WO1993019426A1 - Method of detecting and mapping minerals and other geological phenomena utilizing airborne imaging spectrometers - Google Patents
Method of detecting and mapping minerals and other geological phenomena utilizing airborne imaging spectrometers Download PDFInfo
- Publication number
- WO1993019426A1 WO1993019426A1 PCT/US1992/000593 US9200593W WO9319426A1 WO 1993019426 A1 WO1993019426 A1 WO 1993019426A1 US 9200593 W US9200593 W US 9200593W WO 9319426 A1 WO9319426 A1 WO 9319426A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- neural network
- data
- functions
- layer
- difference
- Prior art date
Links
- 229910052500 inorganic mineral Inorganic materials 0.000 title claims abstract description 41
- 239000011707 mineral Substances 0.000 title claims abstract description 41
- 238000000034 method Methods 0.000 title claims abstract description 39
- 238000013507 mapping Methods 0.000 title claims abstract description 13
- 238000003384 imaging method Methods 0.000 title abstract description 34
- 238000013528 artificial neural network Methods 0.000 claims abstract description 75
- 238000012545 processing Methods 0.000 claims abstract description 11
- 238000001228 spectrum Methods 0.000 claims description 53
- 230000006870 function Effects 0.000 claims description 37
- 239000000463 material Substances 0.000 claims description 23
- 230000008569 process Effects 0.000 claims description 17
- 230000001537 neural effect Effects 0.000 claims description 10
- 238000002310 reflectometry Methods 0.000 claims description 10
- 230000005670 electromagnetic radiation Effects 0.000 claims description 6
- 238000012549 training Methods 0.000 description 41
- 238000010521 absorption reaction Methods 0.000 description 21
- 230000003595 spectral effect Effects 0.000 description 20
- 239000013598 vector Substances 0.000 description 18
- VYPSYNLAJGMNEJ-UHFFFAOYSA-N Silicium dioxide Chemical compound O=[Si]=O VYPSYNLAJGMNEJ-UHFFFAOYSA-N 0.000 description 14
- 238000004422 calculation algorithm Methods 0.000 description 12
- 210000002569 neuron Anatomy 0.000 description 8
- 229910052934 alunite Inorganic materials 0.000 description 7
- 239000010424 alunite Substances 0.000 description 7
- BERDEBHAJNAUOM-UHFFFAOYSA-N copper(I) oxide Inorganic materials [Cu]O[Cu] BERDEBHAJNAUOM-UHFFFAOYSA-N 0.000 description 7
- LBJNMUFDOHXDFG-UHFFFAOYSA-N copper;hydrate Chemical compound O.[Cu].[Cu] LBJNMUFDOHXDFG-UHFFFAOYSA-N 0.000 description 7
- 239000000377 silicon dioxide Substances 0.000 description 7
- 230000004913 activation Effects 0.000 description 6
- 238000005259 measurement Methods 0.000 description 6
- KPZTWMNLAFDTGF-UHFFFAOYSA-D trialuminum;potassium;hexahydroxide;disulfate Chemical compound [OH-].[OH-].[OH-].[OH-].[OH-].[OH-].[Al+3].[Al+3].[Al+3].[K+].[O-]S([O-])(=O)=O.[O-]S([O-])(=O)=O KPZTWMNLAFDTGF-UHFFFAOYSA-D 0.000 description 6
- 229910021536 Zeolite Inorganic materials 0.000 description 5
- HNPSIPDUKPIQMN-UHFFFAOYSA-N dioxosilane;oxo(oxoalumanyloxy)alumane Chemical compound O=[Si]=O.O=[Al]O[Al]=O HNPSIPDUKPIQMN-UHFFFAOYSA-N 0.000 description 5
- 230000000694 effects Effects 0.000 description 5
- 239000010457 zeolite Substances 0.000 description 5
- 238000000701 chemical imaging Methods 0.000 description 4
- 239000000523 sample Substances 0.000 description 4
- 230000004075 alteration Effects 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 238000003331 infrared imaging Methods 0.000 description 3
- 238000005065 mining Methods 0.000 description 3
- 238000000985 reflectance spectrum Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- 229910001595 Buddingtonite Inorganic materials 0.000 description 2
- CURLTUGMZLYLDI-UHFFFAOYSA-N Carbon dioxide Chemical compound O=C=O CURLTUGMZLYLDI-UHFFFAOYSA-N 0.000 description 2
- XEEYBQQBJWHFJM-UHFFFAOYSA-N Iron Chemical compound [Fe] XEEYBQQBJWHFJM-UHFFFAOYSA-N 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 230000033558 biomineral tissue development Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000007613 environmental effect Effects 0.000 description 2
- 229910052595 hematite Inorganic materials 0.000 description 2
- 239000011019 hematite Substances 0.000 description 2
- LIKBJVNGSGBSGK-UHFFFAOYSA-N iron(3+);oxygen(2-) Chemical compound [O-2].[O-2].[O-2].[Fe+3].[Fe+3] LIKBJVNGSGBSGK-UHFFFAOYSA-N 0.000 description 2
- NLYAJNPCOHFWQQ-UHFFFAOYSA-N kaolin Chemical compound O.O.O=[Al]O[Si](=O)O[Si](=O)O[Al]=O NLYAJNPCOHFWQQ-UHFFFAOYSA-N 0.000 description 2
- 229910052622 kaolinite Inorganic materials 0.000 description 2
- 230000000644 propagated effect Effects 0.000 description 2
- ZAKOWWREFLAJOT-CEFNRUSXSA-N D-alpha-tocopherylacetate Chemical compound CC(=O)OC1=C(C)C(C)=C2O[C@@](CCC[C@H](C)CCC[C@H](C)CCCC(C)C)(C)CCC2=C1C ZAKOWWREFLAJOT-CEFNRUSXSA-N 0.000 description 1
- 244000007853 Sarothamnus scoparius Species 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 230000003542 behavioural effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 229910002092 carbon dioxide Inorganic materials 0.000 description 1
- 239000001569 carbon dioxide Substances 0.000 description 1
- 230000019771 cognition Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 229910052742 iron Inorganic materials 0.000 description 1
- 238000011545 laboratory measurement Methods 0.000 description 1
- 239000006101 laboratory sample Substances 0.000 description 1
- VUZPPFZMUPKLLV-UHFFFAOYSA-N methane;hydrate Chemical compound C.O VUZPPFZMUPKLLV-UHFFFAOYSA-N 0.000 description 1
- 230000001089 mineralizing effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- -1 playa Chemical compound 0.000 description 1
- 230000010287 polarization Effects 0.000 description 1
- 239000011435 rock Substances 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V8/00—Prospecting or detecting by optical means
- G01V8/02—Prospecting
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01J—MEASUREMENT OF INTENSITY, VELOCITY, SPECTRAL CONTENT, POLARISATION, PHASE OR PULSE CHARACTERISTICS OF INFRARED, VISIBLE OR ULTRAVIOLET LIGHT; COLORIMETRY; RADIATION PYROMETRY
- G01J3/00—Spectrometry; Spectrophotometry; Monochromators; Measuring colours
- G01J3/28—Investigating the spectrum
- G01J3/2823—Imaging spectrometer
Definitions
- This invention relates to a method of detecting and mapping minerals and other geological phenomena utilizing data gathered by airborne imaging spectrometers.
- information is collected via an airplane or satellite from an identified area to be mapped or analyzed.
- the information is in the form of solar electromagnetic energy reflected and radiated from the earth's surface in the identified area.
- the reflected electromagnetic energy is dispersed in a spectrometer and used to form congruous registered spectral images of the area.
- the reflected and radiated electromagnetic spectrum is measured at a number of discrete wavelengths or bands by the imaging spectrometers.
- the spectra are measured at many spatial locations which are termed pixels. Not all wavelengths are reflected with equal intensity. Some wavelengths are preferentially absorbed, as shown in FIG. 1, wherein line 30 is by a United States
- line 32 is by an Airborne Visible Infrared Imaging Spectrometer (AVIRIS)
- line 34 is by a Geophysical Environmental Research Imaging Spectrometer (GERIS)
- line 36 is by an Airborne Multispectral Scanner
- AMSII Landsat Thematic Mappper
- LandsatTM Landsat Thematic Mappper
- the wavelength at which the absorption feature is a minima, the full width at half maxima (FWHM) 42, and the intensity and shape or symmetry of the feature can be diagnostic of different minerals which reflect the incident electromagnetic radiation, as shown in FIG. 2.
- FWHM full width at half maxima
- FIG. 2 there is also shown a continuum 44 which may be removed to normalize the absorption data at intensity 46.
- the instant invention is concerned pri arily with using absorption features in the range of electromagnetic wavelengths 0.5 to 2.5 microns as shown in FIG. 3. This is termed the visible and near infrared region. It is believed that the first images from the prototype Airborne Imaging Spectrometer (AIS) were acquired in 1983. See Vane, 1985, High Spectral Resolution Remote Sensing of the Earth, Sensors, Vol. 2, pp. 11-22.
- AIS Airborne Imaging Spectrometer
- the spectrometer utilized a small, two dimensional chip with more than a thousand discrete detector elements.
- a scanning mirror scans the detector across the surface of the earth perpendicular to the flight path. Each detector converts the reflected electromagnetic energy into an electrical signal.
- Line areas of detectors rather than discrete detector elements have also been used to increase diagnostic detection.
- Scanning, or whisk broom imaging spectroscopy as it has been called, has also been used with a line array of detectors. See Vane and Goetz, 1988, Terrestrial Imaging Spectroscopy: Remote Sensing of, Environment, Vol. 24, n. 1, pp. 1-29.
- the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) has been built on this principle and records 224 bands between the wavelengths of 0.41 to 2.45 u . Also, area arrays as shown in FIG. 4 have been used.
- spectral characteristic identification algorithms include binary encoding and spectral ratioing for minerals with spectra lacking sharp absorption features.
- Binary encoding is fast, sensitive to the position of the absorption peaks, insensitive to albedo effects and obviously works poorly for minerals that do not possess well defined absorption peaks.
- the spectrum ratioing is sensitive to albedo effects and the position and strength of absorption peaks.
- the instant invention can be used with any of the above imaging spectrometer systems, although for purposes of explanation the GERIS data was used.
- the instant invention uses trained neural networks to aid in the detection of mapping of minerals.
- a neural net consists of a set of weights or numbers with a defined connectivity. The weights or numbers are derived by a process called training (for the purpose of this invention) .
- training for the purpose of this invention
- a neural net is trained in particular to recognize certain patterns in the reflected electromagnetic . spectrum (0.5 - 2.5 um) .
- These training patterns can be derived from:
- Airborne imaging spectrometer data itself with or without a continuum removed.
- Field or laboratory spectrometer measurements of material that are desired to be mapped, but which are not from the area to be mapped e.g. if a certain alteration mineral assemblage has been noted in a locale that is associated with . a desired mineralization, then spectrometer measurements of this mineral assemblage can be used to map minerals in airborne imaging spectrometer data from another locale.
- the process of mineral identification uses two stages.
- the first is the training stage.
- Training of a neural net requires a training set and a training algorithm or paradigm.
- a training set consists of many training pairs.
- a training pair consists of an input vector, which in this case is spectrum reflectivity or absorption data or part thereof and a desired output vector.
- the form of the desired output vector is dependant upon the digital representation arbitrarily designated by the user for the particular mineral the neural net is being trained to recognize.
- the spectra used as input vectors in the training set can either be laboratory spectra, convolved to the resolution of the imaging spectrometer, spectra from the imaging spectrometer survey itself, or from in-the-field ground spectrometer measurements, again convolved to the resolution of the imaging spectrometer used in the mapping.
- a linear or nonlinear continuum can also be removed from the spectrum before training.
- a neural network can be trained to recognize one mineral or several minerals.
- a neural network can also be trained using one or several absorption features in a spectrum which are diagnostic of one mineral.
- Training consists of a forward pass through the neural network, resulting in an error or delta value being calculated based on the difference between the neural network calculated output vector and the desired output vector. This error is then back propagated through the neural network and used to adjust the weights of the neural network according to the generalized delta rule. Training is considered complete when the error values for each training pair (the difference between the neural network calculated output and the desired output vector) are considered to be a minimum. Training results in a trained neural network which consist of a set or net of weights or numbers, which are connected by a neural network topology.
- the second stage of the process of mineral identification is to apply the trained neural network to imaging spectrometer data.
- the imaging spectrometer data is calibrated to reflectance values for the cases in which laboratory or in-the-field measured ground reflectance spectra are used for training. See Clark, et al., 1988, Calibration and
- the trained neural net is then used to process the airborne imaging spectrometer data or its calibrated and/or continuum removed equivalent.
- the output of the neural network indicates a vector (or scalar if only one number is output from the network, depending on user choice of the desired output vector) that is value dependant on the presence of the particular designated mineral to be mapped.
- each pixel in the data set is given a scalar number or vector indicating the presence of the mineral or the absence of the mineral so that detection or mapping can be achieved.
- a calculated output vector similar to a designated output vector for the mineral to be mapped indicates the presence of that mineral.
- FIG. 1 is the reflected electromagnetic spectrum for the mineral alunite, as sampled by different imaging spectrometer systems.
- FIG. 2 is an idealized and hypothetical depiction of geometrical parameters which describe an absorption feature.
- FIG. 3 is the electromagnetic spectrum between 1 and IO 20 Hz showing the different nomenclature for windows of the spectrum and physical phenomena that give rise to electromagnetic radiation in the respective windows.
- FIG. 4 is a schematic perspective view of a type of imaging spectrometer, not drawn to scale.
- FIG. 5 is a block diagram of an artificial neuron or node used in a neural network.
- FIG. 6 is a diagram showing output layer weight adjustment, in a 3 layer, multilayer backpropagation neural network during training.
- FIG. 7 shows a flow chart of a method of invention in which airborne imaging spectrometer data is used for training a neural network.
- FIG. 8 shows a flow chart of a method of invention in which field or laboratory measured spectra are used to train a neural network.
- FIG. 9 is an alteration map for the Cuprite mining district in Nevada.
- FIG. 10 is a library of spectra derived from the image data of the area mapped in FIG. 9.
- FIG. 11 is a map of the extent of silica as derived from an inversion recognition program.
- FIG. 12 is a map of the extent of alunite as derived from an inversion recognition program.
- FIG. 13 is a map of the extent of zeolite as derived from an inversion recognition program.
- FIG. 14 is the offset value spectrum for
- FIG. 15 is the gain factor spectrum for Cuprite GERIS data with standard errors superimposed.
- FIG, 16 is the image derived from a trained neural network to recognize silica using all spectral bands except for 24-27.
- FIG. 17 is the image derived from the trained neural network to recognize silica using the last 32 spectral bands.
- FIG. 18 is the image derived from the trained neural network to recognize alunite using the last 32 spectral bands.
- FIG. 19 is the image derived from the trained neural network to recognize zeolite using the last 32 spectral bands.
- a back propagation algorithm was used to compare the calibrated imaging spectrometer data to laboratory generated spectra of the same resolution. It is understood that the data could also be compared to actual ground measurements or that other algorithms (or paradigms) or variants of back propagation could be used to train the neural network.
- the training of a neural network of weights was achieved using a back propagation algorithm although other training algorithms could be used.
- a neural network can be made of layers of neuron, or nodes, each performing the below function as shown in FIG. 6. 1 The first layer of input nodes are referred to herein as "layer 1" or the "input layer”.
- Neural nets herein are constructed in such a way as that every node in the immediately higher layer (the input layer is the highest) is connected to every node in the next lower layer, but not to any other layer. Nodes are connected to other nodes via weights. This is not the only type of neural network that can be used, as input nodes can be connected directly to the output nodes in a three layer neural network for example.
- the connectivity is generally termed topology. Other modifications of network topologies are also possible.
- the algorithm has an artificial neuron or model as its most fundamental identity.
- a neuron or node is shown in FIG. 5.
- the activation function, F should have the following properties: * It is everywhere differentiable
- Training requires a matched set of inputs and desired outputs.
- the input values of calibrated data can be normalized such that their range lies between 0 and 1 (for example), and are fed into the network through the input nodes. This normalization can optimize the performance of the training. Assuming a three (input, middle and output layers) layer network (FIG. 6), then for each middle layer node, each input node value is multiplied by a respective weight value
- NET - sum of the products of node values and weight values for the layer of nodes immediately above. This is the net variable for node p in layer j.
- the error(s) are then back propagated up through the neural net.
- the error value is multiplied by the derivative of the activation function using the NET value of the particular output node as the argument for the function, to obtain a delta weight value.
- the delta weight value is then multiplied by a learning coefficient, LC, (which is user defined and which can be different for different layers and for different nodes within layers if so desired) and then by output value of an immediately above connecting node.
- the weight to be changed is only that weight that connects the particular hidden layer node to the particular output layer node, and it should be noted that this weight is termed the output weight.
- the product is the change that is to be added to the pre-existing weight value. This process is then performed for weights immediately above the output layer, thus modifying all weights between the output layer and the hidden layer for this example.
- W k (n+l) adjusted weight for the above connecting weight
- D k delta value associated with the output node q in layer k (output layer);
- DW pqk delta weight value or change in weight from node p in the hidden layer to node q in the output layer k;
- the method of derivation of the delta value for a weight in a hidden layer involves the summation of the delta values for each connected output node.
- the delta value is the product of the error for each output node and the derivative of the activation function for each output node again using the NET of the output node as the argument in the function.
- the delta value for each output node is then multiplied by the updated output layer weight and these products are summed.
- the sum is then multiplied by the derivative of the activation for the hidden layer node, using the NET value of the hidden layer node as the argument to the function.
- This delta value is, then multiplied by the learning coefficient and the output of the connecting node in the above layer. In the case of a three layer network, this output would in fact be the normalized input value to the network.
- D k delta calculated for the connecting output node, Q, in layer k
- This forward and backward pass through the network continues, modifying weights until an acceptable error (difference between calculated output vector and desired output vector) is obtained.
- the acceptable error level is user-defined.
- This sequence of iterations or presentations is termed “training” or “learning”.
- a common error measure is the mean square error (MSE) which is the mean of the errors for each complete pass through a training set.
- MSE mean square error
- checks can be made on whether it has been effective. For example, the input only of each training set can be presented to the network, and the output then calculated. The calculated output values should show excellent correspondence to the desired output values.
- Other checks involve addition of random noise (different levels thereof) to input training sets prior to presentation to the trained network.
- Other checks can be made by presenting features that are considered to be important to the problem, to the network, and to assessing the network output.
- FIG. 4 illustrates an imaging spectrometer 2 as it scans the area 3 to be analyzed or mapped.
- Objective 4 directs the reflected electromagnetic radiation through slit 5 where it is again directed through collimator 6 to dispersing element 7 (typically a grating) .
- Lens 8 directs the light onto a detector array 9. Not all wavelengths are reflected with equal intensity. Some wavelengths are preferentially absorbed, creating absorption features which are defined in a close enough sampled spectra.
- the instant invention primarily uses absorption features in the range of wavelengths 0.5 to 2.5 microns. This is termed the visible and near-infrared region.
- a date set is formed from the signals received by the detector array 9.
- each pixel (or picture element) on the ground has a sample of the electromagnetic spectra consisting of digital numbers. These numbers are then converted to ground reflectance values (the process called calibration) .
- a continuum is then removed. There are several ways to define continuum, including (1) using a segmented upper hull which may be adequate for high signal and low noise conditions, which are not that common in the currently used instruments, and (2) using a least squares best fit linear continuum for each different absorption feature.
- the calibrated data or calibrated continuum part of the spectrum is then compared to a trained neural network of weights to give a positive value as to a designated mineral for each pixel in the area to be mapped.
- the absorption feature or combination of absorption features which is diagnostic of a particular mineral is used.
- signals from the detector of the imaging spectrometer 11 are converted into calibrated surface reflectance in the particular sample given. This is called the calibrated data 12.
- the spectra has its continuum removed, which is essentially taking out a background or long wavelength from the spectra, leaving a spectra of more distinctive absorption peaks.
- the calibrated data 12 is compared to a neural network of weights 13 derived using selected spectra from the imaging spectrometer data set.
- the output 14 indicates whether a particular mineral to be studied is present in a given pixel of the area to be mapped.
- Mineral mapping 15 can be achieved from the output 14 and the extent of the mineral can be measured.
- the training set is derived from either field and/or laboratory spectrometer measurements from the area 21 to be mapped.
- the field or laboratory samples can be from an area other than the area to be mapped.
- area 21 will typically contain a number of discrete sample locations to produce data in the form of raw spectra.
- the raw spectra data is processed to convolve to the spectral resolution and to remove any unwanted continuum and is then processed by the neural network.
- the processed data can then be applied to imaging spectrometer data 23 which is first calibrated to remove reflectance 24, and a continuum may be removed 25.
- the output 26 from the calibrated data with the continuum removed 25 is then processed by the neural network to produce a neural network output 26 which can be mapped 27 to show mineralization.
- the following example illustrated in FIG. 9-19 utilizes the neural network software to process airborne GERIS data of the Cuprite mining district of Nevada.
- a summary geology map of the Cuprite mining district of Nevada is shown in FIG. 9.
- the spectra used for training were derived from the imaging spectrometer data set over an area of known geology and ground spectra control. No attempt was made to isolate or analyze individual absorption features in this example. Eight spectra were used and nets were designed to recogni ⁇ e one of the eight spectra. The spectra used were those of basalt, alunite, buddingtonite, kaolinite, playa, silica, hematite and zeolite as shown in FIG. 10.
- the output from the trained neural nets can be compared with the output from a spectral recognition program based on an inversion algorithm as shown in Boardman, 1989, Inversion of Imaging Spectrometry using Singular Value Decomposition, IGARSS Proc. 1989,
- FIG. 11 is the map of the extent of silica as derived from the inversion recognition program.
- FIG. 12 is a map of similarly derived alunite and FIG.
- FIGS. 14 and 15 show the standard errors for each of the bands.
- FIG. 14 shows the offset value spectrum for Cuprite GERIS data with standard errors superimposed. Bands 24-27 have been deleted due to high error.
- FIG. 15 shows the gain factor spectrum for Cuprite GERIS data with standard errors superimposed. Bands 24-27 have been deleted due to high error. It is obvious for each gain and offset factor, the most errors occur for bands with wavelengths greater than 1.8 micrometers. Unfortunately, this is the part of the spectra which contains many of the diagnostic absorption peaks.
- FIG. 11 is similar to FIG. 17 where only the last 32 spectral bands were used. Similarities can also be seen between FIG. 12 and FIG. 18 for alunite and FIG. 13 and FIG. 19 for zeolite. In FIGS. 16-19 a linear stretch has been applied to the output recognition elements.
- neural network means any system that is
- the "trained” to manipulate a set of input data indicative of the reflectivity or absorption characteristics at one or more electromagnetic spectra using a plurality of functions (which preferably are “weights” or sealer functions) to produce an output indicative of a mineral or other material.
- the “training” is accomplished by assigning a value to a material to be detected, inputting data to the neural network (from laboratory or other known sources) that corresponds to such material, manipulating the data to obtain an output, and then adjusting the manipulating functions based on the difference between the assigned value and the actual obtained value.
- the training process may be performed repetitively to increase the precision of the neural network.
- the neural network may include multiple layers of functions.
Abstract
A method and apparatus for detecting and mapping minerals and other geological phenomena (21) using data gathered by imaging spectrometer (23) and processing the data with a neural network (26). The neural network (26) is trained to recognize minerals by processing the data from known minerals and adjusting a set of interconnected weights in the neural network (26) to obtain a predetermined output value (27).
Description
METHOD OF DETECTING AND MAPPING MINERALS
AND OTHER GEOLOGICAL PHENOMENA UTILIZING
AIRBORNE IMAGING SPECTROMETERS
Background
This invention relates to a method of detecting and mapping minerals and other geological phenomena utilizing data gathered by airborne imaging spectrometers.
Typically in airborne imaging spectrometry, information is collected via an airplane or satellite from an identified area to be mapped or analyzed. The information is in the form of solar electromagnetic energy reflected and radiated from the earth's surface in the identified area. The reflected electromagnetic energy is dispersed in a spectrometer and used to form congruous registered spectral images of the area. The reflected and radiated electromagnetic spectrum is measured at a number of discrete wavelengths or bands by the imaging spectrometers. The spectra are measured at many spatial locations which are termed pixels. Not all wavelengths are reflected with equal intensity. Some wavelengths are preferentially absorbed, as shown in FIG. 1, wherein line 30 is by a United States
Geological Survey laboratory spectrum (USGS Spec), line 32 is by an Airborne Visible Infrared Imaging Spectrometer (AVIRIS), line 34 is by a Geophysical Environmental Research Imaging Spectrometer (GERIS), line 36 is by an Airborne Multispectral Scanner
(AMSII), and line 38 is by a Landsat Thematic Mappper (Landsat™) . The data shows absorption features 40.
The wavelength at which the absorption feature is a minima, the full width at half maxima (FWHM) 42, and the intensity and shape or symmetry of the feature, can be diagnostic of different minerals which reflect the incident electromagnetic radiation, as shown in FIG. 2. In FIG. 2 there is also shown a continuum 44 which may be removed to normalize the absorption data at intensity 46. The instant invention is concerned
pri arily with using absorption features in the range of electromagnetic wavelengths 0.5 to 2.5 microns as shown in FIG. 3. This is termed the visible and near infrared region. It is believed that the first images from the prototype Airborne Imaging Spectrometer (AIS) were acquired in 1983. See Vane, 1985, High Spectral Resolution Remote Sensing of the Earth, Sensors, Vol. 2, pp. 11-22. The spectrometer utilized a small, two dimensional chip with more than a thousand discrete detector elements. A scanning mirror scans the detector across the surface of the earth perpendicular to the flight path. Each detector converts the reflected electromagnetic energy into an electrical signal. Line areas of detectors rather than discrete detector elements have also been used to increase diagnostic detection. Scanning, or whisk broom imaging spectroscopy as it has been called, has also been used with a line array of detectors. See Vane and Goetz, 1988, Terrestrial Imaging Spectroscopy: Remote Sensing of, Environment, Vol. 24, n. 1, pp. 1-29. The Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) has been built on this principle and records 224 bands between the wavelengths of 0.41 to 2.45 u . Also, area arrays as shown in FIG. 4 have been used.
A 63 channel spectrometer which measures between 10.4 to 2.5 um wavelengths has also been developed and is described in Collins et al., 1983, Airborne Biogeochemical Mapping of Hidden Mineral Deposits, Econ. Geol.f Vol. 78, pp. 737-749. This latter system is referred to as the "Geophysical and Environmental Research Imaging Spectrometer" (GERIS). Common to all these new imaging spectrometers are better spectral and spatial sampling and necessarily rapid increases in data volumes.
The increased spectral sampling allows causative mineral recognition to become possible. As a
result of improved spectral characteristic information there have been several attempts to extract critical spectral characteristic information from the measured imaging spectrometer data to facilitate causative mineral identification. These have resulted in limited success due to low signal to noise ratios of the data. Other spectral characteristic identification algorithms include binary encoding and spectral ratioing for minerals with spectra lacking sharp absorption features. Binary encoding is fast, sensitive to the position of the absorption peaks, insensitive to albedo effects and obviously works poorly for minerals that do not possess well defined absorption peaks. The spectrum ratioing is sensitive to albedo effects and the position and strength of absorption peaks.
Recently, neural network algorithms have become a topic for intensive research in pattern recognition problems. See, e.g., McClelland et al., 1986, Parallel Distributed Processing: Exploration in the Microstructure of Cognition, Vol. 2, Physiological and Biological Models, Cambridge, MA, MIT Press, Bradford Books. In particular, the back propagation paradigm (Webos, 1975, Beyond Regression: New Tools for Prediction and Analysis in the Behavioral Sciences, PhD Thesis, Harvard University) has been shown to produce a trained neural net, which has the ability to recognize patterns under low signal to noise conditions. See Williams, 1990, The Use of Neural Networks in Recognizing Spatial Patterns in Dipole-Dipole Induced Polarization Data, Colorado School of Mines, Golden, CO, Internal Rpt. After what can be long training times, a neural network trained by the back propagation paradigm can offer an extremely fast method of automatically recognizing spatial patterns in digital data sets.
It is noted that the instant invention can be used with any of the above imaging spectrometer
systems, although for purposes of explanation the GERIS data was used.
Summary of Invention It is an object of the invention to increase the rate at which data from an imaging spectrometer can be analyzed to recognize different minerals.
It is a further object of the instant invention to provide a fast and efficient method of mapping mineral deposits either directly or indirectly from alteration minerals formed primarily from mineralizing processes utilizing imaging spectrometer data.
It is an additional object of the invention to use a trained neural network approach as described herein to recognize different minerals by their reflected electromagnetic spectra.
The instant invention uses trained neural networks to aid in the detection of mapping of minerals. In essence, a neural net consists of a set of weights or numbers with a defined connectivity. The weights or numbers are derived by a process called training (for the purpose of this invention) . In this invention a neural net is trained in particular to recognize certain patterns in the reflected electromagnetic. spectrum (0.5 - 2.5 um) . These training patterns can be derived from:
1. Airborne imaging spectrometer data itself with or without a continuum removed. 2. Ground field spectrometer measurements after calibration of the data to reflectance, convolution of the data to the band widths of the airborne imaging spectrometer data and continuum removal. 3. Laboratory measurements on samples taken from the area to be mapped by the airborne imaging spectrometer. Similar processes
as in 2 above apply. 4. Field or laboratory spectrometer measurements of material that are desired to be mapped, but which are not from the area to be mapped (e.g. if a certain alteration mineral assemblage has been noted in a locale that is associated with . a desired mineralization, then spectrometer measurements of this mineral assemblage can be used to map minerals in airborne imaging spectrometer data from another locale. The process of mineral identification uses two stages. The first is the training stage. Training of a neural net requires a training set and a training algorithm or paradigm. A training set consists of many training pairs. A training pair consists of an input vector, which in this case is spectrum reflectivity or absorption data or part thereof and a desired output vector. The form of the desired output vector is dependant upon the digital representation arbitrarily designated by the user for the particular mineral the neural net is being trained to recognize. The spectra used as input vectors in the training set can either be laboratory spectra, convolved to the resolution of the imaging spectrometer, spectra from the imaging spectrometer survey itself, or from in-the-field ground spectrometer measurements, again convolved to the resolution of the imaging spectrometer used in the mapping. A linear or nonlinear continuum can also be removed from the spectrum before training. A neural network can be trained to recognize one mineral or several minerals. A neural network can also be trained using one or several absorption features in a spectrum which are diagnostic of one mineral.
The training algorithm or paradigm is called back propagation (See Khanna, 1990, Foundations of
Neural Networks, Addison-Wesley Publishing Company, Menlo Park, CA, p. 196). Back propagation uses the Generalized Delta rule. The application of the Generalized Delta rule is described in detail below. Training consists of a forward pass through the neural network, resulting in an error or delta value being calculated based on the difference between the neural network calculated output vector and the desired output vector. This error is then back propagated through the neural network and used to adjust the weights of the neural network according to the generalized delta rule. Training is considered complete when the error values for each training pair (the difference between the neural network calculated output and the desired output vector) are considered to be a minimum. Training results in a trained neural network which consist of a set or net of weights or numbers, which are connected by a neural network topology.
The second stage of the process of mineral identification is to apply the trained neural network to imaging spectrometer data. First, the imaging spectrometer data is calibrated to reflectance values for the cases in which laboratory or in-the-field measured ground reflectance spectra are used for training. See Clark, et al., 1988, Calibration and
Evolution of AVRIS Data: Cripple Creek in October 1987, Proceedings, Airborne Visible/Infrared Imaging Spectrometer (AVRIS) Performance Evolution, JPL Publ., no. 88-38, pp. 49-61. If uncalibrated spectra from the airborne imaging spectrometer is used for training, then there is no need to use such a calibration in the data to be processed. If a continuum as in FIG. 2 has been removed in the training output spectra, then a continuum should be removed from the airborne imaging spectrometer data. See Clark et al., 1987, Automatic
Continuum Analysis of Reflectance Spectra, Proceedings, Third AIS Workshop, JPL Publication 87-30, JPL,
Pasadena, CA, pp. 138-142. The trained neural net is then used to process the airborne imaging spectrometer data or its calibrated and/or continuum removed equivalent. The output of the neural network indicates a vector (or scalar if only one number is output from the network, depending on user choice of the desired output vector) that is value dependant on the presence of the particular designated mineral to be mapped. Thus, each pixel in the data set is given a scalar number or vector indicating the presence of the mineral or the absence of the mineral so that detection or mapping can be achieved. A calculated output vector similar to a designated output vector for the mineral to be mapped indicates the presence of that mineral.
Brief Description of the Drawings
FIG. 1 is the reflected electromagnetic spectrum for the mineral alunite, as sampled by different imaging spectrometer systems.
FIG. 2 is an idealized and hypothetical depiction of geometrical parameters which describe an absorption feature.
FIG. 3 is the electromagnetic spectrum between 1 and IO20 Hz showing the different nomenclature for windows of the spectrum and physical phenomena that give rise to electromagnetic radiation in the respective windows.
FIG. 4 is a schematic perspective view of a type of imaging spectrometer, not drawn to scale.
FIG. 5 is a block diagram of an artificial neuron or node used in a neural network.
FIG. 6 is a diagram showing output layer weight adjustment, in a 3 layer, multilayer backpropagation neural network during training.
FIG. 7 shows a flow chart of a method of invention in which airborne imaging spectrometer data
is used for training a neural network.
FIG. 8 shows a flow chart of a method of invention in which field or laboratory measured spectra are used to train a neural network. FIG. 9 is an alteration map for the Cuprite mining district in Nevada.
FIG. 10 is a library of spectra derived from the image data of the area mapped in FIG. 9.
FIG. 11 is a map of the extent of silica as derived from an inversion recognition program.
FIG. 12 is a map of the extent of alunite as derived from an inversion recognition program.
FIG. 13 is a map of the extent of zeolite as derived from an inversion recognition program. FIG. 14 is the offset value spectrum for
Cuprite GERIS data with standard errors superimposed.
FIG. 15 is the gain factor spectrum for Cuprite GERIS data with standard errors superimposed.
FIG, 16 is the image derived from a trained neural network to recognize silica using all spectral bands except for 24-27.
FIG. 17 is the image derived from the trained neural network to recognize silica using the last 32 spectral bands. FIG. 18 is the image derived from the trained neural network to recognize alunite using the last 32 spectral bands.
FIG. 19 is the image derived from the trained neural network to recognize zeolite using the last 32 spectral bands.
Detailed Description of the Invention
In a preferred embodiment, a back propagation algorithm was used to compare the calibrated imaging spectrometer data to laboratory generated spectra of the same resolution. It is understood that the data could also be compared to actual ground measurements or
that other algorithms (or paradigms) or variants of back propagation could be used to train the neural network. The training of a neural network of weights was achieved using a back propagation algorithm although other training algorithms could be used. A neural network can be made of layers of neuron, or nodes, each performing the below function as shown in FIG. 6. 1 The first layer of input nodes are referred to herein as "layer 1" or the "input layer". Neural nets herein are constructed in such a way as that every node in the immediately higher layer (the input layer is the highest) is connected to every node in the next lower layer, but not to any other layer. Nodes are connected to other nodes via weights. This is not the only type of neural network that can be used, as input nodes can be connected directly to the output nodes in a three layer neural network for example. The connectivity is generally termed topology. Other modifications of network topologies are also possible.
The algorithm has an artificial neuron or model as its most fundamental identity. A neuron or node is shown in FIG. 5. A simple neuron can have a number of inputs (i.e., the calibrated data), with each value simply being an element of a input vector, which after multiplication by a correspondence (same dimension, and multiplication is in the form of a cross product) weight vector, are summed by the neuron and this sum is then multiplied by a squashing or activation function factor. The resulting output is then passed on. Mathematically, this can be expressed as:
OUT = F(NET) Where:
Xk = input vector for node Wk = weight vector
= summing function F = activation function
The activation function, F, should have the following properties: * It is everywhere differentiable
* It has a simple, easily calculable derivative
* It compresses the output values such that they lie between a convenient range, generally 0 and 1 (and in doing so provides an effective automatic gain control) .
Training requires a matched set of inputs and desired outputs. The input values of calibrated data can be normalized such that their range lies between 0 and 1 (for example), and are fed into the network through the input nodes. This normalization can optimize the performance of the training. Assuming a three (input, middle and output layers) layer network (FIG. 6), then for each middle layer node, each input node value is multiplied by a respective weight value
(generally initially set to some random number) and the products summed and then multiplied by the squashing function to form an output for that particular middle layer node. In mathematical notation: 0U p,j = vcι+e-NETP-*i) where
NET - = sum of the products of node values and weight values for the layer of nodes immediately above. This is the net variable for node p in layer j.
OUT • = output for node p in layer j This process is then repeated for all nodes in the last layer, and thus an output value is derived for each output node. This value is then compared to the desired or target value (user defined) to derive an error for each output node.
The error(s) are then back propagated up
through the neural net. For each output node, the error value is multiplied by the derivative of the activation function using the NET value of the particular output node as the argument for the function, to obtain a delta weight value. The delta weight value is then multiplied by a learning coefficient, LC, (which is user defined and which can be different for different layers and for different nodes within layers if so desired) and then by output value of an immediately above connecting node. The weight to be changed is only that weight that connects the particular hidden layer node to the particular output layer node, and it should be noted that this weight is termed the output weight. The product is the change that is to be added to the pre-existing weight value. This process is then performed for weights immediately above the output layer, thus modifying all weights between the output layer and the hidden layer for this example. In mathematical notation:
Dq,k = OUTqk(l-OUTqιk)*(TARGET-OUTqιk)
DWpq)k = LC*Dq,k*OUTp Wpqk(n+1)= pq,k(n) + DWpq,k* where: w pqk(n) = tne weight from neuron p in the hidden ' layer to q in the output layer (layer, k) at iteration n;
W k(n+l) =adjusted weight for the above connecting weight
D k = delta value associated with the output node q in layer k (output layer);
DW pqk = delta weight value or change in weight from node p in the hidden layer to node q in the output layer k;
OUT k = output value for node q in layer k; 0UT pj = output value for node p in hidden layer j;
LC = learning rate coefficient
For hidden layers - that is layers located between the input and output layers - the adjustment of weights is a little more complicated in the sense that there is no target value for error calculation, and hence the delta value cannot be calculated in the manner described above. The method of derivation of the delta value for a weight in a hidden layer involves the summation of the delta values for each connected output node. The delta value is the product of the error for each output node and the derivative of the activation function for each output node again using the NET of the output node as the argument in the function. The delta value for each output node is then multiplied by the updated output layer weight and these products are summed. The sum is then multiplied by the derivative of the activation for the hidden layer node, using the NET value of the hidden layer node as the argument to the function. This delta value is, then multiplied by the learning coefficient and the output of the connecting node in the above layer. In the case of a three layer network, this output would in fact be the normalized input value to the network.
D : = delta value for the p node in the j layer; 0UT pi = outPut °f node p in layer j;
D k = delta calculated for the connecting output node, Q, in layer k
This forward and backward pass through the network continues, modifying weights until an acceptable error (difference between calculated output vector and desired output vector) is obtained. The acceptable error level is user-defined. This sequence
of iterations or presentations is termed "training" or "learning". A common error measure is the mean square error (MSE) which is the mean of the errors for each complete pass through a training set. Once the training process is completed, then checks can be made on whether it has been effective. For example, the input only of each training set can be presented to the network, and the output then calculated. The calculated output values should show excellent correspondence to the desired output values. Other checks involve addition of random noise (different levels thereof) to input training sets prior to presentation to the trained network. Other checks can be made by presenting features that are considered to be important to the problem, to the network, and to assessing the network output.
FIG. 4 illustrates an imaging spectrometer 2 as it scans the area 3 to be analyzed or mapped. Objective 4 directs the reflected electromagnetic radiation through slit 5 where it is again directed through collimator 6 to dispersing element 7 (typically a grating) . Lens 8 directs the light onto a detector array 9. Not all wavelengths are reflected with equal intensity. Some wavelengths are preferentially absorbed, creating absorption features which are defined in a close enough sampled spectra. The instant invention primarily uses absorption features in the range of wavelengths 0.5 to 2.5 microns. This is termed the visible and near-infrared region. A date set is formed from the signals received by the detector array 9. Thus, after the spectra is measured, each pixel (or picture element) on the ground has a sample of the electromagnetic spectra consisting of digital numbers. These numbers are then converted to ground reflectance values (the process called calibration) . To help isolate the spectral feature, a
continuum is then removed. There are several ways to define continuum, including (1) using a segmented upper hull which may be adequate for high signal and low noise conditions, which are not that common in the currently used instruments, and (2) using a least squares best fit linear continuum for each different absorption feature. The calibrated data or calibrated continuum part of the spectrum is then compared to a trained neural network of weights to give a positive value as to a designated mineral for each pixel in the area to be mapped. Thus, for mapping, the absorption feature or combination of absorption features which is diagnostic of a particular mineral is used.
The process of the instant invention will be described briefly with particular reference to FIG. 8a and then described in more detail using GERIS data.
With reference to FIG. 7, signals from the detector of the imaging spectrometer 11 are converted into calibrated surface reflectance in the particular sample given. This is called the calibrated data 12. The spectra has its continuum removed, which is essentially taking out a background or long wavelength from the spectra, leaving a spectra of more distinctive absorption peaks. The calibrated data 12 is compared to a neural network of weights 13 derived using selected spectra from the imaging spectrometer data set. The output 14 indicates whether a particular mineral to be studied is present in a given pixel of the area to be mapped. Mineral mapping 15 can be achieved from the output 14 and the extent of the mineral can be measured.
With reference to FIG. 8, a similar process is shown, but in which the training set is derived from either field and/or laboratory spectrometer measurements from the area 21 to be mapped.
Alternatively, the field or laboratory samples can be from an area other than the area to be mapped. The
- 15-
area 21 will typically contain a number of discrete sample locations to produce data in the form of raw spectra. The raw spectra data is processed to convolve to the spectral resolution and to remove any unwanted continuum and is then processed by the neural network. The processed data can then be applied to imaging spectrometer data 23 which is first calibrated to remove reflectance 24, and a continuum may be removed 25. The output 26 from the calibrated data with the continuum removed 25 is then processed by the neural network to produce a neural network output 26 which can be mapped 27 to show mineralization.
The following example illustrated in FIG. 9-19 utilizes the neural network software to process airborne GERIS data of the Cuprite mining district of Nevada. A summary geology map of the Cuprite mining district of Nevada is shown in FIG. 9. The spectra used for training were derived from the imaging spectrometer data set over an area of known geology and ground spectra control. No attempt was made to isolate or analyze individual absorption features in this example. Eight spectra were used and nets were designed to recogniεe one of the eight spectra. The spectra used were those of basalt, alunite, buddingtonite, kaolinite, playa, silica, hematite and zeolite as shown in FIG. 10. The library of spectra derived from the image data is as follows: ALU - alunite, TUF= tuff, BUD = buddingtonite, IRO =• iron/hematite, KAO = kaolinite, PLA = playa, SIL = silica. Bands 24-27 were deleted due to high noise.
The output from the trained neural nets can be compared with the output from a spectral recognition program based on an inversion algorithm as shown in Boardman, 1989, Inversion of Imaging Spectrometry using Singular Value Decomposition, IGARSS Proc. 1989,
Vancouver. FIG. 11 is the map of the extent of silica as derived from the inversion recognition program.
FIG. 12 is a map of similarly derived alunite and FIG.
13 is zeolite. All GERIS spectra were calibrated by reference first to well known atmospheric carbon dioxide and water absorption bands, to calibrate the wavelength of the bands. The spectra were then converted to reflectance by reference to two known standard homogenous areas covered by the survey. An empirical correction is then designed, by assuming a constant gain and offset for each band to obtain a best fit (in the least squares sense), between the known sets of ground spectra and image field spectra for these standard areas. This calibration removes atmospheric effects (attenuation and scattering), viewing geometry effects and any residual instrument effects.
From this calibration and the inherent solving of an over-determined set of linear equations, estimations of the standard error for the offset and gain for each band can be obtained. FIGS. 14 and 15 show the standard errors for each of the bands., FIG.
14 shows the offset value spectrum for Cuprite GERIS data with standard errors superimposed. Bands 24-27 have been deleted due to high error. FIG. 15 shows the gain factor spectrum for Cuprite GERIS data with standard errors superimposed. Bands 24-27 have been deleted due to high error. It is obvious for each gain and offset factor, the most errors occur for bands with wavelengths greater than 1.8 micrometers. Unfortunately, this is the part of the spectra which contains many of the diagnostic absorption peaks.
Given this noise in the reflectance spectra and based on the results of the noise tolerance tests for the neural networks, it would appear that the neural nets should provide a suitable spectral recognition algorithm.
Training of the neural nets proved to be a linearly separable problem in that no hidden layer in
the neural network was required. Initial training used the entire spectra (57 useful spectral bands) and took typically less than 3 minutes elapsed time on a VAX 780 brand computer. The results of the application of the trained nets were variable. One of the better results are shown in FIG. 16 for silica recognition. When compared to FIG. 11 it is obvious that there is a good deal of similarity. (One should be aware that the geometrical aspect of the pixel sampling has not been corrected for in the images and hence the FIGS. 16-19 are compressed in the north-south direction) . Upon consideration of the more critical differences between the spectral types (FIG. 11) it is obvious that there are more distinguishable features in the bands with wavelengths greater than 1.8 micrometers. Consequently the training exercise was repeated using the last 32 bands which sample the reflectance in this critical portion of the spectra. Again, the training required no hidden layer and training took typically less than 2 minutes elapsed time. Similarly, FIG. 11 is similar to FIG. 17 where only the last 32 spectral bands were used. Similarities can also be seen between FIG. 12 and FIG. 18 for alunite and FIG. 13 and FIG. 19 for zeolite. In FIGS. 16-19 a linear stretch has been applied to the output recognition elements.
Using the inversion recognition algorithm or another expert system style of classification required processing times in the order of four hours while the neural network took approximately fives minutes of elapsed time to run on a VAX 780 brand computer. Real time processing is possible with the newly developed neural chips.
Using the process outlined above, mineral spectral types indicative of vegetation types, rock types, geomorphological features, healthy crops and others can also be recognized given sufficient data sets (i.e. spatial resolution and ground control).
It can be appreciated that the training and use of the neural network can be accomplished in a variety of ways using a desired combination of neurons, weights and layers. For purposes of the claims herein, the phrase "neural network" means any system that is
"trained" to manipulate a set of input data indicative of the reflectivity or absorption characteristics at one or more electromagnetic spectra using a plurality of functions (which preferably are "weights" or sealer functions) to produce an output indicative of a mineral or other material. The "training" is accomplished by assigning a value to a material to be detected, inputting data to the neural network (from laboratory or other known sources) that corresponds to such material, manipulating the data to obtain an output, and then adjusting the manipulating functions based on the difference between the assigned value and the actual obtained value. The training process may be performed repetitively to increase the precision of the neural network. The neural network may include multiple layers of functions.
Claims
1. A method for detecting at least one material in an area, comprising receiving reflected electro- magnetic radiation from said area, converting said received electromagnetic radiation to data indicative of reflectance, processing said data with a neural network trained to recognize said material to produce a neural network actual output value, and comparing said actual output value with a predetermined assigned value indicative of the presence of said material.
2. The method of claim 1, wherein said data includes the reflectivity of the area of a plurality of spectra, and the neural net includes a plurality of functions in a first neural network layer whereby the spectra reflectivity data is divided into discrete bands that are processed by said functions and then the data processed by the first layer is combined to established a calculated output of said neural network first layer.
3. The method of claim 2, wherein said functions are numbers and wherein the combining of the processed data is by addition.
4. The method of claim 2, wherein the neural network includes a second plurality of functions in a second neural network layer, whereby said second plurality of functions processes the calculated output of said first layer and then the data processed by the second layer is combined to establish a calculated output of said second layer.
5. The method of claim 4, wherein said functions are numbers and wherein the combining of the processed data is by addition.
6. The method of claim 1, wherein said material to be detected is a mineral.
7. The method of claim 2, wherein said neural network is trained to recognize said material by establishing a predetermined output value indicate of the presence of said material, processing reflectivity data that is known to be indicative of such calculated material through the neural network to produce a neural network output value, determining the difference between the predetermined value and the calculated output value, and adjusting the functions in the neural network based on said difference.
8. The method of claim 7, wherein said known reflectivity data is processed by the neural network with the adjusted functions to establish a new calculated neural network output value, determining the difference between the predetermined value and the new calculated output value, adjusting the functions -in the neural network based on said difference if said difference exceeds a threshold difference, and repeating the steps of this claim until said difference does not exceed the threshold difference.
9. The method of claim 8, wherein said functions are numbers.
10. The method of claim 9, wherein the functions that are adjusted are those that will most reduce the difference between the predetermined output value and the calculated output value.
11. The method of claim 1, further comprising establishing an image of said area indicating the extent of said material.
12. The method of claim 11, wherein said method is used to detect a plurality of different materials, each- material having a different predetermined assigned value indicative of the presence of said material.
13. An apparatus to detect at least one material in an area, comprising means for processing data indicative of reflected electromagnetic radiation from said area, the processing means including a neural network trained to recognize said material to produce an output, and means for comparing said output with a predetermined assigned value indicative of the presence of the material.
14. The apparatus of claim 13, wherein said data includes the reflectivity of the area at a plurality of spectra, and the neural net includes a plurality of functions in a first neural network layer whereby the spectra reflectivity data is divided into discreet bands that are processed by said functions and then the data processed by the first layer is combined to establish a calculated output of said neural network first layer.
15. The apparatus of claim 14, wherein said functions are numbers and wherein the combining of the processed data is by addition.
16. The apparatus of claim 14, wherein the neural network includes a second plurality of functions in a second neural network layer, whereby said second plurality of functions processes the calculated output of said first layer and then the data processed by the second layer is combined to establish a calculated output of said second layer.
17. The apparatus of claim 16, wherein said functions are numbers and wherein the combining of the processed data is by addition.
18. The apparatus of claim 13, wherein said material to be detected is a mineral.
19. The apparatus of claim 14, wherein said neural network is trained to recognize said material by establishing a predetermined output value indicative of the presence of said material, processing reflectivity data that is known to be indicative of such calculated material through the neural network to produce a neural network output value, determining the difference between the predetermined value and the calculated output value, and adjusting the functions in the neural network based on said difference.
20. The apparatus of claim 19, wherein said reflectivity data is processed by the neural network with the adjusted functions to establish a new calculated neural network output value, and wherein said data processing means determines the difference between the predetermined value and the new calculated, output value and adjusts the functions in the neural network based on said difference if said difference exceeds a threshold difference and repeats the steps of this claim until said difference does not exceed the threshold difference.
21. The apparatus of claim 20, wherein said functions are numbers.
22. The apparatus of claim 21, wherein the functions that are adjusted are those that will most reduce the difference between the predetermined output value and the calculated output value.
23. The apparatus of claim 14, further comprising mapping means for mapping or image indicate of said material.
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/US1992/000593 WO1993019426A1 (en) | 1992-03-25 | 1992-03-25 | Method of detecting and mapping minerals and other geological phenomena utilizing airborne imaging spectrometers |
AU36552/93A AU3655293A (en) | 1992-03-25 | 1992-03-25 | Method of detecting and mapping minerals and other geological phenomena utilizing airborne imaging spectrometers |
CA002101519A CA2101519A1 (en) | 1992-03-25 | 1992-03-25 | Method of detecting and mapping minerals and other geological phenomena utilizing airborne imaging spectrometers |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/US1992/000593 WO1993019426A1 (en) | 1992-03-25 | 1992-03-25 | Method of detecting and mapping minerals and other geological phenomena utilizing airborne imaging spectrometers |
CA002101519A CA2101519A1 (en) | 1992-03-25 | 1992-03-25 | Method of detecting and mapping minerals and other geological phenomena utilizing airborne imaging spectrometers |
Publications (1)
Publication Number | Publication Date |
---|---|
WO1993019426A1 true WO1993019426A1 (en) | 1993-09-30 |
Family
ID=25676428
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/US1992/000593 WO1993019426A1 (en) | 1992-03-25 | 1992-03-25 | Method of detecting and mapping minerals and other geological phenomena utilizing airborne imaging spectrometers |
Country Status (1)
Country | Link |
---|---|
WO (1) | WO1993019426A1 (en) |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO1997034169A1 (en) * | 1996-03-12 | 1997-09-18 | Neural Mining Solutions Pty. Ltd. | Method for analysing data in the exploration for minerals |
WO2000016126A1 (en) * | 1998-09-15 | 2000-03-23 | Scientific Prediction, Inc. | System and method for delineating spatially dependent objects, such as hydrocarbon accumulations from seismic data |
US6574565B1 (en) | 1998-09-15 | 2003-06-03 | Ronald R. Bush | System and method for enhanced hydrocarbon recovery |
US7991717B1 (en) | 2001-09-10 | 2011-08-02 | Bush Ronald R | Optimal cessation of training and assessment of accuracy in a given class of neural networks |
WO2014177180A1 (en) * | 2013-04-29 | 2014-11-06 | Mumelter Michael | A method and apparatus for detection of natural resources |
CN110060173A (en) * | 2019-04-27 | 2019-07-26 | 烟台市牟金矿业有限公司 | A kind of deep gold deposit method of prospecting |
RU2706515C1 (en) * | 2018-08-08 | 2019-11-19 | Автономная некоммерческая образовательная организация высшего образования "Сколковский институт науки и технологий" | System and method for automated description of rocks |
WO2021107869A1 (en) * | 2019-11-25 | 2021-06-03 | Ams Sensors Singapore Pte. Ltd. | Spectrometer calibration |
WO2023070915A1 (en) * | 2021-10-26 | 2023-05-04 | 核工业北京地质研究院 | Method for determining distribution of uranium mine |
CN117191827A (en) * | 2023-11-01 | 2023-12-08 | 核工业北京地质研究院 | Method for defining sandstone-type uranium ore-forming favorable section in area to be surveyed |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4809347A (en) * | 1986-07-18 | 1989-02-28 | Hughes Aircraft Company | Computer vision architecture |
US4906940A (en) * | 1987-08-24 | 1990-03-06 | Science Applications International Corporation | Process and apparatus for the automatic detection and extraction of features in images and displays |
US4912647A (en) * | 1988-12-14 | 1990-03-27 | Gte Laboratories Incorporated | Neural network training tool |
US5046020A (en) * | 1988-12-16 | 1991-09-03 | E. I. Du Pont De Nemours And Company | Distributed parallel processing network wherein the connection weights are generated using stiff differential equations |
US5052043A (en) * | 1990-05-07 | 1991-09-24 | Eastman Kodak Company | Neural network with back propagation controlled through an output confidence measure |
US5124915A (en) * | 1990-05-29 | 1992-06-23 | Arthur Krenzel | Computer-aided data collection system for assisting in analyzing critical situations |
US5155801A (en) * | 1990-10-09 | 1992-10-13 | Hughes Aircraft Company | Clustered neural networks |
-
1992
- 1992-03-25 WO PCT/US1992/000593 patent/WO1993019426A1/en active Application Filing
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4809347A (en) * | 1986-07-18 | 1989-02-28 | Hughes Aircraft Company | Computer vision architecture |
US4906940A (en) * | 1987-08-24 | 1990-03-06 | Science Applications International Corporation | Process and apparatus for the automatic detection and extraction of features in images and displays |
US4912647A (en) * | 1988-12-14 | 1990-03-27 | Gte Laboratories Incorporated | Neural network training tool |
US5046020A (en) * | 1988-12-16 | 1991-09-03 | E. I. Du Pont De Nemours And Company | Distributed parallel processing network wherein the connection weights are generated using stiff differential equations |
US5052043A (en) * | 1990-05-07 | 1991-09-24 | Eastman Kodak Company | Neural network with back propagation controlled through an output confidence measure |
US5124915A (en) * | 1990-05-29 | 1992-06-23 | Arthur Krenzel | Computer-aided data collection system for assisting in analyzing critical situations |
US5155801A (en) * | 1990-10-09 | 1992-10-13 | Hughes Aircraft Company | Clustered neural networks |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO1997034169A1 (en) * | 1996-03-12 | 1997-09-18 | Neural Mining Solutions Pty. Ltd. | Method for analysing data in the exploration for minerals |
WO2000016126A1 (en) * | 1998-09-15 | 2000-03-23 | Scientific Prediction, Inc. | System and method for delineating spatially dependent objects, such as hydrocarbon accumulations from seismic data |
US6236942B1 (en) | 1998-09-15 | 2001-05-22 | Scientific Prediction Incorporated | System and method for delineating spatially dependent objects, such as hydrocarbon accumulations from seismic data |
US6411903B2 (en) | 1998-09-15 | 2002-06-25 | Ronald R. Bush | System and method for delineating spatially dependent objects, such as hydrocarbon accumulations from seismic data |
US6574565B1 (en) | 1998-09-15 | 2003-06-03 | Ronald R. Bush | System and method for enhanced hydrocarbon recovery |
US7991717B1 (en) | 2001-09-10 | 2011-08-02 | Bush Ronald R | Optimal cessation of training and assessment of accuracy in a given class of neural networks |
WO2014177180A1 (en) * | 2013-04-29 | 2014-11-06 | Mumelter Michael | A method and apparatus for detection of natural resources |
RU2706515C1 (en) * | 2018-08-08 | 2019-11-19 | Автономная некоммерческая образовательная организация высшего образования "Сколковский институт науки и технологий" | System and method for automated description of rocks |
CN110060173A (en) * | 2019-04-27 | 2019-07-26 | 烟台市牟金矿业有限公司 | A kind of deep gold deposit method of prospecting |
CN110060173B (en) * | 2019-04-27 | 2021-05-18 | 烟台市牟金矿业有限公司 | Deep gold deposit forming and prospecting method |
WO2021107869A1 (en) * | 2019-11-25 | 2021-06-03 | Ams Sensors Singapore Pte. Ltd. | Spectrometer calibration |
WO2023070915A1 (en) * | 2021-10-26 | 2023-05-04 | 核工业北京地质研究院 | Method for determining distribution of uranium mine |
CN117191827A (en) * | 2023-11-01 | 2023-12-08 | 核工业北京地质研究院 | Method for defining sandstone-type uranium ore-forming favorable section in area to be surveyed |
CN117191827B (en) * | 2023-11-01 | 2024-01-30 | 核工业北京地质研究院 | Method for defining sandstone-type uranium ore-forming favorable section in area to be surveyed |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Keshava | A survey of spectral unmixing algorithms | |
Benediktsson et al. | Classification of multisource and hyperspectral data based on decision fusion | |
Briottet et al. | Military applications of hyperspectral imagery | |
WO1993019426A1 (en) | Method of detecting and mapping minerals and other geological phenomena utilizing airborne imaging spectrometers | |
Kerekes et al. | An analytical model of earth-observational remote sensing systems | |
Wiersma et al. | Analytical design of multispectral sensors | |
AU3655293A (en) | Method of detecting and mapping minerals and other geological phenomena utilizing airborne imaging spectrometers | |
Kerekes et al. | Parameter trade-offs for imaging spectroscopy systems. | |
Wang et al. | Simulated and observed backscatter at P-, L-, and C-bands from ponderosa pine stands | |
Makki | Hyperspectral imaging for landmine detection | |
El_Rahman | Big data analysis: hyperspectral image processing for agriculture applications | |
Endsley | Spectral unmixing algorithms based on statistical models | |
Manolakis et al. | Statistical models for LWIR hyperspectral backgrounds and their applications in chemical agent detection | |
Wang et al. | Rocks/minerals information extraction from EO-1 Hyperion data base on SVM | |
Maksimenko et al. | Applying of methods of processing of hyperspectral data for identification of traces of explosives | |
Qamar et al. | The effects of atmospheric modeling covariance on ground-based hyperspectral measurements of surface reflectance | |
Graña et al. | On endmember detection in hyperspectral images with morphological associative memories | |
Picon et al. | Biologically-inspired data decorrelation for hyper-spectral imaging | |
Pargal | Hyperspectral subspace identification and endmember extraction by integration of spatial-spectral information | |
Goldberg | A performance characterization of kernel-based algorithms for anomaly detection in hyperspectral imagery | |
Haliloğlu | Dimension reduction methods for hyperspectral imagery | |
de Carvalho et al. | Spectral change detection | |
Klein et al. | Physics-guided neural networks for hyperspectral target identification | |
Mishra et al. | Neurally augmented object identification scheme for hyperspectral images | |
Yatsenko et al. | Estimation of chlorophyll concentration in vegetation using global optimization approach |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
WWE | Wipo information: entry into national phase |
Ref document number: 2101519 Country of ref document: CA |
|
AK | Designated states |
Kind code of ref document: A1 Designated state(s): AU BR CA FI JP |
|
AL | Designated countries for regional patents |
Kind code of ref document: A1 Designated state(s): AT BE CH DE DK ES FR GB GR IT LU MC NL SE |
|
DFPE | Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101) | ||
122 | Ep: pct application non-entry in european phase |