Research About Urban Land Dynamic Detection Based On DJI Unmanned Aerial Vehicle

With the continuous improvement of the economy and the rapid progress of urbanization,the population has expanded the population at high rate.The research about the dynamic changes of urban areas based on DJI unmanned aerial vehicle(UAV)images will be worth exploring.However,it is an interesting and urgent task to choose a stable and feasible classification strategy to accurately detection the potential difference information.In addition,the differences in lighting,atmospheric conditions,multi-sensor,and ground humidity between the two UAV images make it difficult to align multi-temporal image pairs(e.g.,noise).To address aforementioned problems,based on DJI UAV,this paper develops a two-stage urban land dynamic detection framework by combining registration and detection.The main contributions of this work are:(ⅰ)a SIFT dynamic threshold processing model is designed to gradually identify outliers and maximize the number of reliable inlier pairs,while helps to build a coarse to fine transformation;(ⅱ)a local spatial structure similarity preservation is proposed to constrain the local structure of putative inliers during registration,while works with a global constraint to refine the warping field by coherently moving putative control points:(ⅲ)a dynamic Gaussian kernel is developed to control the displacement distances of feature points such that the transformation is gradually changed from rigid to non-rigid for assisting the above coarse to fine search strategy.(ⅵ)the joint matrix of similar neighborhood pixels is generated by fuzzy c-means(FCM)classifier.Then,filtering technology is used to weaken the limitation of multi-temporal information.And the pre-trained difference matrix of deep restricted Boltzmann machine(RBM)neural network is constructed to realize fast and automatic generation of change result graph.A series of comparative experiments based on DJI UAV dataset are performed.The experimental results on feature matching and image registration are performed and our method outperforms five current methods.And experiments on change detection methods are performed and our method outperforms three current methods.Our method demonstrates higher registration quality in all scenarios.