基于面向对象多特征学习的无人机影像农作物精细分类方法

Crop Classification Method from UAV Images based on Object-Oriented Multi-feature Learning

获取农作物高分辨率影像特征,探究多特征学习训练对实际作物分类效果的影响,对农业部门掌握作物种植精细化结构信息,高效实施生产管理具有重要意义。针对无人机获取的高分辨率可见光遥感影像,提出一种基于面向对象多特征学习的农作物精细分类方法。首先借助HSI(Hue,Saturation,Intensity)模型对RGB图像进行色彩空间变换,以进一步挖掘影像中潜在色彩结构信息,然后利用ESP(Estimation of Scale Parameter)算法和CART(Classification And Regression Tree)决策树分别确定影像最佳分割尺度及构建最优特征学习空间,最后采用面向对象随机森林(Random Forest)分类算法对多特征空间进行学习训练,以实现作物精细分类,并结合验证数据集进行精度评价。结果表明:该方法对研究区作物的分类总体精度达到90.18%,Kappa系数达到0.877,均大于像素级和单特征学习的分类精度,能够有效区分出不同的作物类型;所构建的最优特征学习空间对研究区内棉花、玉米、西葫芦、葡萄的分类效果较好,各作物类型的生产者精度均大于89%。研究结果可为农业精准管理和作物种植结构优化提供参考。

Obtaining high-resolution image features of crops and exploring the influence of multi-feature learning on actual crop classification are of great significance for agricultural departments to grasp the information of crop planting fine structure and efficiently implement production management.For the high-resolution RGB image acquired by UAV,this paper proposes a new classification method of crops based on object-oriented and multi-feature learning.Firstly,the HSI model is used to transform the colour space of RGB images to mine the potential information of images further.Secondly,the ESP algorithm and CART are used to determine the optimal image segmentation scale and construct the optimal feature learning data set of classification.Finally,object-oriented Random Forest classification algorithm was used to learn and train the multi-feature space,so as to achieve fine crop classification,and accuracy evaluation was carried out in combination with validation data set.The experimental results show that the overall accuracy of the classification in the study area reached 90.18%,and the Kappa coefficient reached 0.877,both of which were greater than the accuracy based on pixellevel and single-feature learning.The optimal feature learning space constructed in this paper have a good classification effect on cotton,corn,cocozelle,grape and other major crops in the study area.The producer's accuracy of each crop type is greater than 89%.This research can provide a reference for agricultural precision management and planting structure optimization.