一种深度学习土地利用图斑影像核查方法

A Method of Deep Learning to Verify Land Use Classification of High-resolution Image Patches

鉴于深度卷积神经网络(deep convolutional neural networks,DCNN)的矩形感受野与土地利用图斑不规则形状范围的套合程度是影响土地利用图斑识别精度的重要因素,文章提出一套基于高分辨率影像的DCNN土地利用类型核查方法。该方法采用图斑掩膜裁切高分辨率影像,滤除矩形感受野内不套合部分,降低背景噪声,提高信噪比,从而准确识别图斑影像语义,通过检查语义与土地利用类型的符合性实现图斑的土地利用类型核查。在广水市第三次国土调查土地利用类型核查工作中,该方法获得了召回率为93.54%、准确率为93.57%的结果,为国土调查核查工作自动化提供了技术支撑。

The degree of fit between the regular rectangular receptive field of deep convolutional neural networks(DCNN)and the irregular shape range of land use patch is an important factor that affects the accuracy of land use patch recognition.This paper proposes a DCNN land use classification verification method based on high-resolution images.The high-resolution image is cropped by using patch mask to filter out the parts that do not fit in the rectangular receptive field,which can reduce background noise and increase the signal-to-noise ratio,so as to accurately identify the semantics of the patch image.At last,it verifies the land use classification of the patches by checking the consistency between the semantics and the land use classification.In the third land use survey of Guangshui city,the proposed method obtains a recall rate of 93.54%and an accuracy rate of 93.57%,which provides technical support for the automation of national land survey and verification.