WO2006120724A1 - Geographic information system using neural networks - Google Patents

Abstract

[PROBLEMS] To provide a geographic information system which has observation data having geographic information for an SST estimation and a parameter as space data and which can execute the SST estimation over a wide range while reducing the amount of calculation and the error. [MEANS FOR SOLVING THE PROBLEMS] The geographic information system is given a teacher set and made to learn it thereby to constitute a neural network, and then is made to do an estimation with input data. Thus, the geographic information system is enabled to make a subsequent estimation, by inputting the teacher set and causing it to be learned, but not by the user's application of an estimation type algorithm from much space data, so that the system can acquire an analytic result easily.

Description

 Specification

 Geographic information system using neural network

 Technical field

 The present invention relates to a geographic information system that accumulates and utilizes spatial data, and more particularly to a geographic information system to which a neural network is applied.

 Background art

 [Calculation method for sea surface temperature (SST) estimation: thermal infrared radiometer]

 Sea surface temperature (SST) estimation has been performed using satellite observational data. Various methods have been proposed as methods for estimating SST. Thermal infrared radiometers have been mainly used in SST. This is because the atmospheric influence is small in the observation wavelength band.

 In the case of thermal infrared light, since the wavelength is long, the influence of scattering is not so important, and the influence of azimuthal direction may not be considered.

 From the radiative transfer equation treated by thermal infrared radiation, the observed luminance Ii in the i wavelength band sensitive to the thermal infrared wavelength range is expressed by the following equation (1).

[0004] [Number 1]

Moth

 Where Bi (t): spectral Planck function, Ts: earth surface temperature [Κ], τ i: transmittance, Z: atmospheric upper end height, μ: observation angle, ζ: altitude, Τ (ζ): temperature Represent.

In equation (1), there is a method of estimating SST by solving the inverse problem of finding the earth surface temperature Ts that is the factor from the observed luminance Ii. In this case, the radiation transfer equation (1) is essentially non-linear. If it is a linear inverse problem, there are methods such as least squares method and orthogonal expansion and solving to prevent the divergence of the solution and solve the inverse problem. However, the solution of the nonlinear inverse problem is limited either by linearization with a limited range of solution or solving iteratively. Therefore, the calculation for one data is very complicated and the amount of calculation increases. It is too difficult to use the method of solving this equation (1) for the huge amount of data in all sea area every time. [Calculation method for sea surface temperature (SST) estimation: Split-Window method]

 Besides SST estimation with a thermal infrared radiometer, there is another method called Split-Window method (hereinafter, simply referred to as SW method). This is a multiple regression analysis that uses observation data for each observation wavelength band where the atmospheric effects are different and a regression equation that also provides parameter power for each observation wavelength band. This can be easily done SST estimation if the parameters of the sea area are obtained in inverse problems.

 Although the influence of the atmosphere is different in each wavelength band, the earth surface temperature is the same. As a result, the influence of the atmosphere can be obtained from the difference in the observed brightness temperature in each wavelength band.

 Assuming that the number of observed frequency bands is n, the brightness temperature of i wavelength band (n) is Bi (T), and the corresponding parameters are obtained, the required earth surface temperature To is as shown in the following regression equation (2) become.

[0008] [Number 2]

[0009] In the regression equation (2), the value of the earth surface temperature actually observed in To is substituted for the corresponding observation data into Bi (T) to obtain the inverse square to obtain the parameter Ci in the least square Solve the problem. Thus, the known parameter Ci and the newly-observed Bi (T) force can also easily determine the unknown To.

 [0010] Since the parameter Ci changes depending on the location of the sea area and the season, it is necessary to determine an appropriate parameter for each sea area and each time. However, to calculate the parameters of each sea area on the whole earth, the amount of calculation is very large. Therefore, until now, the parameters obtained in a specific sea area were used for other sea areas, and errors occurred in each sea area.

 Patent Document 1: Japanese Patent Application Laid-Open No. 2005-52045

 Disclosure of the invention

 Problem that invention tries to solve

[0011] Due to the difference in weather conditions (eg, climate, ocean current, etc.) in each sea area (latitude, longitude), the parameters differ. Therefore, it is necessary to determine the parameters for each sea area.

However, in order to determine the optimum parameters for the sea surface, which accounts for 70% of the earth's surface, the problem is that the amount of data and the amount of calculation become too large. Also avoid calculation It is also possible to use the parameters obtained in a specific sea area as a simple measure for other sea areas

1S Negligible, have a problem if there is an error of a certain degree.

The present invention has been made to solve the above-mentioned problems, and has observation data and parameters for SST estimation having geographical information as spatial data, while reducing the amount of calculation and errors. The purpose is to provide a geographic information system that can carry out estimation widely.

 Furthermore, the present invention is not limited to the implementation of SST estimation, and it is another object of the present invention to be able to extensively implement other estimation using spatial data other than spatial data for SST estimation. Means to solve the problem

 The regression equation of the SW method is for obtaining an output from the correlation of a plurality of input data. To do this, we incorporate the -Uural network (NN) into the GIS for the processing of the SW method.

 Among neural networks (NN), backpropagation (BP) is suitable for regression correspondence. The NN learns to obtain an empirically low error solution by giving repetitive data. In the SW method, measured values at sea level and satellite data force also determine parameters. BP can correct the internal parameters automatically by comparing it with the estimated value from the satellite data, using the measured value on the sea surface as the teaching data. The introduction of NN can improve accuracy and automate.

 Furthermore, GIS enables optimization of parameters. SST is estimated locally and depends on climate, longitude and latitude. Conversely, if geographical information such as climate, latitude, and longitude matches, parameters can be handled collectively. By using appropriate parameters, it is possible to improve the SST estimation of other sea areas. Therefore, it is required to create a database suitable for integrated search and management of geographical information.

[0014] Also, using these systems, it becomes possible to simulate SST estimation in a sea area where measured values and satellite data are unknown. Even if there is no actual measurement value, if the geographical conditions agree with the sea area where the parameters are already known, that parameter can be used.

By using the present invention, it is possible to provide a system for performing SST estimation simulation using SST information by parameter search by GIS. Also, by using the present invention, geographic information Using a system-Constructing a eural network and analyzing method and parameter estimation when using image information with geographical information as input-easing by eural network and optimizing analysis parameters according to geographical conditions Can be done.

 [0015] (1)

 According to the present invention, a geographic information system using a eural network includes a recording unit for recording spatial data, which is data having information on position, and analysis means for analyzing the spatial data of the recording unit. In the analysis means, input data of the teacher set is read into the input layer, forward calculation is performed, and backward calculation is performed by using the output data of the output layer of the calculation result and the teacher data of the teacher set to learn. -Reads the input data to the input layer of the Euler network, and finds the estimation result of the read input data. As described above, according to the present invention, in the geographic information system, a teacher set is given and learned, and after constructing a eural network, input data is input and estimated, so in the geographic information system, a large amount of spatial data is generated. The force user can input and learn a teacher set that does not apply the algorithm of the estimation equation, so that subsequent estimation can be performed and analysis results can be easily obtained. In particular, the input data is recorded in the recording unit of the geographic information system, and the output data is recorded in the same recording unit, thereby constructing a system in which the neural network and the geographical information system are integrated integrally, It enables quick analysis.

 As in the conventional case, if SST is calculated using a non-linear equation or a regression equation, the algorithm becomes complicated and the amount of calculation increases. Compared to the prior art, the present invention can obtain an estimation result quickly as a simple algorithm.

[0016] (2)

According to the present invention, the geographic information system using the eural network is provided with a spatio-temporal search means for performing a spatio-temporal search as necessary, and estimation of a certain time of a certain position is performed at the same time before the last year. It is estimated using weights. As described above, in the present invention, it is searched whether or not there is a weight at the same position and the same position as the position designated by the space-time search means, and if there is a matching weight, the weight used is used. In this way, estimation can be performed without learning. Weights are sometimes called combined load, load . Here, “before last year” means before “last year” and is used in the sense including last year

[0017] (3)

 According to the present invention, the geographic information system using the eural network is provided with similar area detection means for detecting an area meeting the detection condition specified by the user as the similar area, if necessary, The target data of the condition is recorded as spatial data, and when a part of the similar area has already been learned, it has already been learned and has already been constructed-using the eural network and the other in the similar area It is to estimate for the part. As described above, according to the present invention, since the similar area detection means detects a similar area that matches the detection condition, if a part of the similar area has already been learned, learning is performed in other similar areas. It can be estimated. In addition, even if learning has not been conducted in the similar area, it is possible to efficiently estimate the similar area by learning only a part of the similar area. In addition, it is common practice to give a predetermined initial value as the initial value of the weight-it is a general method in the case of the eural network. It is also possible to use the existing-eural network weight as the initial value of the naive weight, and the convergence of the weight can be performed quickly by the teacher set, and the existing-eural net leak weight is used as it is. Inference can be performed with higher accuracy than the inference used.

[0018] (4)

According to the present invention, the geographic information system using the eural network can adjust the detection condition for detecting the similar area as needed. As described above, in the present invention, the detection condition for detecting the similar area can be adjusted, and the user can freely adjust the similar area and the accuracy. That is, if the accuracy is not required, the detection condition is relaxed and the similarity region is expanded, and if the accuracy is required, the detection condition is strict and the similarity region is reduced. For example, the detection condition is "an area 50 to 70 km from land, the Kuroshio or the Tsushima Current, the latitude 20 to 50, and the longitude 120 to 150". If there is a problem, it is possible to adjust by changing the value of the numerical condition or eliminating the condition itself in cases other than the numerical condition. Make adjustments easy In order to present the user with similar areas of detection results on the display unit, it is preferable to perform layer display to facilitate comparison with similar areas detected earlier. .

 [0019] (5)

 According to the present invention-The geographic information system using the eural network detects at least one sample point for each of the obtained similar areas, receives the input of the teacher set for the sample points, and inputs the input teacher set. Data has already been constructed-reasoning is performed using a eural network, output data is compared with teacher data of a teacher set, and the power can be regarded as a similar area depending on whether the error is within a threshold or not To determine the As described above, according to the present invention, the sample point is detected for the area detected as the similar area only by detecting the similar area based on the detection condition set by the user, and the teacher set for the carious sample points The user is requested, the user infers the input data of the teacher set inputted in response to the request, inference is made using the existing -Uural network, the output data is compared with the teacher data, and the error is within the threshold. If so, it is regarded as a similar area, so eligibility is confirmed as a similar area, and it is possible to prevent inference for false similar areas in advance.

[0020] (6)

According to the present invention, a geographic information system using a eural network includes: a storage unit for recording, as spatial data, a far-infrared image corresponding to a position obtained by remote sensing by an artificial satellite; The analysis means is provided with an analysis means for analyzing the spatial data, and the analysis means gives a far infrared image of the position at the input layer, performs forward calculation, and outputs the sea surface temperature of the output layer and the sea surface temperature as teaching data. It is constructed for each position by performing backward calculation by learning-A far-infrared image is input to the input layer of the Euler network, and the sea surface temperature of the input far-infrared image is determined. As described above, according to the present invention, in the geographic information system, the far-infrared image is recorded as spatial data in the recording unit, the analysis unit learns using the teacher set, and the far-infrared image is input. An appropriate estimate can be made using the constructed-Eural network to determine the sea surface temperature corresponding to the input far-infrared image. [0021] (7)

 According to the present invention, the geographic information system using the eural network includes two far infrared images at the same position in different frequency bands, and the same teacher data as the far infrared image. It is the measured data or the satellite observation data of the sea surface temperature at the location. As described above, in the present invention, since the sea surface temperature is determined using different far-infrared images of frequency bands at the same position, the temperature can be determined more accurately.

[0022] (8)

 According to the present invention, the geographic information system using the eural network can, if necessary, determine the position already learned when there is a position where the user specified position and the detection condition match. The sea surface temperature is determined using the estimation. As described above, according to the present invention, when the user designates the position when obtaining the sea surface temperature, the neural network has been trained in the other position where the corresponding position and the detection condition match. In the case where C is constructed, learning is not performed directly for the specified position. Since the sea surface temperature is estimated using the Euler network at other positions where the detection conditions are met, unnecessary learning Can quickly estimate the sea surface temperature

[0023] (9)

 A method applied to a geographic information system using a eural network according to the present invention is applied to a geographic information system that analyzes spatial data of a recording unit that records spatial data, which is data having information on position. The process of reading the input data of the teacher set in the input layer and performing forward calculation, and the process of the process of causing the forward calculation, performs backward calculation using the output data of the output layer and the teacher data of the teacher set. And the step of reading the input data into the input layer after the step of learning to obtain an estimation result.

[0024] (10)

According to the present invention, the geographic information program using the eural network comprises: a processor for storing spatial data of a recording unit for recording spatial data which is data having information on position; A geographic information program that analyzes data, in which the processor reads the input data of the teacher set in the input layer and makes forward calculation, and the output of the output layer is the calculation result of the procedure in which the processor makes the forward calculation. A procedure for performing backward calculation and learning based on the data and the teacher data of the teacher set, and a procedure for reading the input data to the input layer after the learning procedure and for obtaining an estimation result are executed.

Brief description of the drawings

FIG. 1 is a block diagram of a geographic information system according to a first embodiment of the present invention.

FIG. 2 is a diagram of a euron model according to the first embodiment of the present invention.

FIG. 3 is a structural diagram of a hierarchical-eural network according to the first embodiment of the present invention.

FIG. 4 is a structural diagram of an interconnection-eural network according to the first embodiment of the present invention.

 FIG. 5 is a structural view of a perceptron according to the first embodiment of the present invention.

 FIG. 6 is an explanatory view of forward calculation of back propagation according to the first embodiment of the present invention.

 FIG. 7 is an explanatory diagram of backward calculation of back propagation according to the first embodiment of the present invention.

 FIG. 8 is a processing image diagram of SST estimation by the geographic information system according to the first embodiment of the present invention.

 FIG. 9 is an explanatory diagram of a geographic information system according to the first embodiment of the present invention.

 FIG. 10 is an operation flowchart of the geographic information system according to the first embodiment of the present invention.

FIG. 11 is an operation flowchart of the geographic information system according to the first embodiment of the present invention.

FIG. 12 is an explanatory diagram of similar area detection in the geographic information system according to the second embodiment of the present invention.

 [Fig. 13] This is an observation image of channel 4 at 17:11 on May 20, 2002.

[Fig. 14] This is an observation image of channel 5 at 17:11 on May 20, 2002.

[Figure 15] It is MCSST at 17:11 on May 20, 2002. [Fig. 16] This is the double error of the number of times of learning and output from the data off Ogura at 17:11 on May 20, 2002.

 [Figure 17] This is the double error of the number of times of learning and output from the data at Amagasaki offshore at 17:11 on May 20, 2002.

 [Fig. 18] This is an observation image of channel 4 at 16:45 on Nov. 30, 2002.

 [Fig. 19] This is an observation image of channel 5 at 16:45 on Nov. 30, 2002.

 [Figure 20] It is MCSST at 16:45 on November 30, 2000.

 [Fig. 21] This is the double error of the number of times of learning and output from the data off Ogura at 16:45 on November 30, 2002.

 [Fig. 22] This is the double error of the number of times of learning and output by the data off Kashiwazaki at 16:45 on Nov. 30, 2002.

 FIG. 23 is a program list of modules of the neural network according to the embodiment.

 FIG. 24 is a program list of modules of the neural network according to the embodiment.

 FIG. 25 is a program list of modules of the neural network according to the embodiment.

 FIG. 26 is a program list of modules of the neural network according to the embodiment.

 FIG. 27 is a program list of modules of the neural network according to the embodiment.

 FIG. 28 is a program list of modules of the neural network according to the embodiment.

 FIG. 29 is a program list of modules of the neural network according to the embodiment. Explanation of sign

[0026] 10 databases

 20 control unit

 21 Display means

 22 Search method

 23 Analysis means

 30 Display

 40 input unit

 BEST MODE FOR CARRYING OUT THE INVENTION

The invention can be implemented in many different forms. Therefore, in each of the following embodiments, The description should not be interpreted alone. In addition, the same symbols are attached to the same elements throughout the respective embodiments.

 In each embodiment, the system will be mainly described, but as is apparent to those skilled in the art, the present invention can also be implemented as a computer usable program and method. Also, the present invention can be implemented in hardware, software, or software and hardware embodiments. The program can be recorded on any computer readable medium such as a hard disk, a CD-ROM, a DVD-ROM, an optical storage device or a magnetic storage device. In addition, programs can be recorded on other computers via a network

First Embodiment of the Present Invention

 The geographic information system according to the first embodiment of the present invention will be described with reference to the drawings. The geographic information system according to the present embodiment in FIG. 1 is a far infrared image corresponding to a position obtained by remote sensing by artificial satellites. A database 10 for recording spatial data, an analysis means 23 for analyzing the spatial data in the database 10, and a search means 22 for searching the spatial data. It is constructed for each position by providing an infrared image, performing forward calculation, and performing backward calculation and learning from the sea surface temperature of the output layer of the calculation result and the sea surface temperature serving as teacher data. Then, the far-infrared image is input, and the sea surface temperature of the input far-infrared image is obtained.

 The control unit 20 reads spatial data from the database 10 through the user's operation of the input unit 40 in addition to the search means 22 and the analysis means 23 in addition to the search means and the analysis means 23. It has display means 21 for displaying. The search result of the search means 22 may be displayed on the display unit 30 by the search means 22 itself, or may be displayed by passing the search result to the display means 21. The analysis means 23 is also the same.

The search means 22 can perform attribute search and space search. Attribute search is a search for attribute data in spatial data, and the search expression specified by the user by operating the input unit 40 differs depending on the attribute. Spatial search means map data other than attribute data in spatial data Search for etc. In addition, attribute search and spatial search can both be performed in terms of time, and in particular, the search performed in consideration of time in spatial search is called space-time search. In the geographic information system, these attribute search, space search and space-time search are well known techniques and can be applied to the system as needed by those skilled in the art.

 The analysis means 23 can perform vector overlay analysis, point-in-polygon analysis, knocking, geographical measurement, point operation, neighborhood operation, topographic analysis, network analysis, Borony division and the like. Vector overlay analysis is an analysis that performs polygon processing on geometric elements of points, lines, and faces. Point-in-polygon analysis is processing to determine whether the points that make up a line or surface are included in the interior of the polygon. Knocking is about creating the required zones around a specific feature. Geographical measurements are distance measurement, area measurement, and volume measurement, which measure Euclidean distance according to the Pythagorean theorem. Point operations perform mathematical operations (addition operation, multiplication operation, maximum value operation) on the attribute values of cells in the same position. The neighbor operation is to perform various operations on the neighboring cells of one cell and the active cell. The Boronoli division is to divide the area in the sphere of power. These analyzes are also well known techniques and can be applied to the system as appropriate by those skilled in the art.

 The database 10 records, as spatial data, input data of the input layer, intermediate data of the intermediate layer, output data of the output layer, and a teacher set. Therefore, these data can also be searched and analyzed in geographic information systems. In addition, it is possible to achieve excellent performance by recording data in the database 10 in the system rather than inputting data from the outside as external data.

 Next, data used for sea surface temperature (SST) estimation will be described. The SW method requires a plurality of thermal infrared observation wavelength band data. There is a meteorological observation satellite NO AA series operated by the United States Ocean Atmosphere Administration (NOAA). For the NO AA series

 There are five, NOAA12, NOAA14, NOAA15, NOAA16, NOAA17, which are used for various earth observation tasks.

The optical sensor mounted on NOAA is AVHRR (Improved High Resolution Radiometer). Its performance is: resolution l.lkm, observation width 2800km, number of pixels 2048 pixels / line, density gradation 10 bits. The observation wavelength bands of each channel and the main observation targets are as shown in Table 1 below. With NOAA12 NOAA 14 does not have channel 3A, NOAA 15 NOAA 16 NOAA 17 is a channel

Switch between 3A and 3B day and night.

 [3⁄41]

Among the above six channels, channels 4 and 5 are used in the present invention.

 In the following, first, the neural network and the GIS will be described one by one.

Neural Network (NN: Neural Network)

 A neural network (NN) is a model that is modeled on neurons in an organism. The junction between a neuron and another neuron is called a synapse. Each-Euron learns by linking each neuron by synapse and updating the synapse connection weight.

<o

 Given the evaluation function, the feature is that NN learns to automatically output appropriate output.

 Unlike the commonly known solution method and the programmed system, the solution method is learned empirically.

 Next, synapse learning will be described. Make up the NN-The euron model looks like Figure 2.

 The sum of the products of the input xi and the connection weight wi of each synapse is called a net value (equation 3).

[0038] [Number 3] i ', w; Each-euron has a value Θ called a threshold.

 By the non-linear function net) which also has a net value and a threshold repulsion, there are various forces in the function net). Most of the forces have saturation characteristics so that the value of the output y falls within a certain range.

 Back-propagation (BP), described below, requires a continuously changing differentiable response function. Thus, a continuous monotonically increasing sigmoid function is used. Specifically, a logarithmic function (equation 4) is used as the sigmoid function.

 [0040] [Number 4]

> j = f {nrl I.

1 "end f ^: J'p (", r-'t-Θ)))

[0041] The synapse learning is performed based on Heb's learning rule "When cell A is excited, if cell B is always excited, the synaptic weight from cell A to B is large." We use the squared error as an evaluation formula for learning. Assuming that the teacher signal is t, the error E is

 [0042] [Number 5]

1 _t · · · · (5)

 /? = yf

[0043] From equations 3, 4 and 5, E is a function of wi. Learning neural networks is equivalent to finding wi so that the error E is minimized. Therefore, let the small positive constant 7? Be a learning coefficient, and let the correction value Awi of the connection weight be the following equation 6.

 [0044] [Number 6]

Δ, — —... (6)

<) h 8 rh t

:::: ry y} f {nt) T ₃

 At this time, in equation 6, (net) is derived from equation 4

[0045] [Number 7] f'inet) --- y {\-) · · · · (7 / Thus, equation 6

[Number 8]

«« ',-> 2> i (t-y iyi i-y)., ... (8)

 It becomes.

 Equations 6 and 8 are called Generalized Delta Rules.

Next, classification of neural networks (NN) and various features will be described (see FIGS. 3 and 4). There are several types of NNs that have been repeatedly improved. Depending on the type of NN, there is a problem suited to the solution method. A back propagation neural network is suitable for the SW method and the present invention which handles a large amount of data. Each NN is divided into two according to the following four features.

 <Classification by Learning Algorithm> Correct data to be compared with output data at the time of learning is called teacher data. Depending on whether teacher data is prepared, it is classified as having a teacher signal and without a teacher signal.

 Classification by structure> NNs, each of which includes N.Euron in each layer and processes each layer, is called a hierarchical-eural network. Without being divided into each layer-von Eun is linked so that the NN is called an interconnected-Eural network.

 <Flow of input signal> An NN that flows data forward only is called a feed-forward-user network. Furthermore, NN that flows data backward is called a feedback-type eural network.

 <Analog and Digital> There are hardware-like analog and digital NNs.

 However, the analog type is not used so much at present, and the mainstream is digital type.

Next, perceptron will be described. Originally developed-the eural network. Has a teacher signal, performs hierarchical, feedforward processing. As shown in Fig. 5, it consists of an input layer, multiple intermediate layers (hidden layers), and an output layer. The input data propagates -Eron of each layer and the output is finally determined. If there is a teacher signal, the output layer-Eron updates the synapse coupling weight, comparing with the output. The synapse learning is only the output layer-Euron, which is in direct contact with the teacher signal, and learning is not performed in the middle layer. Therefore, even if learning is repeated, nonlinear problems can not be solved. Next, knock propagation (BP) will be described. It is currently the most representative-eural network. The error back propagation algorithm is incorporated into the perceptron and the feedback type is improved. The operation for input is the same as perceptron, and it is called by BP as forward operation (see Fig. 6). The learning of the newly made middle layer and input layer-euros is performed from the output layer toward the input layer, and is called backward calculation (see Fig. 7). It requires a lot of input and output data for learning. It is suitable for pattern recognition and can solve nonlinear higher-order problems.

 As for the operation and learning of knock propagation (BP), first, when there is an input, a forward operation is performed. The synapses of each layer determine the output of Eron by the sigmoid function with respect to the sum of the inputs and transmit it to the next layer.

 Next, a teacher signal corresponding to the input vector is provided to the output layer. And learning of the weight by error reverse propagation is performed.

 Let K be the number of euros in the output layer, and some be ユ ー ロ k and the output ok and the teacher signal tk. The output of the output layer, the teacher signal and the error sum are defined as the following equation.

[0052] [Number 9]: '—

 [0053] Here,

 [Number 10]

<3⁄4 = / ".) (Output! New output of 3) · · · · '(1 0)

[Equation 11] ίΜ,: ■: ■ :: 'u' 3⁄4 (output! G of a neuron's force fn) '· · · (1)) change in synaptic weight between the middle layer and the output layer According to the general delta delta rule (Equation 8), the quantity is as shown in Equation 12.

[0055] [Number 12]

_{? Δι _ = i] (t} k one ^{_{o k) f f (ri €}} i k) y t · · · - (1 2)

::-- _K ) y, In the same way

 [0056] [Number 13]

(Intermediate question β output of neutron j) (1 3)

[0057] [Number 14]

: Assuming the torture neuron (14), the amount of weight change between the middle layers can be obtained as follows.

[0058] [Number 15]

(15)

 '

[0059] Geographic Information System (GIS)

 GIS uses maps based on specific themes called thematic maps, organizes data from each area, and extracts quantitative information on phenomena from multiple thematic maps (see Figures 8 and 9).

At the Geographical Survey Institute of the Ministry of Land, Infrastructure, Transport and Tourism, “Geographic Information System (GIS) comprehensively manages and processes data (spatial data) with information on location based on geographical value location, It is a technology that visually displays and enables advanced analysis and quick judgment. ” In order to manage and analyze each theme in an integrated manner, it is necessary to construct a database that stores various information about geography [0060] Next, thematic maps (thmatic maps) will be described. A map of special purpose made on a specific subject, such as vegetation, artificial, railway network, religion etc., is called thematic map. Multiple pieces of information (land use, settlements, major buildings, transportation networks, administrative district boundaries, place names, etc.) are included.

, It is distinguished from the general map (general map) which serves various purposes.

 In the case of sea surface temperature estimation maps, thematic maps deal only with the water temperature distribution for each region, and remote sensing data from satellites deal with far infrared images.

 Next, the database will be described. It systematically organizes and stores the stored data centrally and makes it available from programs. Link to other databases to enable complex analysis and processing.

[0061] Roads and structures on a map and those with spatial locations such as municipal boundaries are collectively referred to as features.

 A digital representation of information about features is called spatial data. All spatial data carry some form of spatial location information.

 In spatial data, information representing spatial position and shape is called graphic data. Besides graphic data, various information related to the feature is called attribute data. Attribute data can also include photos and videos, etc. that are not powerful in text or numeric format.

 In order to associate attribute data with graphic data, a symbol generally called an ID is used. In the case of using a relational model database, for example, drawing table (ID, left corner origin coordinates), layer table (layer ID, layer name), sea surface temperature (ID, layer, sea surface temperature, coordinates), ocean current ( It is a schema of ID, layer ID, ocean current name, coordinates) and coast (ID, layer ID, chain coordinates). The schemas listed here are merely examples, and persons skilled in the field of geographic information systems can appropriately modify, add, and delete as appropriate.

 In GIS, attribute data is referenced, aggregated, and various analyzes are performed via ID. This time, when dealing with thematic maps in GIS, the overlapping latitude and longitude in the spatial method will be dealt with in an overlapping manner.

 In the present embodiment, parameters for each sea area are stored.

Next, the operation of the geographic information system according to the present embodiment will be described. First, normal display is performed, for example, as follows. The user inputs the keyboard, mouse, etc. 40 When displaying a designated point (hereinafter referred to as a required point) by using, first, the coordinates of the required point on the display means 21 are obtained, and the drawing segment ID (required drawing segment) to which the force coordinate belongs is specified. Also, the map data and attribute data relating to the required drawing section, and further, the map data and attribute data relating to the adjacent drawing section of the required drawing section are also read out from the database 10 and stored in the buffer. Since what is currently required to be displayed is an area (display area) of a certain distance that can be obtained from the request point, only the geographic information corresponding to the display area from the buffer is displayed on the display screen (this A series of processing is called display processing). Thereafter, when the user moves the required point, the same display processing is performed. Here, even if the required point is moved, display is performed without accessing database 10 if it can be covered by the geographical information in the knotter, and if it can not be covered by geographical information in the knotter, the database 10 is accessed again. Refer to map data and attribute data for the required drawing sections and store in the buffer. When the user requests the enlargement or reduction of the display, it can be treated simply as the enlargement or reduction of the display area. In general, if the user requests the display to be expanded, it can be covered by the geographical information in the current browser, and if the user requests the reduction of the display, it can be covered by the geographical information in the current buffer, Many things.

 The operation of the Euler network function, which is a feature of the present invention, is initially performed by the user using the input unit 40 to designate an estimated position (see FIG. 10, step 101). The analysis means 23 determines whether or not learning has been performed for the designated estimated position (step 102). If it is determined that learning has not been completed, a request for a learning set is issued to the user (step 103). The user specifies the learning set as required (step 104). When the designation of the learning set from the user is received, the analysis means 23 performs defined processing learning using the powerful learning set as input data (step 200).

[0065] If it is determined in step 102 that learning has been completed, or after step 200, a request for observation data is made to the user (step 111). In response to this request, the user specifies observation data (step 112). After receiving the designation of observation data, the analysis means 23 carries out estimation, which is a defined process, using the learned two-eural network (step 300). After estimation, the estimation result is displayed on the display unit 30 (step 121). In the defined process learning (step 200), first, observation data of the learning set is read (see FIG. 11, step 201 below), and learning data is read (step 202). The product of the read observation data and the weight of the input layer and the middle layer is calculated (step 203). Operation of the sigmoid function is performed with the result of the product operation as an argument (step 204). The product of this operation result and the weight of the intermediate layer and the output layer is calculated (step 205). The operation of the sigmoid function is performed using the result of the product operation as an argument (step 206). Using the calculation result (output value of the output layer), the output value of the intermediate layer, and the teacher data as arguments, the amount of change in load between the intermediate layer and the output layer is calculated (step 207). Subsequently, using the output value of the output layer, the weight of the intermediate layer and the output layer, the output value of the intermediate layer, and the teacher data as arguments, the amount of change in load between the input layer and the intermediate layer is calculated (step 208). From this step 207 and step 208, the weights of the new intermediate layer and output layer, and the weights of the input layer and intermediate layer can be obtained and reflected. It is determined whether or not the teacher set is the last (step 209), and if it is the last, the defined processing learning is ended. If it is determined in step 209 that it is not the last one, the process moves to step 201.

 The predefined process estimation (step 300) is performed through steps 201 to 206.

 As described above, according to the geographic information system according to the present embodiment, learning is performed and learning is performed by giving a set of teachers that is not limited to the functions of the usual geographic information system using display, analysis, and search using spatial data. By inputting observation data later, excellent SST estimation can be performed on the geographic information system, and more rapid estimation can be performed. In addition, image data related to SST can be displayed on the geographical information system, and comparison with the estimation results can be easily performed. Furthermore, input data in the input layer, intermediate data in the intermediate layer, output data in the output layer, and a teacher set are recorded in the database 10, which enables rapid learning and estimation.

 [0067] [Data format of observation data, teacher set]

In the present embodiment, the observation data used for estimation and the teacher set used for learning may be a set of numerical data or image data. In the case of image data, it is converted into numerical data by a predetermined conversion formula and used in estimation and learning. [Application target]

 In the present embodiment, the estimation of SST in remote sensing-force applying the Euler network geographic information system The present invention is not limited to this estimation, and can be applied to other remote sensing, and used for analysis other than remote sensing. You can also For example, it can be applied to vegetation, chemical content in the atmosphere, and the like. However, there must be a relationship between the input data for estimation and the estimation results. For example, the present invention can not be applied because it has nothing to do with SST and population density. Whether there is a relationship or not can be found by heuristics, or analysis may find some relationship. In general, if the relationship is strong, even one intermediate layer may require multiple intermediate layers if the relationship is weak.

[Neural Network]

 In the present embodiment, in the operation explanation part, the number of nodes changes according to a force estimation method in which the nodes of the input layer are 2 nodes, the nodes of the intermediate layer are 2 and the nodes of the output layer are 1. For example, the node of the input layer may be 1, the node of the middle layer may be 1, and the node of the output layer may be 1.

 Further, the intermediate layer is not limited to one layer, and may have a multilayer structure. The more accurate the estimation, the more the number of middle classes tends to increase.

[Intermediate layer output value]

 In this embodiment, the layer of the geographic information system and the layer of the neural network can be regarded as the same one, and the database as the graphic data or attribute data of the geographic information system without discarding the output value of the middle layer. It can also be recorded in 10. Also, a set of weights can be treated as a layer. The change of weight itself can be analyzed.

[Estimated result]

 In the present embodiment, the estimation result refers only to numerical data after numerical calculation, but the numerical data is visualized and displayed on the display unit 30 so that the user can compare and evaluate more visually. It can also be displayed.

Second Embodiment of the Present Invention

A geographic information system according to a second embodiment of the present invention will be described. This embodiment The geographic information system according to is configured in the same manner as the first embodiment, and is configured to include a similar area detection unit that detects a similar area that is an area that can be estimated using the weight of the same-eural network.

 As for SST, we have empirically found that it is possible to obtain SST by the same estimation method in a region at a predetermined distance from land and in the same area. Therefore, if it is possible to search for the distance from land and whether it is the same area or not, it is possible to detect the similar area, and the similar area detection means performs an intensive search to detect the similar area. It is carried out. Here, it is necessary to record in advance the data necessary to determine the distance from the land and whether the area is the same area, in the database 10 in advance. By storing this data in a database on the geographic information system, similar area detection can be performed quickly.

 The operation of the similar area detection means is started by the user designating the learned area and instructing the detection of the similar area through the input unit 40. An attribute search or a spatial search is performed for a part having the same area as the designated area and the same area from the land (the search means 22 may be used for the search here). The similar area detection means outputs the area as the search result as the similar area and displays it on the display unit 30. With regard to the strongly similar area, already learned-since the Euranole network can be used, learning that inputs the learning set is unnecessary, and estimation is performed by inputting observation data used for estimation into the geographic information system. can do. Therefore, the learning time can be significantly reduced.

 For example, when similar region detection is performed by specifying the position where the user needs to work when there is a learned region a using the teacher set shown in FIG. 12, the learned region is first included. The similar region b is detected for the region to be detected, and further, the coast of Kashiwazaki, which is only off Fukuoka, will be detected as the similar region c if the detection conditions are met.

Then, for the area detected as a similar area, observation data is collected using an artificial satellite when there is no observation data, and when there is observation data, it is read from the database 10, By performing processing estimation (step 300), the sea surface temperature in the similar area can be determined. Determine the sea surface temperature, and complete the sea surface temperature distribution map for the image and similar area part. It is desirable that all of these series of operations be performed by automation after receiving a detection condition from the user. According to this configuration, After specifying the detection conditions, the user can acquire the sea surface temperature for the area that can be obtained without doing anything. Conversely, if there is a part where sea surface temperature can not be acquired, a teacher set is prepared to perform learning on the powerful part, the user learns, and the user specifies the detection condition, and the same as above, The sea surface temperature can be obtained for the required part. Here, by repeating such a series of operations, the sea surface temperature may be obtained for each of the overlapping portions of one weight-the Euler network and the other weights-the Euler network. In this case, it is possible not to determine the sea surface temperature for the part already obtained, but also to apply the filter matrix or the weighted average filter matrix used in the image processing field to obtain the sea. The distribution map of surface temperature can be made into a smooth image. And from the distribution map of the sea surface temperature, it is also possible to obtain the value of the surface temperature in reverse.

 If the detection condition of the similar area detection means is relaxed, the error will be large and the accuracy will be low. If the detection condition is strict, the error will be small and the accuracy will be high. That is, there is a trade-off between the detection condition and the accuracy.

 The detection conditions described above in the similar area detection means are similar to the search conditions specified in the attribute search or the spatial search, and are known techniques in the geographic information system, Although a detailed description is omitted, various designations can be made for the user to make as will be apparent to those skilled in the art. Also, image processing with a filter matrix and a weighted average filter matrix is a well-known technique, and although detailed description is omitted, various filtering processes can be applied as will be apparent to those skilled in the art.

 [Detection target]

 In the present embodiment, since the detection condition is the same area or not, the detection condition is limited to the use only in Japan, and the distance from the land is the same, and the detection condition is that the currents are similar. If so, similar region detection can be extended to the whole world.

 [Application to other than SST]

Further, in the present embodiment, if the force-like area limitedly described with respect to SST is limited to SST, spatial data to be detected is recorded in the database 10 if detection conditions are established. Detecting similar regions can be done easily [Not Learned !, Specifying a Position or an Area]

 Further, in the present embodiment, although the explanation has been made that the similar area detection is performed on the already learned area, the similar area detection can also be performed on the non-learned area. In the case where the total area of the similar area is large, if the force is large, the estimation system can be efficiently constructed by performing learning V in order.

 Further, in the present embodiment, the similar area is detected by the similar area detection means, but when learning is performed, when a certain position or area is designated, the position or area to be touched and the position already learned are already detected. Or, it is possible to judge whether the area and the force match the detection condition, and it is possible to apply the weight of the already learned position or area only to the position or area needed by the user. An estimate of the desired location or area without learning can be performed.

[Detected when CPU usage is low]

 In the present embodiment, similar area detection can also be performed in the background when the CPU usage rate is low. When the target area for similar area detection is wide, the amount of processing for detection becomes excessive, so that the similar area can be detected smoothly by appropriately performing the process in the background. Generally, it is preferable to perform similar area detection when the CPU utilization is low because the operation of the geographic information system is high load.

 [Detection when changes in time series are equivalent]

 Further, in the present embodiment, the SST case is specifically described as the detection condition, but regions having the same change in time series can often be regarded as similar regions. Even in the case of SST, if the observation data or teacher data to be input data has the same change over time, it can be detected as a similar area.

[Support for Specification of Detection Condition]

Although it has been stated that the user sets a settable condition detection condition in the spatial data, what setting can be made for what attribute data or map data can be set in the database 10. It is necessary to know if the data is recorded, and it is difficult to set it up unless you are an expert. Therefore, data on the position or area specified by the user It is possible to selectably present what kind of data is recorded in the base 10, and to be able to input or select conditions for the selected one. For example, a text box or a pull-down menu can be configured as a window in which a text box or a pull-down menu appears by designating an attribute of data, placing a check box next to each attribute name, and activating the check box. Dialog Parts may be used.) ₀ ) If so, what kind of data is recorded in the database 10 at the user or the position or area to be analyzed of the beginner who has started to use this system. Even a user who does not know the presence can easily set the detection conditions and can perform similar area detection smoothly.

 [Confirmation process of similar area]

 When the user sets the detection conditions, the user's excellent insight is often required. Therefore, it is necessary to make sure that the detection conditions set by the user are appropriate and whether they are correct as similar areas. As this confirmation method, when the similar area is a plurality of closed areas, the user is asked for a teacher set at a certain position on each similar area, and inference is made using the input data of the teacher set, If the difference between the output data and the teacher data of the teacher set is within a predetermined threshold, it can be regarded as a similar area. Conversely, if it is outside the threshold, it is not regarded as a similar area, and the user is informed that the current detection condition is not appropriate. When the similar area is one large closed area, for example, the similar area is applied to the grid to obtain the grid point or the grid center as a sample point, and a teacher set of such sample points is obtained from the user and input. It can be confirmed in the same manner as above based on the teacher set. An appropriate similar area can be detected by confirming with such small sample points and judging whether it is a similar area or not.

 Although the present invention has been described by the above respective embodiments, the technical scope of the present invention is not limited to the scope described in the embodiments, and various changes or improvements may be added to the respective embodiments. Is possible. Also, embodiments to which various changes or improvements are added are also included in the technical scope of the present invention. This is also apparent from the scope of claims and the means for solving the problems.

 Example

[Temporal-Spatial Adaptation by Geographic Information System (NN—GIS) with Neural Network Function Automatic Estimation of Response Sea Surface Temperature Estimation Parameters]

 The coefficients of the sea surface temperature estimation equation optimal for sea surface temperature estimation differ depending on the sea area and the season, and accurate estimation using the sea surface temperature estimation equation compatible with all sea areas and seasons is impossible. On the other hand, the geographic information system is capable of displaying superimposed maps with topographic maps, thematic maps, etc. associated with a plurality of position information, and has the feature that spatial search, quantitative analysis and simulation are possible. . In particular, by connecting each layer of the geographic information system with a weighting factor, this can be regarded as a hierarchical-eural network. This is called neural network geographic information system (NN-GIS). Due to this, input necessary for estimating the sea surface temperature of AVHRR band 4, 5 etc. is set in the input layer of the geographic information system, and the truth data as the correct value in the output layer (Truss data obtained in each sea area By setting buoy data, etc.), it is possible to determine the optimal weight coefficient for each sea area, and use it to estimate sea surface temperature for all data in the surrounding sea area. In the coastal area where the time change of sea surface temperature is steep, if the time change is slow for each time zone, or if the time change is slow like the sea surface temperature of the open ocean, then a season-by-season estimation formula can be obtained. The sea surface temperature estimation formula can be generated automatically. The specific usage is as follows.

 NN 1 Set AVHRR data in the input layer of GIS and display it. Set V, ト truth data obtained in the relevant sea area to the output layer. Train the NN-GIS weighting factor from the above data. Estimate the sea surface temperature of the surrounding sea area including the location of the truth data. Change the sea area and change the time zone and season of the acquisition AVHRR and repeat the above procedure. This will enable us to estimate the sea surface temperature that is optimal for all time and area.

 [Experiment]

 In order to confirm the function of NN-GIS, the data on May 20, 2005 at 17:11 and the data on November 30, 2004 at 16:45 were used. Actually, the main or module of the geographic information system calls modules other than the main part.

The data and output results for May 20, 2002 are as shown in Figures 13-15. The data for May 20 takes the image of channel 4 (Fig. 13) and channel 5 (Fig. 14), and 16 * 16 pixels from (289, 126) off Ogura of MCSST (Fig. 15) as the correct value. ,Less than Calculated based on the teacher set (Table 2). The squared error of its output is shown in (Figure 16) [Table 2]

(Teacher set based on data from Ogura off May 20, 2002)

 Number of Layers INPUT = 2 HIDDEN = 2 OUTPUT = l PATTERN = 25 eta = 2.4

 alpha = 0.8 beta = 0.8 where eta and alpha are coefficients for updating the steepest descent weighting coefficient, respectively

 [Equation 16]

d _w ί (i * d

Find the distance (variation) in the solution space to advance to the next solution by multiplying the variation of the current weighting factor and the variation of the previous weighting factor by eta and alpha There is. Further, beta is a non-linear function (sigmoid) in the output layer, and is defined by 1.0 = (1.0 + _eX p (-bet _a * u).

[0087] In FIG. 16, there are sharp increases and decreases up to 2.8 by 100,000 times. The constant state of about 0.01 continues up to 200,000 times, and then gradually decreases. It converges sufficiently by 700,000 times.

Also, take 16 * 16 pixels from (375, 325) off Amagasaki, and calculate based on the following teacher set (Table 5). The squared error of the output is shown in FIG. [Table 3]

 (Teacher set based on the data off Miyazaki, May 20, 2002)

 [0089] In FIG. 17, the convergence is sufficiently fast in about 30,000 times as compared with other data. There is no further rise in the error after that and it is very stable.

On the other hand, first, the data and output results on November 30, 2004 are as shown in Figure 18 to Figure 20. Ru. The image of channel 4 (Fig. 18) and channel 5 (Fig. 19), and the teacher of (Table 4) taking 16 * 16 pixels from (235, 173) off Ogura in MCSST (Fig. 20) respectively. Calculated based on the set, the squared error of the output is shown in (Figure 21).

[Table 4]

[0091] INPUT = 2 HIDDEN = 2 OUTPUT = 1 PATTERN = 22 eta = 2.4 alpha = 0.8 beta = 0.8 In Fig. 21, the error slightly decreases for a while and then rises rapidly to 0.1 and then gradually rises up to 0.015 at the maximum. Do. After that, it decreases smoothly to 400,000 times and converges sufficiently.

 In addition, taking 16 * 16 pixels from (335, 380) off Kashiwazaki, respectively, the teacher set in Table 5 is calculated, and the squared error of the output is shown in FIG.

[Table 5]

(Four

OC n too too

0TC800 / S00idf / X3d 63 1 ^ _0Ζΐ / 900Ζ O In Fig. 22, the error converges rapidly after the maximum error of 0.47. However, with this data, sufficient convergence takes the most time.

[Consideration]

 The results of the experiment in the previous chapter show that the transition of convergence of the set of teachers and the error that are extracted differ depending on each sea area and each period. Also, in any of the experimental results, the square error is sufficiently reduced by the repetition of learning. Therefore, it can be understood that the optimal SST estimation formula can be automatically and easily determined by NN-GIS in each sea area and each period.

 Teacher data in learning can be measured on site by installing a satellite data receiver on a fishing boat. Based on the acquired teacher data and satellite observation data, it is possible to easily obtain the estimation formula that is optimal for the relevant area and the relevant period.

 Geographic information such as climatic conditions and geographical conditions for each sea area can be added to the input data, and learning data of the eural network can be efficiently used for each similar sea area. Even in areas where observation data are lacking, it is possible to simulate estimated values using observation data from other areas.

This time, the power that showed the effect of NN-GIS, taking the sea surface temperature as an example of physical quantity, can derive the optimum physical quantity estimation formula similarly with other vegetation and physical quantities such as chemical substance content in the atmosphere. Application to earth observation such as meteorological observation and environmental problems is possible.

 23 to 29 show program lists used in this embodiment.

Claims

The scope of the claims

 [1] A recording unit for recording spatial data, which is data having information on position, and analysis means for analyzing the spatial data of the recording unit,

 In the analysis means, the input data of the teacher set is read into the input layer, and forward calculation is performed, and it is constructed for each position by learning backward with the output data of the output layer of the calculation result and the teacher data of the teacher set. -Read input data into the input layer of the eural network and obtain estimation results of the read input data.-Geographical information system using the eural network.

 [2] A space-time search means for performing space-time search is provided,

 Estimate the time of a certain position using the same time weight of the same position before last year

 A geographic information system using a eural network according to claim 1.

[3] A similar area detection means for detecting an area meeting the detection condition specified by the user as a similar area is provided, and the recording section records data targeted for the detection condition as spatial data.

 If a part of the similar area has already been learned, it is already learned and has already been constructed-estimation is performed on other parts in the similar area using the Euler network. -Geographic information system using eural network.

[4] The detection condition for detecting the similar area can be adjusted

 A geographic information system using a eural network according to claim 3.

[5] Detect each at least one sample point of each similar region found, receive teacher set input for sample point, and have already been constructed for input teacher set input data-Ural Make an inference using a network, compare the output data with the teacher data of the teacher set, and decide whether or not to consider as a similar area depending on whether the error is within the threshold or not.

 A geographic information system using a eural network according to claim 3.

[6] A storage unit for recording, as spatial data, a far-infrared image corresponding to a position obtained by remote sensing by a satellite, and an analysis means for analyzing the spatial data of the storage unit Equipped with

 In the analysis means, a far infrared image of the position at the input layer is given, and forward calculation is performed, and learning is performed backward by learning from the sea surface temperature of the output layer of the calculation result and the sea surface temperature serving as teacher data. To be built in-to input the far-infrared image to the input layer of the eural network, and find the sea surface temperature of the input far-infrared image-the geographic information system using the eural network.

[7] The input far-infrared image is two far-infrared images at the same position in different frequency bands, and the teacher data is the measured data of the sea surface temperature at the same position as the far-infrared image or satellite observation data

 A geographic information system using a eural network according to claim 6.

[8] When there is a position where the user specified position and the detection condition match, the sea surface temperature is obtained using the already-learned position-Eural network estimation.

 A geographic information system using a eural network according to claim 6.

[9] A method applied to a geographic information system for analyzing spatial data of a recording unit for recording spatial data, which is data having information on position,

 Reading the input data of the teacher set to the input layer, performing a forward operation, and performing the backward operation and learning with the output data of the output layer, which is the calculation result of the forward operation, and the teacher data of the teacher set Reading the input data into the post-process input layer to be learned, and determining an estimation result.

[10] A geographic information program in which a processor analyzes spatial data of a recording unit that records spatial data, which is data having information on position,

 The processor reads the input data of the teacher set in the input layer and performs forward operation, and the processor returns the output data of the output layer which is the operation result of the forward operation and the teacher data of the teacher set. A geographic information program that executes a procedure for computing and learning, and a procedure for reading the input data to the input layer after the procedure.