Abstract
Land-cover/land-Use has become crucial basis work to carry out the prediction to the dynamical change of land-use, prevention to natural disaster, environment protection, land management and planning. With rapid development of remote sensing technology, especially deeply studies in remote sensing geo-analysis modes, remote sensing land-cover/land-use classification has become the most credible, rapid and effective measure to monitor the condition and changing of land-cover/land-use in the ;global surface. In this work, firstly authors will give the review of traditional land-cover/land-use RS classification. Traditional land-cover/land-use remote sensing classification methods mainly include three models. They are mathematical and statistical theory based classification model, neural computation theory based classification model, and symbolic logic reasoning based classification model Under these methods, authors then propose the remote sensing intelligent Geo-Analysis model names remote sensing intelligent geo-interpretation model (RSIGIM).In so that the Geo-Decision knowledge can be structurally and parametrically fused in. Based on this model we will build the new systematic structure of land-cover/land-use RS classification with the example case in Hong Kong Island by multi-dimensional spatial; data fused between the multi-platform remotely sensed data and ancillary geography data.
Introduction
Land-Cover/Land-use, being the new concept developing with the remote sending technology, has become a crucial item of basic tasks in order to carry through a series of important works, such as the p0rediction of land-use change, prevention of nature disaster, management and plan pf lam -use, protection of environment, etc,. With the more thorough development of remote sensing technology and Geo-Analysis model, using remotely sensed data to monitor the status and dynamical change of land-cover/land-use is become the one of the one of the most rapid, credible and effectual method.
Land-cover and Land-use are two different concepts in its intrinsic signification .Land-cover emphasize particularly on its nature properties and it is the synthetically reflection of various elements in global surface covered with natural body or manual construction. Using remote sensing classification method, whatever used or non\-used covering object in surface can be separated. However, Land-use, emphasizing more on land's social properties, is the output of reconstruction activities that human adopts a serials of biologic, technologic measure to manage and regulate the land chronically and periodically according to determinate economic and social purpose. Thus, land-use is a process of turning natural ecosystem into social ecosystem, and the process is a complicated procedure by the synthetic effect from nature, economy and society. The manner, degree, structure, area distributing and benefit of land-use are not only affected by natural condition nut also restricted by diversified natural, economic and technologic condition, and in sometimes among all factors the social production form is determinant Land-use is the most direct and leading driving factor to the land-cover change.
In carrying out research and application of the land-covert and land-use remote sensing investigation, the uniform classification system is usually built up by combining the two concepts under one system, which is called Remote Sensing Land-Cover/Land-Use classification system. IN this paper, we will initially build the intelligent land-cover/land-use classification system that is supported with the proposed model named Remote Sensing Intelligent Geo-Interpretation Model (RSIGIM) basic on the data fusion among multi-platform remote sensing data, geographical ancillary data and geo-knowledge. The multi-platform remote sensing data includes TN,SPOT-HRV, and the geographical ancillary data include topological data and land-cover/land-use classification methods are reviewed. Then in third part, the RSIGIM will be presented out and its land-cover/land-use classification system also is discussed in this part. The application and output evaluation will be talked about in part four. In the end, we will tine the conclusions and prospects.
Conventional Land-Cover/Land-Use Classification Methods
Conventional land-cover/land-use remote sensing classification methods mainly include traditional classification methods, neural networks based classification methods and symbolic knowledge based logic reasoning classification method .
Traditional Land-cover/Land-use RS Classification
Traditional remote sensing image classification can be classified into rationalization method, hierarchical classification method, decision-tree classification method, and statistical classification method, etc. Every method has its geo-background. For example, rationalization and hierarchical classification, mainly according to the geo-difference rule in remote sensing image, are to step wisely classify the image from sketchy degree to subtle degree to subtle degree by different classification decision rules in different hierarchy, such as the decision tree mode by landscape difference in remote sensing image. These traditional classification method require that the one who want to do classification should master the geo-regularity in the remote sensing image to a sufficient degree, or the outcome is difficult to reflect to reflect the true geo-distribution. Therefore, the limitation to the traditional method is the difficulties to hold the criterion of classification and qualitative component in the method is primary.
Based on mathematical and statistical analysis model, a lot of remote sensing classification and information extraction methods already has been developed The typical methods include ISODATA clustering, minimum distance methods, maximum like classification, etc, which are all based on parametrically or non parametrically statistical distribution model to do the basic processing and analytical process to the remote sensing image by mathematical computing and statistical analysis. With MSC model and the fact that the statistical difference exists between datum in feature space, the primary cognition to the RS image including the basic land-cover unit classification and information extraction can be realized. But for no integrating with geo-knowledge MSC is difficult to truly reflect come special geo-distribution and especially the defect of being difficult to gain the parameters of the parametric model is existing when processing or analyzing the complicated spatial information.
Neural Networks Based Land-Cover/Land-Use Classification
Comparing to the traditional classification methods, artificial neural networks (AQNN)based classification method simulates the human vision and neural processing system to get the cognition to the RS image. Neural-network classifiers are non-parametric and therefore may be more robust when distribution are strongly non-Gaussian. During training, the network to capable of forming arbitrary decision boundaries in the feature space. With distributed knowledge represented in the neural network and the knowledge got by training the priori samples of datum the neural network, can be get complicated and nonlinear mapping ability to realize the land-cover hierarchical classification from coarse degree to subtle degree with scale space changed. Thus, ANN method generally can get more high accuracy of the outcome and have been widely used in land-cover/land-use classification. Especially ANN's superiority is showed to the complicated land type.
Since the publication of Rumellhart and McClelland(1986),there has been a renaissance of interest in using neural network, particularly back-propagation neural networks, for land-cover/land land-use classification. Howald(1989).McClelland (1989), Hepner (1990), T.yoshida(1994),K>S>Chen(1995),J.D.Poala(1997),etc, all used ANN method in the land-cover classification by TM image, and in vary degree the accuracy is improves. Kanellopoulos(1992) used FANN method to classify the land -cover into 20 type with SPOT-HRV image, and its outcome got better accuracy than general statistical method. By ANN method, G.M.Foody(1996) separated the mixed pixel in land-cover. Under the basis of TM remote sensing data, spatial context information data, ancillary information (including DEM and slope), L.Bruzzpne (1997) classified the complicated land-cover using ANN method and 9%improvement in accuracy comparing to the maximum like hood classifier.
Despite its strong non linear mapping ability to processing and analyzing the remote sensing information with complicated spatial distribution, most ANN methods are also just a computing behavior like general computing models. Therefore, ANN method without being supported by geo-knowledge also can nota truly reflect the distribution character of some special type of phenomenon in remote sensing image.
