| Developing modern and efficient agriculture,accelerating the progress of agricultural science and technology,improving the level of agricultural mechanization and informatization,and making agriculture becomes a leading industry.these are important measures to ensure the realization the dream of agriculture,countryside and farmers.Agricultural greenhouse represents the development direction of modern agriculture.Timely and accurate monitoring and acquisition of spatial distribution information of agricultural greenhouse can provide decision-making basis for agricultural management,environmental protection,soil pollution and other issues.Aiming at the characteristics of high spatial resolution and rich texture information of GF-2 satellite image,this paper proposes an object-oriented multi-feature fusion method to extract the area of agricultural greenhouse.Comprehensive utilization the features of spectrum,texture,edge detection,threshold segmentation and mathematical morphology.Firstly,enhancing the image,which the buildings and roads are removed by combining the spectral and texture features of the greenhouse.Threshold segmentation is used to delete the "noise" after edge detection.Morphology is used to improve the efficiency of image segmentation.Finally,the five shape feature parameters of Ar,Per,Rd,Pwl,Pr are used for some difficult remove "noise" to eliminate.This method and the support vector machine algorithm module in ENVI are used to compare and analyze the accuracy in the same experimental area,which proves that the extraction method designed in this paper has great advantages.In the GF-2 satellite images,210 true-color multi-class samples and 210 false-color multi-class samples with the size of 1000 × 1000 were produced respectively,which marked 7608 agricultural greenhouses,130 true-color single-class samples and 130 false-color single-class samples with the size of 1000 × 1000 were produced respectively,which marked 7426 agricultural greenhouses.SSD algorithm model based on TensorFlow framework to identify agricultural greenhouses.Using the transfer learning and the mature VGG-16 model,four sample sets were trained in two versions of SSD-300 and SSD-512,respectively.Finally obtained the true color one-class optimal model of SSD-300 and the true color multi-class optimal model of SSD-512.Random selection of easy,medium,difficult three types of size of 1000 *1000 each 20 image data.Through statistical averaging,obtaining the accuracy of two optimal models in the three types of image detection and recognition of agricultural greenhouse.The research results show that using object-oriented multi-feature fusion method to extract agricultural greenhouses and using transfer learning and SSD algorithm to detect and identify agricultural greenhouses have great application potential.This is a hot research direction in the future.Which can provide technical support for statistics of the overall spatial distribution of agricultural greenhouses,scientific and effective management,guidance of agricultural greenhouse construction and policy formulation. |
