Research On Automatic Extraction Method Of Tea Planting Area From High-resolution Remote Sensing Images

China is a large tea producing country.The monitoring of tea planting areas have a great significance to China's economic development.However,the traditional artificial field survey method needs a lot of manpower and material resources,poor timeliness,low precision,can not obtain the spatial distribution information of tea planting area in time and effectively,but also has a relatively high error.Remote sensing technology has the natural advantage of accurate and timely access to information,so it is feasible to use remote sensing monitoring method to realize the automatic extraction of tea planting area.But at the same time,due to the similarity of tea plant spectral characteristics with other crop planting areas,the current tea extraction algorithm is difficult to achieve good results.In this paper,based on remote sensing technology,combined with a variety of algorithms to achieve the automatic extraction of tea planting area,the main research contributions are as follows:(1)An object-oriented and variogram based automatic extraction method for tea planting area is proposed.Firstly,the object-oriented method is used to realize the high-precision extraction of non-vegetation area,which eliminates the interference of non-vegetation area,and lays the foundation for the next step of tea planting area extraction.Then,based on the difference of variogram texture features between tea planting area and other vegetation areas,a decision tree classification model is constructed to extract tea planting area with high precision.The experimental results show that this method is feasible and advantageous for tea planting area extraction from high-resolution remote sensing images.(2)Aiming at the difference of target size in high-resolution remote sensing image,that is multi-scale problem,and the false alarm problem caused by within class inconsistency of high-resolution remote sensing image,an automatic extraction method of tea planting area in GF-2 image based on MF-DFNet(multi scale feature and discriminative feature network)was proposed.To solve the multi-scale problem,firstly,the improved VGG16 network is used as the main network of the encoder,which removes the full connection layer and uses Hierarchical-Split residual block instead of the convolution layer.The Hierarchical-Split residual block effectively improves the multi-scale feature extraction ability of the network.Then,the residual RFB module is added at the top of the backbone to expand the receptive field and realize the dense multi-scale context information aggregation;To solve the problem of false alarm,the foreground-scene relation module is introduced between the encoder and decoder to enhance the discriminative power of tea area features.Then the channel attention module(CAB)is introduced into the decoder to further select more discriminative features.Experimental results show that the method can effectively improve the tea extraction accuracy of high-resolution remote sensing images.The F1-score and Io U can reach 0.969 and 0.940 respectively.(3)In order to improve the extraction speed of tea in high resolution remote sensing image effectively,the extraction accuracy can meet the basic practical needs,a fast extraction method of tea planting area from GF-2 image based on SPRRD-Shuffle Net V2 is proposed.Firstly,Shuffle Net V2 network is used as the backbone of encoder which the last 1×1 convolution layer,global pooling layer and full connection layer are removed,and the decoder is added to achieve pixel-level classification.And then,without increasing the number of parameters and affecting the inference speed,improved strip pooling module(I-SPM)and mix pooling module(MPM)are added to the encoder to capture global and local dependencies.A residual refinement block(RRB)is added to the decoder to optimize the output features.The experimental results show that this method can meet the basic extraction accuracy requirements,F1-score and Io U can reach 0.940 and 0.887 respectively,and inference speed can reach 0.006s/image,which effectively improves the extraction speed.(4)A high resolution remote sensing image extraction system is designed and developed by using pyqt5 library of Python.The system includes the functions of training data preparation,model training and tea area detection.The detection results can be obtained from the high-resolution remote sensing images input by users,which can replace the heavy field survey work,save a lot of time and energy,and can be used in the actual tea garden detection work.