AU2004297326A1 - Method and device for the at least semi-automated evaluation of remote sensing data - Google Patents

Abstract

The invention relates to a method and a device for the at least semi-automated evaluation of remote sensing data, by means of identification parameters. According to said method, objects are identified and/or classified using image characteristics, the remote sensing data exists in the form of a time series of multi-spectral remote sensing data (11 to 17) and the identification parameters are determined from the remote sensing data in a multi-spectral and multi-temporal manner.

Description

VERIFICATION OF TRANSLATION 1, Dr. Joachim Brunotte, Patent Attorney of Effert, Bressel und Kollegen Radickestrasse 48 12489 Berlin Germany verify that the following document is a true translation of International Patent Application No. PCT/EP2004/014052 (and amendments) to the best of my knowledge and belief (Signature) DATED this 2. day of May 2006 DLR AP 22/103 Method and apparatus for at least partially automated evaluation of remote sensing data The invention relates to a method and an apparatus for 5 at least partially automated evaluation of remote sensing data, with objects being identified and/or classified on the basis of image features. Features which can be acquired by optical sensors are referred to as image features. Image features which can be used 10 for identification and/or classification are, for example, the gray shade, the color, the multispectral signature, the surface structure (texture), the environment and/or the shape of a recorded object. 15 It is known for remote sensing data to be preprocessed automatically for identification and/or classification. Known methods for automatic atmosphere correction are used, for example, in the programs DurchBlick or SMAC (Simplified Method for Atmosphere Correction). 20 Automation of the identification and/or classification of remote sensing data is becoming increasingly important against the background of the ever better access to remote sensing data and the databases which 25 are becoming ever larger. A distinction can be drawn between two procedures for identification and/or classification. On the basis of a first procedure, images to be investigated are 30 segmented on the basis of a statistical rule of principle or with the aid of training regions, and individual segments are then associated with specific classes and/or objects. On the basis of a second procedure, the images are searched through specifically 35 for specific classes and/or objects.

DLR - 2 AP 22/103 defined in advance for automatic identification and/or classification. Identification parameters are criteria in feature space, on the basis of which it is possible to decide whether a pixel is or is not associated with 5 an object. Identification parameters can be defined on the basis of empirical values and/or known object characteristics. 10 However, the constraints must also be at least approximately constant for use of automatic identification and/or classification of objects on the basis of static, that is to say permanently defined, 15 identification parameters. Since constant constraints cannot be assumed either for an individual scene or for an entire time series, rigid identification parameters must be defined very widely 20 in order to allow largely all recording conditions to be taken into account. The options for automation on the basis of rigid identification criteria are thus severely restricted. 25 Further interactive programs are known, for example the programs ENVI/RSI or ERDAS/IMAGINE. In this case, the identification parameters are matched to changing constraints by interaction with a user. The user defines the classification model by definition of 30 rejection radii or by the selection of test regions. The classification result is generally influenced highly subjectively. Furthermore, this work is associated with a large amount of manual effort, and the user must be appropriately trained in advance. 35 Apart from requiring a considerable amount of time, a DLR - 3 AP 22/103 Furthermore, (semi)automatic methods are known, for example ECOGNITION/DEFINIENS, which allow even relatively large amounts of data to be interpreted. The 5 program is trained for a predetermined field by optimization of various interpretation parameters. The lack of parameters that need to be controlled manually associated with this during direct classification results in this program package producing 10 comprehensible objective results, and being suitable for inclusion in automatic process chains. However, the use of this software in automatic process chains is limited by the fact that, although it is based on highly developed scientific imaging algorithms, it 15 essentially does not support any physical and biological approaches. Information about constraints that are known a priori thus cannot be taken into account in the optimization process. Without additional optimization, a defined classification model can be 20 used only for a restricted time period, in which the constraints remain unchanged. Furthermore, methods are known for identification and/or classification using neural networks. Neural 25 networks also require complex training, in a similar manner to the method by means of optimization. Remote sensing signatures of known landscape objects are used for this purpose, in order to develop a classification model. However, it has been found that neural networks 30 can produce incorrect results if they have not been trained sufficiently, and/or have been trained incorrectly. Furthermore, neural networks cannot be generalized, that is to say a neural network is applicable only to the value range of the training 35 data.

DLR - 4 AP 22/103 of morphologically and structurally complex objects in an object space, which comprises the following steps: a) recording of topometric data and information in the object space, b) evaluation of the topometric data and 5 information and c) reproduction of the evaluated topometric data and information as machine-legible data records, or data records which can be perceived by the sensors. When analyzing data records which are intended to be investigated, reference data records can be 10 produced in order to check the results. Furthermore, analyses can be carried out using object, pattern and shape recognition methods. If the results produced from the analyses are used further, the investigated objects can be classified on the basis of the results of the 15 analyses, from the respective specialist viewpoints. The publication "Interpreting ERS SAR Signatures of Agricultural Crops in Flevoland, 1993 - 1996", by Paul Saich and Maurice Borgeaud, IEEE Transactions on 20 Geoscience and Remote Sensing, Vol. 38, No. 2, March 2000, pages 651 - 657, describes the interpretation of radar signatures of agricultural plant systems. Indications have been found that rules, temperature and incidence angle interfere with multitemporal radar 25 signatures. The derivation of direct physical relationships between the radar signatures and the characteristics of the vegetation and of the ground requires that these effects be corrected. However, it has not been possible to carry out appropriate 30 corrections. The invention is therefore based on the technical problem of providing a method and an apparatus by means of which objects can be identified and/or classified 35 more reliably when the constraints vary.

DLR - 5 AP 22/103 features in patent claims 1 and 9. Further advantageous refinements of the invention can be found in the dependent claims. 5 For this purpose, identification parameters are defined multitemporally for at least partially automated evaluation of remote sensing data, with objects being identified and/or classified on the basis of image features. The remote sensing data is in the form of a 10 multispectral remote sensing data time series. The definition of the identification parameters depending on function of time makes it possible to take into account effects which are dependent on the time of day and/or the time of year. Effects which are dependent on 15 the time of day are, for example, shadows, reflection characteristics of objects on the basis of the position of the sun, etc. Effects which are dependent on the time of year are, for example, the coloring of leaves on deciduous trees and/or the changes in the plant 20 growth on soily ground with the time of year. Identification parameters may be optimized for specific relationships. The matching to temporal effects can be automated without any interaction by a user. 25 In a further embodiment, data is reconstructed in the time/space domain. For data reconstruction in the time/space domain, pixels which are contaminated by temporary disturbances are identified and, if possible, are replaced by interpolation methods. Pixels in a time 30 series are contaminated, for example, by clouds, mist, and/or temporarily by snow. The interpolation process is preferably carried out by means of time-series analysis and geostatic methods. This means that the interpolation is carried out not only in the time 35 domain over successive data items in the remote sensing DLR - 6 AP 22/103 into account in this process, if possible. In a further embodiment, identification parameters are defined at least dependent on spectral signatures and a 5 temporal land usage characteristic. The temporal profile of the land usage characteristic is, for example, the phenology of vegetation. In a further preferred embodiment, identification 10 parameters are reclassified iteratively, with at least the temporal dependency of land usage characteristics being reclassified. The iterative reclassification over the time profile of the land usage characteristics allows further improvement to the classification 15 results. In a further preferred embodiment, a time-of-year profile is documented in a data time series, and land usage characteristics are adapted to the acquired time 20 of-year profile. In a further preferred embodiment, changes in the land usage characteristics which are dependent on the time of year are predicted. 25 In a further preferred embodiment, an object is identified and/or classified on the basis of three dimensional fuzzy functions. In this case, it is feasible to use suitable correlation methods if a 30 sufficient number of support points distributed throughout the time of year are available. A combination such as this allows automatic determination of the landscape characteristics without interactive interventions and/or expert knowledge. The completely 35 automatic method can be used universally for remote DLR - 7 AP 22/103 satellite images themselves. The following text relates in particular to the identification and/or classification of an object (or 5 of a plurality of objects) which is (are) recorded by remote sensing but is (are) not known. In an ideal evaluation of the radiation characteristics (in particular the reflection behavior), the available measured values are compared over the entire spectral 10 range with the radiation spectrum of possible different object types or object classes. The object can be allocated to the object type or to the object class on the basis of this comparison. In practice, however, measured values are generally available only in a 15 finite number of spectral ranges of different width. This exacerbates the association process. In particular, at least one spectral signature of an object (preferably a plurality of spectral signatures 20 of the object) is thus evaluated, for example in different suitable spectral ranges, with each of the spectral signatures describing radiation characteristics of the object. When evaluating the individual spectral signatures, it is proposed that 25 fuzzies (also referred to as probability functions) be used in each case. These fuzzies can be used to derive an association measure in each case from the values of the radiation characteristics or values of a variable derived from them (for example for each of the possible 30 object types) . These derived association measures from the various spectral ranges of each object type can then be linked to one another in a suitable manner (for example by addition or multiplication) to form an overall association measure. For example, the acquired 35 object can be associated with a specific object type or DLR - 8 AP 22/103 In addition, the spectral signatures can describe the radiation characteristics of the objects in a class or in a type at in each case one defined time or in one 5 defined time period. This makes it possible to distinguish between different object types additionally by association of one object type with a defined time (for example a vegetation time) . Furthermore, spectral signatures which describe the radiation characteristics 10 of an object at different times and/or in different time periods can be evaluated, and this makes it possible to improve the reliability of the object association process. 15 Preferred details in conjunction with the fuzzies will also be described. In this case, the weighting measures are referred to as probabilities that the evaluated object would correspond to an expected object type, defined by the fuzzies. 20 In particular, the fuzzies may be specific probability functions which have a maximum of the weighting measure which extends over a range of values of a radiation characteristic of the object. In order to take account 25 of widely differing constraints, the maximum occurs not just at a single value of the radiation characteristic. In particular, when a specific point in the spectrum is considered, the maximum may extend over a range of the values. Alternatively or additionally, the fuzzies may 30 be specific probability functions in the sense that the probability of a maximum does not decrease suddenly, but continuously or with a functional profile which is determined in some other way over a range of values of the radiation characteristic to a lower probability 35 level, in particular to zero.

DLR -9 AP 22/103 of the probability makes it possible on the one hand to identify with a specific probability objects with characteristics which are similar but can be distinguished as being associated with the same object 5 type. Different objects of a specific type or class may, however, on the other hand have very similar characteristics in the investigated spectral ranges, or the effect of the various constraints on the measurement signal in these spectral ranges may lead to 10 association being possible with only two or more object types at the same time. Further methods or spectral information which can be differentiated may be required in order to further increase the association probability with one object type. 15 The invention will be explained in more detail in the following text using one preferred exemplary embodiment. In the figures: 20 Figure 1 shows a schematic illustration of multitemporal (for example seasonal) spectral signature of a landscape object, 25 Figure 2a shows a schematic illustration of the spectral signature of one object type which, for example, can be determined, but only in the spectral ranges A, B and C, 30 Figure 2b shows a schematic illustration of a model of the spectral signature from Figure 2a at the wavelength A by means of a fuzzy, 35 DLR - 10 AP 22/103 Figure 2a at the wavelength B by means of a fuzzy, Figure 2d shows a schematic illustration of a 5 model of the spectral signature from Figure 2a at the wavelength C by means of a fuzzy, Figure 3 shows a schematic illustration of a 10 phenological profile of a field of wheat, and associated transition probabilities between different development and processing stages, and 15 Figure 4 shows a schematic illustration of a classification processor. Figure 1 shows, schematically, spectral signatures 11-17 of a landscape object over the course of a year. 20 Landscape objects are, for example, soils, woods, meadows, deserts, lakes and seas etc. Landscape objects have specific image features which can be used for identification and/or classification. These include, for example, the gray shade, the color, the spectral 25 signature, the surface structure (texture) and/or the shape of an acquired object. The spectral signatures 11-17 show the intensity R as a percentage plotted against the wavelength X. Spectral signatures and/or other image features are in general not time-invariant 30 but change over the course of the day and/or year. According to the invention, multitemporal identification parameters are formed in order to take account of the time dependency. The identification parameters are preferably mapped as a fuzzy set. At 35 least one fuzzy set (this will be described in more DLR - 11 AP 22/103 However, the invention is not restricted to the type of spectral signatures shown in the figures. In fact, the expression "a spectral signature" should be understood 5 as meaning at least one physical measurement variable or a variable derived from it, which describes the radiation received from the observed object and/or an observed scene as a function of the wavelength, as a function of the radiation frequency and/or as a 10 function of an equivalent variable. In the example, the physical variable is the intensity R, for example the effective reflectivity of the scene. Figure 2a shows, schematically, the spectral signature 15 12 from Figure 1. Figures 2b-2d show models of the spectral signature 12 by means of fuzzies for the wavelengths A, B and C. The individual fuzzies are, in particular, probability 20 functions which in each case allocate a probability value to the values (in this case: intensities) of the spectral signature (or, in general, to values of the image feature mentioned above, which may be used for identification and/or classification) . The association 25 relates in particular to one of different object types, that is to say a single fuzzy is defined with respect to one specific object type. In this case, the fuzzies may be defined for a value range (in this case: the areas of the intensities illustrated in Figure 2b to 30 Figure 2d, for example reflectivity from 0 to 100%) of the image feature of one object type. One fuzzy set can thus be used analogously to 2b) to 2d) for evaluation for each further object type. 35 The fuzzies can be used to evaluate in each case one DLR - 12 AP 22/103 (seasonal) time. The same fuzzies may, of course, be used to evaluate a plurality of picture elements, local areas and/or the entire image. In this case, for example, the individual picture element may be 5 evaluated successively and/or at the same time using the fuzzies. Furthermore, the individual fuzzies (in the present example, illustrated in Figure 2b to Figure 2d) can 10 each be defined for one specific wavelength (in general for one point in the radiation spectrum). However, they may also be defined for one or more continuous ranges of wavelengths (in general a spectral area) (for example by a mathematical rule) or may be specified as 15 a field of probability values which quasi-continuously cover a spectral range with a given resolution. This means that a probability value can be specified for each object type for each wavelength in the range by the fuzziness given by the resolution. 20 An association such as this, which allocates probability values not only for one wavelength but for a plurality of wavelengths or for a spectral range, can additionally change over time (for example depending on 25 the time of year), so that this results in a three dimensional association. For example, in the case of the record illustrated in Figure 2b to Figure 2d, the three individual fuzzies for one object type allow their variation over time to be taken into account, 30 that is to say a three-dimensional fuzzy set is defined and used overall for this object type. Individual spectral signatures (that is to say in each case one spectral signature for one observation time) 35 of one object (for example a so-called land coverage DLR - 13 AP 22/103 to determine whether the object is associated with an already defined class of object. For example, the landscape objects mentioned above may be evaluated. In this case, these "fuzzy sets" describe the spectral 5 signature at different wavelengths. Individual fuzzies are linked in the specific exemplary embodiment, which will be explained in more detail here with reference to the figures, for example using known 10 methods to form a common consense. The expression consense means in particular those fuzzy sets which preferably correspond to the various characteristics of one specific object type. In one preferred embodiment of the invention, which takes into account the change 15 over time of the radiation characteristics of objects, the expression "consense" may also mean combinations of fuzzies which correspond to the radiation characteristics at different times. 20 The greater the number of fuzzies which are used for modeling (evaluation) of the spectral signature 12 (or in general for any given spectral signature), the less ambiguously a specific landscape object can be identified at the observation time from the spectral 25 signature. However, the number of the required logic operations (in the simplest case addition or multiplication of the individual probabilities for the association with one object type) of individual fuzzies to form the common consense also rises. 30 As can be seen from Figure 2a, the comparison of the reflection values in the spectral sections A, B and C with the predetermined fuzzy set for one object type is to be performed, with the respective individual 35 probabilities wA, WB and wc being determined and DLR - 14 AP 22/103 different object types, there must accordingly be m different fuzzy sets, analogously to 2b) to 2d) . With the large number and variability of the object types in remote sensing, it is highly probable that these cannot 5 be distinguished from one another by the evaluation of the measured signals at the points A, B and C (see Figure 2). This means that the measurements are carried out in a greater number m of spectral ranges, in which case the m comparisons for each object type are then 10 also linked to one another to form an assignment probability statement. However, in general, the aim is to use the fuzzies to map selected characteristics, for example at least all 15 the relevant characteristics. In the exemplary embodiment, the measured intensities at that wavelength for which the respective fuzzy set is defined in each case correspond to the characteristics. In the present example, the characteristics include a mean intensity R 20 (reflection value) at the wavelength A, a lower intensity R at the wavelength B, as well as a high intensity at the wavelength C. The better the match between the recorded intensity R and an intensity which is modeled by means of a fuzzy and is theoretically 25 associated with the landscape object (in particular the intensity at the maximum probability of a fuzzy) , the greater is the probability that this is that landscape object which is described precisely by the fuzzy in precisely that wavelength range. In the exemplary 30 embodiment, the probabilities are normalized on a scale between 0 and 1 for a logic operation on the individual results. The fuzzies are, in particular, specific association 35 probability functions which (as can be seen from the DLR - 15 AP 22/103 and only one defined time period) associate a maximum with the probability for a range of the image feature values (in this case: intensities R) . Because of the widely differing constraints, the radiation behavior 5 (for example the reflection behavior) of an object may vary to a greater or lesser extent, so that the signal which is measured for this object, in this case for example the reflection behavior, also changes. The assignment probability of "1" should therefore be 10 allowed for a range of R. Outside this range, the probability that this is the object type with precisely this fuzzy, and in comparison with all the measured spectral ranges, the fuzzy set, still in fact is this object type, becomes ever less, and in the end falls to 15 "0". It can then be assumed that a probability of more than "0" can be determined by means of the comparison with fuzzies of other object types. In the case of three-dimensional fuzzies, the time variations of the object types are then also taken into account. In some 20 circumstances, this thus results in a shift on the axis by R for the fuzzy set in Figures 2b to 2d, although the width and the edge gradient of the trapezoidal shape of the fuzzies may, however, also differ considerably. The maximum 1, normalized on a scale 25 between 0 and 1, that is to say the probability of the image feature value being associated with an object class, is preferably 1 (or 100%), however, for each wavelength. In other words, an image feature value completely matches the expectation when it is in the 30 region of the maximum, although, however, a plurality of object types may match this expectation at the same time. Alternatively or additionally, the fuzzies are in 35 particular specific probability functions in the sense DLR - 16 AP 22/103 relationship over a range of image feature values (in this case: intensities) to a low probability level, in particular to zero. The refinement of fuzzies as described in the previous paragraph makes it possible 5 in particular to assess individual characteristics of the observed object (for example intensity not only at one specific wavelength but in addition at one defined time or in one defined time period) to have a probability between 0 and 1. Objects with 10 characteristics which vary with the seasons can thus be identified using the same fuzzies, but with seasonal variations, as being associated with the same object class in one seasonal time period. Tolerances in the characteristics are allowed within the limits defined 15 by the individual fuzzies, so that the probability of the presence of a corresponding object type does not become zero at this stage, because the characteristic differs slightly from a specific, theoretically present image feature value. 20 If only a single characteristic (for example the intensity at one specific wavelength) is evaluated, then there is a high probability of a plurality of object classes having this particular characteristic, 25 so that the object classification leads only to an ambiguous result. The uncertainty can be reduced by carrying out evaluation processes at a plurality of wavelengths (in general: a plurality of characteristics). 30 According to the preferred refinement of the invention, however, as already described above, fuzzy sets which vary over time are defined for objects with characteristics which vary over time, with each of the 35 fuzzy sets corresponding to one combination of the DLR - 17 AP 22/103 is possible to considerably increase the certainty in the classification of objects which are stationary but which vary over time (in particular landscape objects), even though a small number of individual fuzzies should 5 be sufficient for precisely that investigation time period in each of the sets. The expression "a fuzzy set" also means a refinement in which a single fuzzy defines probabilities for one or 10 more continuous spectral ranges, and/or is defined as a field of probability values for more than a single point in the radiation spectrum. A fuzzy such as this contains comparable information to a set of monospectral fuzzies. 15 Using the example of a field of wheat, Figure 3 shows, schematically, how the phenological behavior (varying over the course of the year) of a landscape object is transferred to a "land usage fuzzy". The temporal 20 behavior of different cultivated landscape objects can be mapped by landscape usage characteristics (in this case: for example by the described combination of fuzzy sets) . The expression "land usage fuzzy" may thus be understood as meaning a combination of fuzzy sets such 25 as this, and the landscape usage fuzzies are logically linked with the fuzzies from the spectral signatures to form a three-dimensional fuzzy set. Various development and/or processing stages of the 30 field of wheat are plotted against the time t in a row a in Figure 3. The illustrated time axis extends over one calendar year. Depending on the time of year, overall probabilities can be defined for specific development and/or processing stages. Figure 3 35 illustrates overall probabilities for two different DLR - 18 AP 22/103 the middle of the year). The overall probabilities may be regarded, for example, as a weighted and normalized sum of the probabilities for individual object characteristics. 5 Each of the land usage fuzzies "ground" and "wheat plants" is a trapezoidal overall probability function whose base is the time axis (taking into account the fact that the expected time development at the end of 10 the year is continued again by the expected time development at the start of the year) . For example, the flanks of the trapezoidal functions overlap in the spring and in the autumn, that is to say for probability values between 0 and 1. Strictly speaking, 15 the change from one land usage class to another takes place in these time periods, in which case each individual one could then no longer be assigned the probability 1. 20 For example, the phenological profile of the land usage class "wheat" is governed by a range of factors (for example climate - precipitation, temperature) which are themselves highly variable over time. This effect must be taken into account in an appropriate manner in 25 fuzzies which vary over time. Even if the assignment probability over time for wheat and ground in Figure 3 has a profile similar to a fuzzy, then this should provide only the time transition from idealized "ground fuzzies" analogously to 2a) to 2d) into idealized 30 "wheat fuzzies". However, a three-dimensional fuzzy set in fact makes it possible to describe different phases and precisely such as these, that is to say a three dimensional fuzzy set for the land usage class "wheat" also includes the fact that it contains the "(only) 35 ground" phase at specific times.

DLR - 19 AP 22/103 defined for one time period (in the example: 1 year) By way of example, a distinction could be drawn in the "ground" phase between "uncovered ground" and "snow-covered ground", and a distinction could be drawn 5 in the "wheat" phase between "early development of the plants" and "ripe plants with fruits formed". However, the invention allows an even more general formulation for the evaluation of the rate of change of the radiation characteristics over time using a plurality 10 of three-dimensional fuzzy sets, with each of the fuzzy sets describing one expected state of the radiation characteristics that occurs at one time. In addition to the fuzzies of the spectral signatures, 15 fuzzies can be taken into account for other image features (for example texture or compactness measures in the example in Figure 2b to Figure 2d). In addition to land usage, land coverage is taken into 20 account for identification and/or classification. Landscape objects are covered at times by clouds, snow and/or similar phenomena. This covering is identified by land coverage classes. 25 Deciduous woods, coniferous woods and/or other uncultivated landscape objects can optionally be classified as land covering and/or land usage. The type of classification depends on the desired evaluation results. 30 The appearance of a landscape object can change over a scene in phenological appearance, because of the geographical position. Changes in the state of the sun also lead to a change in the appearance of one and the 35 same object, governed by the angle-dependent reflection DLR - 20 AP 22/103 usage and by natural variability over a year. The temporal behavior of the land usage is thus continuously matched to the actually occurring conditions. For example, matching to climatological 5 changes and/or to growth conditions are/is feasible. The results of the identification and/or classification are for this purpose fed back to a land usage classification. 10 The change in the appearance of a landscape object over the year can also be recorded in a time series, and can thus contribute to estimation of the usage, of the usage intensity and thus to the classification of the respective landscape object. If different years and 15 different geographical zones are considered (for example southern Spain and northern Germany) , time shifts can occur in the phenological profile of the vegetation periods because of climatic changes between the years. 20 Figure 4 shows, schematically, a processor for automated identification and/or classification of remote sensing data 1 in land usage classes 7 comprising the modules "monotemporal land coverage 25 classification (MLBK)" 2, "multitemporal data reconstruction (MDR)" 3, "multitemporal land usage classification (MLNK)" 4 and "determination of the classification accuracy (EKG)" 5, as well as a databank 6. The module MLBK 2 is used for identification of land 30 coverage classes within a scene. Land coverage classes are, for example, clouds, mist, water, vegetation, ground surfaces or sealed surfaces. The process elements "texture analysis" 21, "spectral analysis" 22, "automatic segmentation" 23, "land coverage 35 classification" 24 and "object identification" 25 are - - - - - - - - *-. - I - -I-- - - '1 - AT TZ 1V ) rl'- - -- A-Il - M1ThP DLR - 21 AP 22/103 areas and to the identification of non-varying and varying data areas. The module MDR 3 comprises the processor elements "time/spatial data reconstruction" 31 and "quality control" 32. The module MLNK 4 is used 5 for assignment of the usage intensity to land coverage and/or land usage objects, and thus for identification of the land usage classes. This includes the processor elements "land usage classification" 41 and "matching of the temporal behavior" of the land usage 42. 10 In addition to the actual classification, quality control can be used to determine the classification accuracy. The module EKG 5 is provided for this purpose. The module EKG 5 can initiate a 15 reclassification process iteratively in order to allow the classification result to also be improved retrospectively/iteratively. The modules 2-5 may be provided in a common processor 20 and/or in separate units. They access the databank 6 with methods and parameters indirectly and/or directly. Data from different sensors and/or with a different geometrical resolution is subjected to atmospheric 25 correction and/or is geo-referenced by means of modules and/or methods which are not illustrated. The preprocessed remote sensing data 1 is multispectral data. 30 The result of the evaluation of the remote sensing data 1 is a multitemporal data record of classified images of the land coverage and one or more images of the land usage in so-called land usage classes 7. Furthermore, if there are a sufficient number of multitemporal 35 support points, the method is also used for short-term DLR - 22 AP 22/103 usage and "extrapolation" of the behavior of land usages. This then additionally results in landscape predictions 8. This provides the capability to predict the annual profile of the phenological development of 5 vegetation in the short term. The modules 2 to 5 and the associated process elements will be described in detail in the following text. 10 The module MLBK 2 comprises five process elements. The "texture analysis" 21 and "spectral analysis" 22 are applied to the input data 1 in parallel and are used for evaluation of each of the spectral information items given for one pixel, and for the texture in a 15 defined pixel environment. Both process elements are constructed identically in terms of the integrated sub-modules. This is followed by the "segmentation" 23, the actual "land coverage classification" 24 and the "object identification" 25 of classified image objects. 20 The input data for the spectral analysis 21 is the individual multispectral remote sensing data 1 from a times series. This may comprise not only individual scenes but also composites (weekly or 10-day 25 composites) . First of all, land coverage classes that are certain are identified. Land coverage classes that are certain are classes whose characteristic does not change or changes only minimally spectrally and/or spatially. The identified land coverage classes are 30 identified as objects. The land coverage classes cloud, mist, snow, water, uncovered ground and vegetation are preferably identified. In a similar way to the "spectral analysis" process 35 element 21, the properties which are characteristic of DLR - 23 AP 22/103 analysis" 22. The characteristic properties are textures which are determined on the basis of homogeneity criteria. Ten homogeneity classes are preferably identified, graduated on the basis of a 5 statistical criterion. The input data for the "segmentation" 23 is the results of the "spectral analysis" 21 and of the "texture analysis" 22. Since, even if the numbers of spectral 10 and texture classes are the same, it cannot be assumed that the results of the previously described process elements will be coincident, the two information items are combined in accordance with defined rules, thus resulting in the creation of new segments. These 15 segments are dealt with as image objects from then on. In the actual "land coverage classification" 24, the results of the "segmentation" 23 are compared with preferably predefined fuzzies or else threshold values 20 for the various land coverage classes. The association with the individual classes then takes place on the basis of the predetermined class model. One additionally accepted generalization of the 25 classification results can be used to reduce the "salt and pepper" effect, in order to eliminate individual pixels and to reduce the number of identifiable objects. 30 A class identifier is allocated to each identified image object by the process element "object identification" 24. For multitemporal evaluation, an object identifier is also allocated within the class, in order to track the process of objects falling apart 35 and growing together.

DLR - 24 AP 22/103 "time/spatial data reconstruction" 31 and the subsequent "quality control" 32. The results of the land coverage classification for all 5 the remote sensing time periods are investigated in the "time/spatial data reconstruction" 31. This identifies pixels that are contaminated with clouds, mist and temporarily with snow in the time series. After this, the spectral profile of the contaminated pixels for the 10 most probable land usage class is estimated by using interpolation methods to calculate the most probable spectral profile, and by inserting this into the data time series. The interpolation process is preferably carried out by time series analysis and geostatistical 15 methods. The measured pixels, that is to say the pixels which have been identified as not being "contaminated", are used as support points for the interpolation processes. The results of the two interpolation processes mentioned above are compared with one 20 another. If the two values differ by more than a permissible tolerance for the missing point under consideration, then the results are initially rejected, and the next missing point is processed. This procedure is carried out for one time period until no more 25 improvement can be achieved. The next data record is then processed in the same way. An iterative reconstruction attempt (iteration step) for the time series is carried out only once all of the 30 other data records have been completed as well as possible. The iteration of the time and spatial interpolation process is continued until the results from both calculations are in the confidence interval and can be used, after averaging, as correct values. 35 DLR - 25 AP 22/103 indicates the percentage of missing data points in the time series of a pixel can be used for result error estimation. The greater the number of missing data points per pixel, the greater is the probability of 5 incorrect loading of the resultant time series. The method is used to fill data gaps, without changing existing values. The module therefore does not correct the measured values, but assumes that they are correct 10 and plausible, and leads to a data record without any gaps as the basis for statistical calculations. Overall, this process step produces a remote sensing data time series which is filled over to cover an area, that is to say it is reconstructed and geo-referenced, 15 as well as an error probability image. Since errors can occur in the data reconstruction despite the iterative procedure in the process element 31, the result is subjected to a "quality control" 32. 20 For this purpose, once the values of all of the pixels which have been identified as being contaminated have been completely recalculated in the entire time series, a plausibility test is carried out over the entire time series. In this case, those pixels which have not 25 passed the plausibility test or for which there were not a sufficient number of support points are marked as being blocked for the rest of the data evaluation. The module MLNK 4 uses the successive process elements 30 "land usage classification" 41 and "matching of the temporal behavior" 42 of the land usage to process the reconstructed results from the previous module MDR 3. The remote sensing data time series that is created is subjected to the classification process, based on the 35 derived n-dimensional fuzzy sets.

DLR - 26 AP 22/103 classification" 41 is the reconstructed and interpolated time series from the module MDR 3. The spectral signature, the spectral class and the texture class are thus available as a function of time for each 5 pixel. The fuzzy spectral functions for the various model classes of land usage as well as land coverage are mapped in the associated databank 6. By way of example, all cultivated landscape objects such as grassland, cereals, fields, weather phenomena such as 10 snow, clouds and mist as well as uncultivated landscape objects such as woods are mapped as land coverage classes for land usage. The fuzzies for the spectral areas A, B and C chosen by way of example here are entered for each model class in accordance with the 15 schematic fuzzy functions illustrated in Figure 2. Additional spectral areas may be added for hyperspectral remote sensing data. In addition, the databank 6 contains fuzzy functions 20 for the temporal behavior of classes of land usage. These include, for example, the phenological development of the model classes as shown in Figure 3. Examples of these model classes are summer and winter cereal crops, or else different types of deciduous 25 woods, which can be separated only by their phenology. The process element "time matching of the temporal behavior of land usage" 42 is required, for example, in order to adapt the phenology of a land usage class on 30 the basis of the current climatic development in a specific year, and thus its dynamism. For example, the phenology of maize is dependent on where the maize is grown. The start of the sowing process depends, for example, on geographical and/or climatic conditions, as 35 well as on special land features. One possible way to DLR - 27 AP 22/103 surface temperature as measured by the remote sensing sensors. This means that the process element 42 is used to fit 5 the respective individual scene into the overall context of the annual cycle, so that it is also possible to estimate a subsequent data record in advance (prognosis). 10 All of the quality products that have occurred during the processing are evaluated in the module EKG 5 in order on the one hand to check the capability to associate pixels with one class, and on the other hand also to provide an overall error image as additional 15 data for classification. Quality products are, for example, the error probability image from the "time/spatial data reconstruction" 31 or the "pixel marked as blocked" data record from the "quality control" 32. 20 The modules 2-5 can access the method and parameter databank 6, which is used to store, for example, the training spectra for land coverage classes (for example clouds, mist, water, vegetation, uncovered ground), 25 model classes for usage, spectra profile and annual usage profiles and intensities, and phenologies for individual objects. The databank 6 may also be used to store individual 30 cases and to reassess them in order to achieve an improvement in the classification results.

Claims (13)

1. A method for at least partially automated evaluation of remote sensing data, with objects being 5 identified and/or classified on the basis of image features, by means of identification parameters, characterized in that the remote sensing data is in the form of a multispectral remote sensing data time series, and the 10 identification parameters are determined multispectrally and multitemporally from the remote sensing data.

2. The method as claimed in claim 1, characterized in 15 that the remote sensing data (1) is reconstructed in the time/space domain.

3. The method as claimed in claim 1 or 2, characterized in that 20 identification parameters are defined at least depending on spectral signatures (11-17) and a temporal land usage behavior.

4. The method as claimed in claim 3, characterized in 25 that identification parameters are reclassified iteratively, with at least the temporal dependency of land usage characteristics being reclassified. 30 5. The method as claimed in one of claims 3 or 4, characterized in that a time-of-year profile is documented in a data time series, and land usage characteristics are adapted to the recorded time-of-year profile. 35 DLR - 29 AP 22/103 changes in the land usage characteristics which are caused by the time of year are predicted.

7. The method as claimed in one of the preceding 5 claims, characterized in that an object is identified and/or classified on the basis of n-dimensional fuzzy functions.

8. The method as claimed in one of the preceding 10 claims, characterized in that spectral probabilities are formed from a spectral signature of an object as functions of an image feature, and temporal probabilities for a specific time development stage and/or processing stage of the object are formed, and 15 in that other objects are identified and/or classified using the spectral probabilities and the temporal probabilities.

9. The method as claimed in one of the preceding 20 claims, characterized in that a plurality of spectral signatures of an object are evaluated, with each of the spectral signatures describing radiation characteristics of the object at more than one radiation wavelength, in that fuzzies are used in each 25 case when evaluating of the individual spectral signatures, with the fuzzies associating a weighting measure with values of the radiation characteristics or values of a variable derived from them, in that the weighting measures of the radiation characteristics of 30 the object are found by evaluation, and in that the weighting measures that have been found are used to determine whether the object is associated with a type of object which corresponds to the fuzzies. 35 10. The method as claimed in the preceding claim, with DLR - 30 AP 22/103 which differs from the other spectral signatures, or within a different time period.

11. An apparatus for at least partially automated 5 evaluation of remote sensing data, by means of identification parameters, comprising at least one databank, in which case at least the identification parameters can be stored in the databank, and at least one processor, in which case objects can be identified 10 and/or classified by the processor on the basis of image features, characterized in that the apparatus is designed to determine the identification parameters multispectrally and 15 multitemporally from the remote sensing data which is in the form of a multispectral remote sensing data time series.

12. The apparatus as claimed in claim 9, characterized 20 in that the processor has at least one module (3) for multitemporal data reconstruction MDR, in which case the remote sensing data (1) can be reconstructed in the time/space domain by the module (3). 25 13. The apparatus as claimed in claim 9 or 10, characterized in that identification parameters are defined at least as a function of spectral signatures (11-17) and a temporal land usage behavior. 30

14. The apparatus as claimed in claim 11, characterized in that the processor has a module (5) for determination of the classification accuracy EKG and identification parameters can be reclassified 35 iteratively by the module (5), in which case at least DLR - 31 AP 22/103

15. The apparatus as claimed in one of claims 11 or 12, characterized in that a time-of-year profile is documented in a data time series, and land usage 5 characteristics can be adapted to the recorded time-of year profile.

16. The apparatus as claimed in one of claims 11 to 13, characterized 10 in that changes in the land usage characteristics which are dependent on the time of year can be predicted.

17. The apparatus as claimed in one of claims 9 to 14, characterized 15 in that identification parameters are defined as n-dimensional fuzzy functions.