| Forests play a role in regulating climate and water conservation in terrestrial ecosystems and are an important material basis for national sustainable development strategies.Accurately obtaining the spatial distribution of forest land contributes to the monitoring of forest land change and through the analysis of its changes,the corresponding protection management policies are formulated.The development of remote sensing technology can provide a cost-effective means for ecological monitoring and large-scale forest resource inventory.When using remote sensing technology to obtain forest land information,the traditional visual interpretation is labor-intensive information extraction work,and the resource consumption is huge.Remote sensing technology combined with computer greatly improves the automation of information extraction.Among them,machine learning is one of the important research directions.The current machine learning method is generally applied to the scene-by-view training and scene-by-view classification in the application of specific remote sensing image classification.There are problems such as repeated selection of training samples and consumption of model parameters in the classification process.For a wide range of research areas spanning dozens or even hundreds of remote sensing images,how to achieve high-precision forestland information extraction in a large-scale research area through a machine learning model established in a small area typical area remains to be explored.The ability of a classifier to correctly classify data outside the training set in a machine learning method is called generalization capability,also known as predictive power.Generalization ability is an important indicator to evaluate the quality of the algorithm generated by the classifier.The stronger the generalization ability,the higher the prediction accuracy.Selecting a classifier with high generalization ability to use it effectively for other data sets outside the training set is critical to achieving portability of the machine learning model.In order to extract high-resolution remote sensing imagery forest information from large-area and multi-temporal batch remote sensing images,it is necessary to explore a machine learning algorithm with higher generalization ability.The research is based on two points to explore the generalization ability of machine learning for large-area forest land information extraction.The first is the use of good machine learning classifiers.This paper selects three mainstream machine learning algorithms from the research of many machine learning methods,including support vector machine,random forest and deep neural network.Integrate object-oriented image analysis method to construct the optimal model of machine learning classifier.Secondly,in view of the current scene-by-view training and scene-by-view classification of remote sensing images,consider selecting training samples from a small area in a typical area from the study area,and training the selected machine learning classifiers separately,using the optimal classifier for training.Realize the extraction of forest land information throughout the study area.The generalization ability of the selected machine learning classifier is comprehensively evaluated by the accuracy of the forest land extraction and the consistency of the joints in the whole study area.(1)For most remote sensing images,the information extraction mode of scene-by-view training and scene-by-view classification is adopted.The optimal machine learning classifier established on the representative data of small area is demonstrated by experiments.It can be applied to the whole research area.Data can even be applied to similar data and similar data.Prove the rationality of the remote sensing information extraction framework proposed in the paper.At the same time,the information extraction framework reduces the parameter consumption and sample repetitive selection in the process of classifier establishment.High-precision information extraction for batch remote sensing images can be achieved by small sample selection,which greatly reduces sample dependence and improves the efficiency of remote sensing information extraction.(2)Based on the integrated object-oriented and machine learning method,the optimal model of support vector machine,random forest and deep neural network is constructed based on the typical sample area of Qilian Mountain coniferous forest.The information extraction framework proposed in the paper is used to compare the generalization ability of three machine learning classifiers in time and space.Overall,deep neural networks show the highest generalization ability both in space and time,followed by random forests and support vector machines.At present,most classification regression and other learning methods are shallow structure algorithms.Compared with shallow learning such as support vector machine and integrated learning random forest,deep neural network uses "hierarchical" information to fit complex nonlinear functions.In the paper,the object-oriented image analysis method is integrated in the construction of deep neural network.For the multiple features of the extracted coniferous forest information,the Relief F-Cfs-Pso combined filtering feature optimization method is adopted to achieve dimensionality reduction.Reduce the number of input variables and use more neurons per layer to fit more complex functions.By eliminating redundant information between multiple hidden layers,it is finally determined that the 4 hidden layer networks can achieve higher generalization capabilities.(3)Analysis of the differences in the generalization ability of time and space of the three classifiers shows that the spatial generalization ability of the three classifiers is higher than the generalization ability in time.The reason is that the optimal model is constructed when exploring the generalization ability of space,and the model is directly used in the information extraction of another period of image with different time segments,resulting in a decrease in accuracy.Therefore,in order to optimize the optimal depth neural network,by changing the sample,the typical sample sample of the other image is added,and the deep neural network is used to extract the information of the whole research area,and the extraction result is in the classification accuracy and multi-view image.The accuracy of the joint consistency and the like are improved.It is proved that the new sample area can be added to improve the accuracy of neural network information extraction.(4)Based on the optimal model,the two-stage precision of the Qilian Mountain coniferous forest with the highest accuracy was analyzed.The temporal and spatial changes of the coniferous forest area and the landscape pattern index of the past 20 years were compared.Compared with 2000,the coniferous forest was overall in 2017.Increased by 423.88km2,an increase of 11.14%.From the perspective of the landscape pattern,the overall degree of fragmentation increases and the boundary of the landscape shape becomes complicated.The overall interference is reduced by human interference.The overall distribution of coniferous forests showed a good situation.However,the development of forest land in some areas is uneven.Specifically,the changes in the area and landscape indicators of the administrative divisions of Shandong,Central,West,and counties in Qilian indicate that the growth of forest land between the regions is uneven.The phenomenon of deforestation and land reclamation in the protected areas has always existed,and the contradiction between forests and forests has reduced the forest land of the Sunan Yugu Autonomous County in the western section of the Qilian Mountains by 212.15km2 in the past 20 years.Although the overall forest area of Qilian Shandong section shows an increasing trend,the forest land in the Tianzhu area in the eastern section is seriously degraded,the landscape pattern is broken,and the disturbance of human activities makes the landscape form simple.The area of the policy forest land in the Yuanyuan Hui Autonomous County and the Yuzhong County in Qinghai Province has been greatly increased due to the return of farmland to forests(grass).The plaques in the landscape are more compact and the overall shape is simpler.Therefore,although the coniferous forest has been greatly improved as a whole,the forest land situation in some areas is not balanced,and there is still a deterioration.It is necessary to strengthen the effective implementation of relevant protection management policies according to local conditions and strengthen the monitoring and evaluation in the later period. |
PDF Full Text Request