基于机器学习的高分辨率遥感影像农业大棚提取

Extraction of agricultural greenhouse based on high-resolution remote sensing images and machine learning

农业大棚数量、面积和空间位置信息的及时和准确采集,对于农业结构调整、污染防治有着重要意义。遥感数据被广泛应用于大棚提取,但是在遥感影像上大棚所占的面积往往远少于背景地物,属于少数类,而传统随机森林(RF)机器学习分类器对非平衡数据集中少数类样本分类效果较差。为了使RF更适用于对大棚的提取,优化了RF的样本选择方式,首先随机抽取相同数量的不同类别样本构建初始训练样本集进行RF建模,然后基于投票熵和样本间的广义欧几里得距离循环迭代的测试集中优选样本,并将所选样本添加到训练集。结果表明,优化的RF模型用于基于遥感数据的农业大棚信息获取可以取得满意的结果。

Timely and accurately acquisition of the area and spatial distribution of greenhouse in the agricultural regions would be valuable for the local authorities taking measures to adjust regional agricultural structure and to prevent and control environmental pollution.Remote sensing data are widely used in greenhouse extraction,however,the areas of greenhouses in remote sensing images are often far less than the background objects,and they belongs to the minority class.However,the traditional random forest(RF)has poor classification performance for the minority samples in the unbalanced data set.In order to make the RF more suitable for the extraction of the greenhouses,we improved the sample selection method of RF.First,the same number of samples from the minority class and the majority class were selected randomly to build training set for RF modeling.Then,samples with high quality were added to the training set automatically according to the voting entropy and the generalized Euclidean distance based on the sample characteristic parameters.The experimental results show that the optimized RF model we proposed could achieve excellent classification results for the greenhouses.