基于地理国情的深度学习分类容错性研究

Study on Deep Learning Classification Fault Tolerance Based on Geographical Conditions

地理国情监测获取的地表覆盖分类成果具有覆盖区域全、精细度高、时相新等优势,具有作为深度学习分类模型训练样本的能力和优势,能够大大减少样本获取的成本。但是,受数据源、时相以及采集标准等因素的影响,直接使用地表覆盖数据作为样本,往往与模型训练采用的影像存在一定的误差。研究采用深度学习语义分割算法,比较了人工标注样本以及不同量地表覆盖数据样本的分类结果。结果表明深度学习分类算法具有一定的容错能力,即使地表覆盖样本与训练影像存在一定的误差,当样本量足够大时,同样能够获得有效的分类结果。研究成果说明了地表覆盖数据作为深度学习分类样本的可行性,为如何更好地使用该数据提供了思路,一定程度上解决了深度学习中样本获取难的问题。

Enough accurate label data acquisition is one of the difficulties in deep learning remote sensing image intelligent interpretation.With the advantages of large coverage,high accuracy and yearlyupdate,the land cover from geographical conditions has become potential label data.However,as the data source,acquisition time and data collection standard may be different,the land cover data often don’t match training imag-es very well.We used different kinds of labels(human labels and different numbers of land cover data)to train the deep learning semantic seg-mentation model and compared the results.The result shows that the deep learning method has the fault tolerance ability when label data don’t match the images very well.The study demonstrates the feasibility of using land cover data as labels to train deep learning classification models.At certain extent,it solves the training samples acquisition for deep learning method.