华南地区典型种植园地遥感分类研究

华南地区种植园地广泛分布,类型混杂多样,导致园地分布信息难以正确获取,为农业管理造成了较大困难。本研究基于Landsat8 OLI数据,通过数据融合、特征优化,应用随机森林算法构建面向对象的种植园地分类规则集,对华南地区典型经济作物香蕉、柑橘、葡萄、蒲葵、海枣、番木瓜和火龙果等进行类别识别,同时对比贝叶斯分类法、K最邻近分类法、支持向量机法、决策树分类法的分类效果。结果表明：数据融合会在一定程度上影响分类结果精度;植株形态、光谱特征接近,种植期交错是影响华南地区典型园地分类精度的重要原因;以中分辨率影像为数据源,面向对象的随机森林算法应用于种植园地分类研究总体精度可达88.05%,Kappa系数0.87,可以有效区分华南地区典型种植园地类别;相比于其他算法,随机森林算法在分类精度、可靠性和稳定性上具有一定优势,可为园地作物生长监测和种植管理提供科学依据。

Study on the Classification of Typical Plantations in South China

Plantation refers to the land of perennial woody and herbaceous plants that are planted to collect fruit and leaves, including the land used for nursery. Plantations in southern China is widely distributed, and the types are mixed and diversified. As a result, it is difficult to obtain the information of the garden distribution, which has caused great difficulties for agricultural management. In order to improve the classification accuracy of the remote sensing images of the plant species, using Landsat8 OLI data, we took the random forest algorithm to construct object-oriented plantation classification rule set based on data fusion and feature optimization. This was used to classify the banana, citrus, grape, livistona chinensis, phoenix dactylifera, carica papaya, hylocereus undatus and so on, which are typically planted in southern China. At the same time, we compared the classification effects of Bayesian classification, K nearest neighbor classification, support vector machine method, decision tree classification method and random forest classification, and it verified the applicability of the object-oriented random forest approach for the classification of the garden types in medium-resolution images. Then, based on the clustering matrix of the classification results of random forest classification and the feature distance matrix between two categories, we analyzed the reasons that affect the classification accuracy of the garden by combining the image conditions and field survey results. Finally, we compared the characteristic difference of typically easy-to-mix land before and after data fusion, and evaluated the effect of data fusion on the classification of garden. The results show that: after the data fusion, the difference of the characteristics of water body and aquatic vegetation is reduced, and the spectral difference of livistona chinensis and phoenix dactylifera is reduced in some bands. Although, the data fusion can improve the image spatial resolution, to a certain extent, it weaken the spectral differences of objects. Thus, it would affect the classification effect based on spectral information. The plant morphology and spectral characteristics of the plantation land are close to each other, the planting period of the plantation land is intertwined and there are important factors that affect the classification accuracy of typical garden in southern China. Based on the medium resolution images, the object-oriented random forest algorithm for plantation classification can reach 88.05% and Kappa coefficient of 0.87, which can effectively distinguish the typical plantation land types in South China. The classification results of object-oriented stochastic forest algorithm show that the area of various plantations in the study area: banana covers an area of 700.2 hectares, citrus covers an area of 981 hectares, livistona chinensis covers an area of 81 hectares, phoenix dactylifera covers an area of 68.04 hectares, carica papaya covers an area of 93.24 hectares, hylocereus undatus covers an area of 167.4 hectares, and grape covers an area of 16.2 hectares. For the classification accuracy of the coverage types, the random forest classification was higher than that of the other four algorithms except grape and hylocereus undatus. The random forest algorithm has certain advantages in classification accuracy, reliability and stability, and can provide scientific basis for crop growth monitoring and planting management.