A Classification Method Of Farmland Landscape Small-scale Features Based On UAV Image

Agricultural landscape is composed of cultivated and non cultivated habitat.Non cultivated habitat is a general term of natural and semi-natural habitat in farmland landscape,which is embedded in farmland ecosystem,generally including trees,hedges and margin and grass.Their quantity,quality and spatial allocation play an important role in ecosystem services,biodiversity and sustainable development of farmland.In recent years,with the application of high-resolution remote sensing satellites in China becoming more and more mature,UAV and other aerial images provide the possibility for small-scale object recognition by their high-resolution characteristics.However,at present,the precision of making map of farmland landscape is not high,and the standard of discrimination of small objects in the landscape is low.At present,the research of interpretation in the landscape is generally based on large-scale,and the research of interpretation of small area or single non-farming habitat is less,which is difficult to meet the requirements of farmland micromanagement and landscape structure research.On the basis of building multi-level category structure,this paper takes UAV image as data source,combined with object-oriented and high-efficiency autonomous random forest algorithm(random forest),proposes a universal interpretation scheme for small and medium-sized land features in farmland landscape.Using the characteristics of different habitats in the spatial structure scientifically and reasonably,this paper explores the ability of machine learning algorithm to use feature variables to describe classes and complete automatic and accurate classification.The purpose of building a multi-level category structure is only to reduce the number of variables,so it ensures that the child object level is not affected by the misclassification in father object level,and this has better generalization ability,as well as the interpretation results can have higher accuracy.This paper mainly constructs the method of making small-scale wide map of farmland.In order to verify the effectiveness and generality of the method,three different UAV images are selected as the verification object.The main work and conclusions are as follows:1)For the classification of small-scale features in farmland landscapes,this study constructs an automatic interpretation method based on the combination of object-oriented and random forest.That is to say,based on multi-level category structure,object-oriented,feature space optimization,typical sample set optimization,random forest model design and debugging are constructed.Finally,the model is trained by the optimized feature variables and samples and then applied to subsequent prediction and classification,so as to obtain high-precision farmland landscape map.2)A general multi-level category structure is constructed.Among them,the father object represents wide land use categories,while the child object represents the subdivision categories of the father object on a small scale.Because the features in the child object are selected from a large number of feature sets bythe father object layer,and the child object level is not affected by the misclassification in father object level.This classification structure can produce better classification effect.3)Establish an optimization scheme to improve the typical samples.In order to ensure the selection of more comprehensive typical samples and avoid subjective problems caused by visual interpretation,this study proposes a combination of field sampling,visual interpretation and threshold method to expand and improve the typicality of samples reasonably.4)The modeling and optimization scheme of random forest model is established.This paper proposes a scheme to design a random forest model and optimize its parameters based on R language platform,and analyzes the influence of the number of features when constructing a single tree in the random forest model and the number of decision trees in the forest on the classification efficiency and accuracy.Using the characteristics of randomness,the optimal parameters of the model are selected by the way of calculating the average error of OOB,so that the model can achieve better learning performance and prediction accuracy.5)A scheme is established to solve the problem that a large number of redundant feature variables have bad effects on classification efficiency and accuracy.In this paper,the feature variables are selected based on the descending order of average precision.The influencal feature variables are selected to construct the optimal feature subset,which can optimize the feature space and improve the efficiency and accuracy of subsequent classification.6)In order to verify the validity and feasibility of the method,three research areas are used for demonstration.The results show that the overall accuracy of the model validation is more than 90%.Through the case study of Shifosi Township in Shenbei New District and Laoha river drainage area A and B in Jianping County,the overall accuracy of the model classification in Shifosi township is 91.46%,and the kappa coefficient is 86.93%;the overall accuracy of Laoha river drainage area a in Jianping County is90.88%,and the kappa coefficient is 87.52%;the overall accuracy of random forest in area B is 93.13%,and the kappa coefficient is 89.76%,all of which meet the requirements of understanding the translation accuracy.It shows that this method has reliability and universality.The classification effect of farmland landscape is good,which can show small-scale features such as margin with a width of 0.5m,grass and trees.The technical method proposed in this paper can provide high-precision map support for the future research of farmland ecology.