Research On High Precision Extraction Of Corn Area In GF Remote Sensing Images With Ensemble Learning

Crop recognition is an important content and basis of agricultural monitoring and survey,such as planting area,growth,yield,quality,plant diseases and insect pests,etc.The rapid development of satellite remote sensing technology provides sufficient and high quality images which contain abundant information for agricultural industry application,but automatic extraction of remote sensing information plagued researchers for a long time.Elaborate extraction is still rely on artificial interpretation,this process is low efficiency,constraint by experience.Although machine learning methods continue to improve,but because the complicated differences,such as crop varieties,complicated structure,broken blocks,intercrop planting,phenomenon of same class with different spectrum,automatic extraction algorithms with simple rules and single structures are up against bigger limitations in cases of complex distribution.Meanwhile,various algorithms rely on handmade features and a large number of labeled samples,extraction accuracy of GF images is too low to meet requirements,and then final decision is affected badly.Therefore,improving the extraction accuracy becomes the key point in agricultural application of remote sensing.Ensemble learning can take advantage of multiple algorithms to solve the same problem,integrate the advantages of various algorithms and improve the overall generalization ability.This dissertation takes ensemble learning as leading clue to study high precision extraction algorithms of GF images,synthesize support vector machines(SVM),neural networks or deep learning networks.To improve the extraction accuracy,there are many aspects,such as different basic algorithms,different feature sets,different numbers of training samples to ensure algorithms' diversity.The proposed algorithms can also play a role in information extraction process of another crop.The research work contains four aspects as follow.(i)Study of feature extraction and feature sets constructionOn the basis of remote sensing images correction,fusion and features extraction,theimportance of features should be assessed based on random forest.Then extraction results of different feature groups are taken as evidences to deal with conflict evidence according to modified D-S evidence synthetic rules.The synthetic results of two images experiments are more than 0.84 and 0.87,increased by 4% and 6% respectively.According to above results,some important features are selected and recombined to build different types of feature sets,such as spectral set,texture set and joint set.(ii)Study of homogeneous and heterogeneous ensemble algorithmsClassical supervised shallow algorithms,such as support vector machine and extreme learning machine are studied deeply.The new algorithm named multi-kernel support vector machine adaptive boosting is put forward based on hybrid iteration strategy.And then,based on analyzing differences of classification methods performance,heterogeneous hybrid algorithm is put forward based on SVM and ELM to improve algorithm generalization ability in the situation of complex terrain.(iii)Study of deep networks integrated algorithmsThis section explores the basic principle and structure of stacked auto-coder and deep convolution neural networks,and their setting rules and parameters optimization means respectively.Then,deep networks ensemble methods are constructed based on unsupervised feature learning to learn deep features and decide land-cover objects attributes according to one-dimension pixels and two-dimension images.Meanwhile,with considering special requirement of convolution neural networks and local pixels context,an effective method has been constructed to prepare two-dimension input images to overcome the disadvantages of classical partitioning.(iv)Experiments verificationThis section selects representative regions of GF-1 and GF-2 images with different spatial resolution(2 meters and 0.8 meters),time phase(single phase and two phases),size(512?512?1024?1024 and 1500?1500 etc),feature sets(spectrum,texture and joint feature set)and land-cover distribution,to verify performance of six proposed algorithms.The homogeneous and heterogeneous ensemble algorithms based on SVM and ELM improveoverall accuracy compared with classical supervised algorithms and simple methods,especially,the overall accuracy with joint feature set achieve 0.85,the precision of heterogeneous ensemble algorithm is better than homogeneous algorithm.The overall accuracy of deep convolution neural networks is more than 0.90 basically.From what has been discussed above,ensemble learning algorithms can improve extraction accuracy effectively,especially,the deep networks ensemble algorithms are superior to homogeneous and heterogeneous algorithms,and higher spatial resolution can improve corn extraction precision obviously.