摘要：
【目的】森林火灾识别是避免森林火灾大面积蔓延的一项重要研究。随着深度学习的快速发展,基于卷积神经网络的模型因其在图像识别领域的优异表现,被广泛应用到森林火灾识别任务当中。然而,基于卷积神经网络的方法通常在标签数据不充分时...展开更多 【目的】森林火灾识别是避免森林火灾大面积蔓延的一项重要研究。随着深度学习的快速发展,基于卷积神经网络的模型因其在图像识别领域的优异表现,被广泛应用到森林火灾识别任务当中。然而,基于卷积神经网络的方法通常在标签数据不充分时,难以取得令人满意的森林火灾识别结果。【方法】本研究提出了一种基于视觉变换网络的自监督森林火灾识别模型(Self supervised forest fire identification model based on visual transformation network),来提高模型在标签稀缺情况下的森林火灾识别精度。具体来说,该模型采用视觉变换网络作为主干网络,通过视觉变换网络中的多头自注意力机制来捕获森林火灾图像的全局信息特征。并且引入自监督学习中的图像重建任务来辅助模型训练,从而减少模型对标签数据的依赖。模型通过对掩盖图像的特征恢复和重建学习相关语义信息。同时,本研究还提出了一种基于傅里叶低频混合变换的数据增强方法来提高模型的泛化性和鲁棒性。【结果】通过开展详细的试验来验证模型的有效性,结果表明,与其他常见的网络模型相比,FFDM模型在森林火灾识别任务中取得了最佳的识别效果,其识别准确率为89.51%,比VGG16网络高13.7%,比ResNet50网络高8.2%,比InceptionV3网络高7.2%。【结论】通过自监督学习辅助模型训练的方法,FFDM模型即使在标签稀缺下依然可以取得不错的森林火灾识别效果。收起
通讯机构：
[Jianjun Li] A;Author to whom correspondence should be addressed.<&wdkj&>College of Computer and Information Engineering, Central South University of Forestry and Technology University, Changsha 410004, China
关键词：
aerial imagery;ultra-high spatial resolution orbital imagery;object detection;YOLOv4;vision transformer;deep learning
通讯机构：
[Xu, Piao-Rong] C;Cent South Univ Forestry & Technol, Coll Comp & Informat Engn, Changsha 410004, Hunan, Peoples R China.
摘要：
A numerical model of carrier saturation velocity and drain current for the monolayer graphene field effect transistors (GFETs) is proposed by considering the exponential distribution of potential fluctuations in disordered graphene system. The carrier saturation velocity of GFET is investigated by the two-region model, and it is found to be affected not only by the carrier density, but also by the graphene disorder. The numerical solutions of the carrier density and carrier saturation velocity in the disordered GFETs yield clear and physical-based results. The simulated results of the drain current model show good consistency with the reported experimental data.
作者机构：
[左志远; 谭建灿; 毛克彪] Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Science, Beijing, 100081, China;[赵天杰] State Key Laboratory of Remote Sensing Science, Institute of Remote Sensing and Digital Earth Research, Chinese Academy of Sciences, Beijing, 100101, China;[谭雪兰] College of Resources and Environment, Hunan Agricultural University, Changsha, 410128, China;[李建军] College of Computer Science and Information Technology, Central South University of Forestry and Technology, Changsha, 410004, China
通讯机构：
[Mao, K.] I;Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Science, Beijing, China
摘要：
In this paper, a novel approach for initializing clustering centers of K-Means algorithm is presented.This method is based on the variance of dimension, which is used as keyword to make a full permutation.The results of the full permutation for the primary and secondary sequence of keyword is divided into k subsets to initialize the clustering centers.Four international datesets are used for testing datasets to test the effectiveness of this algorithm.And this algorithm is examined by numerical simulation.Experiments suggest that the initial clustering centers chosen by the optimization method proposed in this paper are very close to the clustering centers of ultimate convergence after clustering iteration.Compared with the traditional K-Means clustering algorithm, this algorithm increase the rationality of algorithm on the initial clustering center selection and improve the accuracy of clustering results, and the clustering results is more stable as well.
作者机构：
[付秀丽] Information Engineering Institute, Beijing Institute of Petrochemical Technology, Beijing, 102617, China;[左志远] National Hulunber Grassland Ecosystem Observation and Research Station, Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing, 100081, China;College of Information Science & Technology, Beijing University of Chemical Technology, Beijing, 100029, China;[谭雪兰] College of Resources & Environment, Hunan Agricultural University, Changsha, 410128, China;[李建军] College of Computer Science and Information Technology, Central South University of Forestry and Technology, Changsha, 410004, China
通讯机构：
[Mao, K.] N;National Hulunber Grassland Ecosystem Observation and Research Station, Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing, China
摘要：
It is hard to program a traditional forest sub-compartment management with a long programming cycle and high energy input and lack of computer technology support, such, this paper presents a visualization technology of the sub-compartment management procedure based on WF. The paper, with the programming theory of forest sub-compartment management and WF technology, abstracted traditional sub-compartment management as management measures modules(TABLE II) and implemented the sub-compartment management procedures of custom design(FIGURE II) and used MOGRE 3D Render Engine to display the results at last. The test data was from the pure forest sub-compartment of Chinese fir in Youxian Huangfengqiao forest farm, Hunan Province. The result shows the 3D scene of sub-compartment and the structure factors before and after the operation(FIGURE V and FIGURE VI), which has the intelligent and visual characteristics. In this way, it can realize the effective detection for management procedure, which sets a foundation for the future intelligent forest management schemes in forest farm level.
摘要：
As a representative method of swarm intelligence, Particle Swarm Optimization (PSO) is an algorithm for searching the global optimum in the complex space through cooperation and competition among the individuals in a population of particle. But the basic PSO has some demerits, such as relapsing into local optimum solution, slowing convergence velocity in the late evolutionary. To solve those problems, an particle swarm optimization with comprehensive learning & self-adaptive mutation(MLAMPSO) was proposed. The improved algorithm made adaptive mutation on population of particles in the iteration process, at the same time, the weight and learning factors were updated adaptively. It could enhance the ability of PSO to jump out of local optimal solution. The experiment results of some classic benchmark functions show that the improved PSO obviously improves the global search ability and can effectively avoid the problem of premature convergence.
摘要：
In this paper, a novel approach for initializing clustering centers of K-Means algorithm is presented.This method is based on the variance of dimension, which is used as keyword to make a full permutation.The results of the full permutation for the primary and secondary sequence of keyword is divided into k subsets to initialize the clustering centers.Four international datesets are used for testing datasets to test the effectiveness of this algorithm.And this algorithm is examined by numerical simulation.Experiments suggest that the initial clustering centers chosen by the optimization method proposed in this paper are very close to the clustering centers of ultimate convergence after clustering iteration.Compared with the traditional K-Means clustering algorithm, this algorithm increase the rationality of algorithm on the initial clustering center selection and improve the accuracy of clustering results, and the clustering results is more stable as well.
