Research On High Accuracy Surface Modeling Based On Parallel Stochastic Neural Network

Surface reconstruction is one of the most fundamental problems in 3D reconstruction,which has a significant impact in the area of GIS.High accuracy surface modeling(HASM)solves the error problems in surface reconstruction that comes from the most commonly traditional surface modeling methods.Despite the strengths of HASM,there were some application defects:implementation of HASM relied on the high accuracy input driving field,the data of which were usually hard to obtain;In HASM,the unknown point values were calculated based on the limited sampling point data,which ignored the basic parameters(historical data)of the model;The generalization ability of HASM would be greatly reduced if there were missing data or big errors.For the questions mentioned above,we conducted the following research works based on Neural Network.(1)To obtain high accuracy driving field data,the parallel stochastic neural network is proposed to fit the complex surface.The neural network has a strong function approximation ability and is able to approximate any continuous function with arbitrary precision,especially the nonlinear relationship,which is the key to reconstruct some complex surface,especially the ones that are difficult for conventional mathematical methods to approach.(2)In order to improve the prediction ability of the model,a reinforcement learning method based on data fusion is came up with to predict the spatial information of the surface.First,the historical data is used as the trained data to incorporate into the calculation process of the initial field to produce the trend data.Then,combined with the existing sample data,an optimal solution is found between the initial field and the sampling point data.Finally,the new observed data are adopted as the parameters of the enhanced signal to verify the model,as well as correct and further improve the prediction ability of the model.(3)To further improve the generalization and robustness of a model,a data weighting strategy based on Dropout is proposed,in which partial sampling data or network nodes are temporarily discarded according to a certain probability.This strategy can eliminate the abnormal data influence,and therefore effectively avoid overfitting,which improve the generalization ability of the model.To sum up,based on parallel stochastic neural networks,we proposes a more robust modeling method for surface modeling which combines both historical data and limited sampling point data.The accuracy of the model is verified by the climate factors such as temperature and precipitation.According to the analysis of the experimental results,the surface modeling method based on parallel stochastic neural network not only improves the accuracy,but also improves the generalization of the model.