Research On Artificial Object Extraction Using High Resolution Remote Sensing Imagery

The development of high-spatial resolution remote sensing(HRRS)imagery makes it possible to extract objects at parcel level,providing a data foundation for refined geographic applications.However,the data complexity also brings huge challenges to traditional object classification algorithms.Among all the types of object,artificial objects represented by buildings and farmland in HRRS imagery have more distinct artificial transformation features,are more complicated,richer in parcel numbers,more difficult to extract,and more closely connected to human life and production.Thus,the realization of the extraction of artificial objects at parcel level and obtaining their information has greater socioeconomic significance.This study addresses the problem of artificial object extraction using HRRS imagery.In order to adapt the classification system to the differences in object visual characteristics in HRRS imagery,artificial objects are divided into two types: construction land and agricultural land.There are three other types of natural objects in water,grassland and rocky land.The visual characteristics of artificial objects in HRRS imagery is analyzed in terms of spatial morphology,internal texture and semantic scene,along with the comparison of the characteristics and applicability of traditional methods and deep learning methods.This study selects the artificial objects that are most important,most representative,and most valuable in research,buildings and farmland,and points out the current problems in the object extraction using deep learning technology: the research and application of building extraction are relatively mature,but the dependence on the pre-trained models limits the improvement possibility on the network structure;the use of mainstream semantic segmentation methods in the extraction of cultivated land is difficult to achieve extraction in farmland parcel scale,resulting in an under segmentation problem.Taking the extraction of buildings and farmland as an example,this study relies on convolutional neural networks in deep learning technology,using the visual features and inherent principles in HRRS imagery to optimize network structures.A building extraction method based on a semantic model and a farmland extraction method based on an edge model and a texture model are proposed respectively.In terms of building extraction,this study uses deep learning semantic models for building extraction from HRRS imagery.DE-Net is proposed to improve the preservation of spatial information during the network calculation process,especially in downsampling,encoding and upsampling.Its network structure innovations are:(1)the downsampling operation combines strided convolution and max pooling to enrich sampled features;(2)the encoding module uses SELU activation functions instead of the commonly used Re LU activation functions;(3)densely upsampling convolution modules are used in upsampling,allowing DE-Net to encode spatial information into downsampled feature maps to achieve fine pixel-level classification;(4)use dice and cross-entropy loss function to solve the imbalance class problem.In the experiment,DE-Net outperformed the best results comparing with other mainstream networks on the public WHU building dataset and a self-built Suzhou building dataset with Gaofen-2 satellite images,which proves designing network structures according to the visual characteristics of building without relying on pre-training models can also bring about an increase in accuracy.In terms of farmland extraction,this study proposes farmland extraction based on edge models and texture models.The edge model removes the last two pooling layers on the basis of the RCF network,which reduces the loss of spatial information and improves the recall rate of parcel contour,makes them less prone to breakage to preserve recognizable parcels as much as possible.This research proposes the concept of cultivated land purity for sample selection and designs an Internal Texture Network(IT-Net)as the texture model.IT-Net uses feature widening modules to expand feature map dimensions,and then using multi-pass convolution modules that combine multiple convolution kernels of different kernel sizes into a three-way convolution scheme to improve the sensitivity to texture features of different sizes.Finally,three fully connected layers are used for classification.The texture model internally samples multiple texture image tiles from parcels extracted by the edge model and averages the prediction results to obtain the farmland probability of the parcels.This study fully compared the accuracy difference between different algorithms of edge model and texture model and the influence of hyper-parameters in experiments.This study extracted 944,800 parcels of Longxi County with 130-times faster than manual extraction in estimation where the extracted farmland parcels have a high degree of contour accuracy and category accuracy.This study also discusses the significance of this method in practical farmland parcel production.On the one hand,the probability heat map produced by the texture model at parcel scale enables parcel classification accuracy to change in real time by adjusting thresholds,making it possible for more intelligent human-computer interactive parcel extraction.On the other hand,the edge model and texture model have a low demand for training samples,so that this method of farmland extraction can be quickly and flexibly applied to new data and object types in practical applications and therefore has broad application prospects.