基于集成深度学习模型的耕地地块提取

CROPLAND FIELD EXTRACTION BASED ON ENSEMBLE DEEP LEARNING MODEL

关键词:  耕地地块  地块提取  卷积神经网络  Bagging抽样  集成学习

Key words:  cropland field  farmland extraction  convolution neural network  Bagging  Ensemble Deep Learning

摘要:

目的 耕地地块空间分布是农业生产管理和农业政策制定的重要基础信息。我国农业耕作方式和种植结构复杂，地块形状多样、均质性较低，基于Landsat 影像的传统方法难以实现地块的准确提取。方法 文章提出一种集成深度学习模型（Ensemble Deep Learning，EDL），可以在高分辨遥感影像中实现地块提取。首先通过随机可放回的Bagging抽样方法得到不同的训练集，然后将训练集用于多个卷积神经网络（FCN、PspNet、SegNet、Unet），逐像素计算相应的耕地边界概率，最后将概率图按照平均值进行集成，获得耕地地块边界，进而实现耕地地块的提取。结果 该文提出EDL方法提取耕地地块的总体精度达到96%，相较于FCN、SegNet、Unet提升了1%， 相较于PspNet提升了2%。相较于单个分类器， 集成深度学习模型可以减小偏差，提高地块提取的准确率结论 集成深度学习模型能够综合多个卷积神经网络的优点，提高分类精度，为耕地地块边界提取提供了新方法。

Abstract:

The spatial distribution of cropland field is an important basic information for agricultural production management and agricultural policy making. The traditional methods based on landsat images are difficult to achieve accurate extraction of field parcels due to the high fragmentation of cropland and complex planting structures in China. In this study, an ensemble deep learning (EDL) model was proposed. Firstly, the various training sets were obtained by random Bagging sampling method, and then the training sets were applied to multiple convolution neural networks (FCN、PspNet、SegNet、Unet) to obtain the cultivated land boundary probability map. Finally, the probability maps were integrated according to the average value to obtain the field parcels. The results showed the overall accuracy of EDL reached 96%, which was 1% higher than FCN, SegNet and Unet, and 2% higher than PspNet. Compared with a single classifier, EDL could reduce the deviation and improve the accuracy of cropland field extraction. It concludes that the EDL model integrates the advantages of multiple convolutional neural networks with better performance, and it can provide an effective method for field parcels boundary extraction in China.