| With the intensification of energy crisis and environmental pollution,a large number of photovoltaic(PV)power plants have emerged in China with the promotion of government related policies.Obtaining PV power plants and dynamic change information quickly and automatically using remote sensing technology has important economic,social and ecological benefits for the government management departments,such as the rational use of land resources,energy utilization,poverty prediction and environmental protection.Automatic extraction of man-made objects from remote sensing images is an important research area in pattern recognition.However,the spectral uncertainty of terrestrial objects resulting from the complexity of the environment and remote sensing image processing presents challenges for traditional algorithms for object extraction based on spectral statistical properties of training samples.The major obstacle is that the algorithm has poor generalization capability.Although deep learning has shown great potential in the field of image scene recognition,the small sample data in reality is still the main problem.In this context,we use support vector machine and convolutional neural network algorithm to extract PV power plants from the perspective of spectral matching,multi invariant features combination and scene recognition.Aiming at the challenges brought by dataset shift in remote sensing images with different geographical locations and different imaging times,a lot of experiments are carried out and related algorithms have been improved.The main results are as follows:(1)Based on the spectral signatures of PV power plants on different space positions and different phase images,the decision tree is used for the extraction test.After that,the advantages and disadvantages of classification and clustering algorithms are introduced,and some improvement strategies are put forward.In order to overcome the deficiency of the optimum band index(OIF)and one-class support vector machines(OC-SVM)combined in the extraction of ground objects,a modified optimum band index(MOIF)is proposed by analyzing the difference between the feature extraction and classification.Experiments show that: compared with the traditional algorithm,the accuracy of the improved algorithm is obviously improved.(2)The auxiliary structure and texture features are used to identify the power plants,and the validity of the spatial derived features in the extraction of ground objects in the middle and low resolution remote sensing images is analyzed.In view of the problems existing in the extraction of objects,the paper introduces the advantages of machine learning algorithm on the basis of fully excavating deeper features,and makes use of complementarity of features to extract PV power plants.In addition,for the challenges brought by dataset shift to time series and large scale remote sensing image processing,we try to explore the problem by combining multi invariant features.The experiment shows that the improved algorithm has obvious advantages in generalization performance,noise resistance and so on.(3)In recent years,deep learning has become a powerful tool to solve the problem of complex and informative images.However,designing a new deep learning model is difficult and requires a large number of samples to train,which is often difficult to meet in the practical application.In this paper,convolution neural network(CNN)is selected as classifier.Taking the scene recognition of power plants in medium spatial resolution image as an example,transfer learning is applied to solve small sample problem.The ability and influencing factors of CNN to identify complex PV power plants scenarios under small sample conditions are studied.Experiments show that: transfer learning can effectively solve the problem of small sample in reality.CNN shows strong recognition ability for background dopant,brightness value change,scale change,occlusion interference,inter class homogeneity and intra class heterogeneous image. |
PDF Full Text Request