基于DBSCAN和BP_Adaboost的农机作业地块划分方法

农业机械(农机)在多个地块作业,费用和效率有时需按地块统计,现有的农机监控系统仅能记录农机定位信息和作业状态信息,难以实现地块的自动精准划分。本文通过研究轨迹点属性特征,分析作业地块数量不确定性和轨迹点分布规律,采用基于密度聚类方法(Density-based spatial clustering of applications with noise, DBSCAN)和分类器集成算法(BP＿Adaboost)结合的方法划分地块。根据DBSCAN算法对农机轨迹点多数有效、识别错误集中的特点,结合BP＿Adaboost算法挖掘多维度信息关联、容错能力强、分类效果好等优势,先利用DBSCAN得到初步的轨迹点状态类别,再利用BP＿Adaboost算法建立训练模型对农机轨迹点状态精准识别,根据时间序列和类别标记划分地块。本文方法既解决了只依靠阈值和经纬度信息聚类不准确的问题,也减少了大量样本标记工作。利用该方法轨迹点状态识别准确率达96.75%,地块划分准确率为97.74%。

Land Division Method for Agricultural Machinery Operation Based on DBSCAN and BP_Adaboost

Agricultural machinery operates in multiple plots, and the cost and efficiency sometimes need to be counted according to the plots. The existing agricultural machinery monitoring system can only record the positioning information and operation status information of agricultural machinery, which is difficult to realize the automatic and accurate division of plots. By studying the attribute characteristics of track points, the uncertainty of the number of work plots and the distribution law of track points were analyzed, and the combination of density clustering method (density-based spatial clustering of applications with noise, DBSCAN) and weak classifier integration algorithm (BP_Adaboost) were used to divide the plots. According to the characteristics that DBSCAN method is effective for most agricultural machinery trajectory points and the recognition error is concentrated, combined with BP_Adaboost method to mine multi-dimensional information association, strong fault tolerance, good classification effect and other advantages. Firstly, DBSCAN was used to obtain the preliminary track point state category, and then the method of BP_Adaboost was used to establish a training model to accurately identify the track point state of agricultural machinery, and divide the land mass according to time series and category markers. The method not only solved the problem of inaccurate clustering only relying on threshold and longitude and latitude information, but also reduced a lot of sample labeling work. Using this method, the accuracy of track point state recognition was 96.75%, and the accuracy of plot division was 97.74%.