm models and the recognition and extraction of ground classes. Secondly, in view of the different characteristics of saline-alkali land in different seasons, the influence of seasonal change on saline-alkali land should be considered in the future research. |
| Key words: GF-6 saline-alkali soil object-oriented feature selection random forest |
