基于多分类器集成的GF-1影像围填海地物识别

Identification of coastal reclamation from GF-1 imagery using ensemble classification strategy

围填海是人类获取海洋资源的重要方式。监测围填海的变化是海岸带管理、海岸带演变研究中一项非常重要的任务。然而,围填海地物复杂多变,给利用遥感技术监测围填海带来困难。为此,通过构造识别地物类别的10个特征因子(GF-1的Band1—4波段的均值特征、波段均值的均值、对象面积、对象周长、外接矩形面积、对象面积与外接矩形面积之比和对象周长与对象面积之比),提出一种识别GF-1影像中围填海地物的多分类器集成算法;对特征因子进行集成,构建出单个特征分类器模型、光谱特征分类器模型、形态特征分类器模型和所有特征集成分类器模型4种组合特征分类器模型;对每种分类器模型进行试验研究,并对比分析4种集成模型的多分类器围填海地物识别精度。结果表明,单个特征分类器模型识别精度最高达到82.03%,光谱特征分类器模型识别精度为63.28%,形态特征分类器模型识别精度为87.50%,所有特征集成分类器模型识别精度为80.47%。本研究结果可为监测围填海变化提供较好的解决方案。

The coastal reclamation is an important way for people to access marine resources. Monitoring the coastal reclamation changes is an important task in coastal zone management and coastal zone evolution study. However,the coastal reclamation feature is complex, and it is difficult for remote sensing techniques to efficiently monitor reclamation. In this paper, the authors propose an ensemble classification algorithm for identifying four categories of reclamation using GF - 1 imagery. The ensemble classification is constructed based on minimum distance algorithm and 10 features from manually extracted image objects. The 10 features include four mean features of each object in the four bands of GF - 1 imagery respectively, mean value of the four mean features, object size, object perimeter, external rectangular area, ratio of object area, external rectangular area, ratio of object perimeter and object area. The proposed method was extensively tested by using two GF - 1 images from 2013 and 2014. The results show that the highest accuracy of single feature model is up to 82.03 %, and the accuracy of spectral features based ensemble model and that of the spatial features based ensemble model are 63.28% and 87.50% respectively, and the accuracy of full feature based ensemble model is 80.47%. This study provides a useful solution for monitoring the coastal reclamation.