CAJ | 학술논문

针对土地利用数据库更新规则复杂、不同更新类型与更新规则自动匹配困难等问题，该文提出并构建了一种基于多层感知器神经网络的土地利用要素自适应更新模型。该模型将土地利用要素的每个变更类型判断及更新行为判断过程均设计成一个神经元，同类神经元组织成一个网络层，所有网络层采用全连接方式构成一个多层感知更新策略判断模型。该模型可以自动完成变更类型与更新规则的正确快速匹配，且可根据更新规则的变化，灵活改变知识库并产生新的推理机。试验表明：该模型明显减少了人工交互环节，综合更新效率较各基地软件可以提高30%左右，一次性更新正确率可以提高5个百分点以上，研究成果可为土地利用数据库的高效自动更新提供一整套新的解决方案。
침대토지이용수거고경신규칙복잡、불동경신류형여경신규칙자동필배곤난등문제，해문제출병구건료일충기우다층감지기신경망락적토지이용요소자괄응경신모형。해모형장토지이용요소적매개변경류형판단급경신행위판단과정균설계성일개신경원，동류신경원조직성일개망락층，소유망락층채용전련접방식구성일개다층감지경신책략판단모형。해모형가이자동완성변경류형여경신규칙적정학쾌속필배，차가근거경신규칙적변화，령활개변지식고병산생신적추리궤。시험표명：해모형명현감소료인공교호배절，종합경신효솔교각기지연건가이제고30%좌우，일차성경신정학솔가이제고5개백분점이상，연구성과가위토지이용수거고적고효자동경신제공일정투신적해결방안。
Land use database is the basis for the government departments at all levels to regulate land use, and the currency and quality of land use database directly determine the level and effect of land supervision. However, at present, the land use database updating technology and means are not advanced enough yet. The currency of the land use database significantly lags our economic development level. In light of the automatic matching complexity of change type and update strategy, artificial neural network is introduced into update strategy judgment field. According to the structure and main updating content, from the horizontal, land use database adaptive updating model is divided into land class polygon, linear feature and isolated feature. Then, in accordance with annual update implementation program of land use database and current updating progress, methods and habits, from the vertical, the above-mentioned updating model is divided into spatial analysis layer, input layer, change type judgment layer, spatial update strategy judgment layer and attribute update strategy judgment layer. Land class polygon is the first and the most important layer for land use database to update, so its updating model is designed on the top of the land use database adaptive updating model. To judge the change type and update strategy, 3 input conditions and 12 neurons are set up in land class polygon updating model, among which 4 neurons are responsible for judging the change type, 6 neurons are responsible for judging the spatial strategy, and 2 neurons are responsible for judging attribute strategy. Compared with land class polygon, linear feature updating model is more complicated. Therefore, linear feature updating model has 6 input conditions and 12 neurons, among which the distribution of neurons is the same as that of neurons in land class polygon updating model. Isolated feature belongs to one-dimensional element, so its updating model is relatively simple. In isolated feature updating model, 4 neurons are arranged to judge the change type, 3 neurons to judge spatial strategy, and 2 neurons to attribute strategy. The same-type neurons of adaptive updating model are organized into a network layer, and all layers are organized into a multi-layer perception network in an all-connected way. In addition, in order to realize the judgment of change type and update strategy, a threshold activation function is arranged in each neuron, and between the neurons connection weight is set up to adjust the input of neuron active function. All kinds of training methods of multi-layer perception neuron network are analyzed in a comprehensive and deep way. Change sample data of each element are collected so that updating model training can be carried out. Spatial and attribute update rules of various land use elements under the conditions of different topological relationships and different properties should be studied. And the knowledge and experience are organized into an update knowledge database. When new change survey data are input into model again, topology, property and linkage effect of the input sample are gained through the spatial analysis function, and then are organized in the form of vector. According to topological, property and linkage characteristics of new input sample together with learned experience, adaptive updating model draws a correct inference to spatial update strategy and attribute update strategy of new input sample, and eventually offers corresponding update strategy of sample in the form of vector. After the research subject is finished, updating application and technology tests are carried out respectively in 7 test bases, including Laoshan, Laixi, Huangdao of eastern Shandong, Liuyang and Yizhang of central Hunan and Chengdu and Wenjiang of western Sichuan, and the desired effect has been achieved. The tests show:Firstly, adaptive updating model of the land use features that the paper constructs can quickly complete change type’s judgment of land class polygon, linear feature and isolated feature under any condition, and correctly offer spatial update strategy and attribute update strategy of all features. Compared with all test bases’ software, our new software based on this updating model can comprehensively improve the updating efficiency by about 30%, and enhance one-time updating correct rate by more than 5%. Secondly, adaptive updating model of the land use features that the paper constructs has a strong adaptability. When update rules vary, it only needs to conduct a second sample training of the model, and the correct update strategy will be given without modifying software, thus improving soft adaptability. Thirdly, different from common updating software, in the adaptive updating model that the paper presents, the update strategies of all features are given not by man, but by experts’ knowledge database. Meantime, change methods model need not be driven by man, but automatically driven by update strategy, thus bettering software automation level.