Method for the computer-aided generation of a data-driven model for the computer-aided processing of digital SAR raw data

Abstract

The invention relates to a method for the computer-aided generation of a data-driven model for the processing of SAR raw data, which comprises the backscattered radar echoes of radar pulses (RP), wherein the radar pulses (RP) were emitted by one or more transmitters of at least one object (SA) and the radar echoes were received by one or more receivers of the at least one object (SA), wherein the SAR raw data contains a raw data set (RD) for each receiver consisting of a plurality of data samples (SP), wherein, in the course of processing the SAR raw data, data blocks (DB) of data samples (SP) of the raw data set (RD) with a bit rate (BR) assigned to the respective data block (DB) are selected using bit rate information (ŷ1, ..., ŷ m ) are quantized, where the bitrate information (ŷ1, ..., ŷ m) specifies the assigned bit rate (BR) for each data block (DB) and the data-driven model for determining the bit rate information (ŷ1, ..., ŷ m ) is provided for, the procedure comprising the following steps: - Generation of training data (TD) using a rule (RU) T1 , ..., RU Tn ); - Training a neural network (NN) with the training data (TD) based on an optimization goal that includes a first sub-goal and a second sub-goal, wherein, according to the first sub-goal, the deviation between the bit rate information (ŷ1, ..., ŷ) output by the neural network (NN) is to be determined. m ) from the training data (TD) is as low as possible and according to the second sub-goal (ŷ1, ..., ŷ m ) the deviation of the respective bitrates (BR) of the bitrate information (ŷ1, ..., ŷ) output by the neural network (NN) m ) from a given bitrate value (R1, ..., R m) is as low as possible.

Description

[0001] The invention relates to a method for the computer-aided generation of a data-driven model for the computer-aided processing of digital SAR raw data, as well as a method and a device for the computer-aided processing of digital SAR raw data, a computer program product, and a computer program.

[0002] SAR systems (SAR = Synthetic Aperture Radar) enable the generation of images of the Earth's surface from the echo of radar pulses emitted by a moving object, particularly an aircraft such as a satellite. SAR data processing utilizes the fact that the object's movement detects corresponding points on the Earth's surface at varying distances from the object, resulting in a frequency shift in the radar echo. Through appropriate processing, a digital image of the Earth's surface can be calculated from this data.

[0003] Since capturing radar echoes for a multitude of different positions in the object's direction of movement (also known as azimuth positions) generates a large amount of raw data, it is a known technique to quantize the SAR raw data after digitization. This reduces the storage space required for the raw data during its temporary storage on the moving object and necessitates a lower data rate (bits per sample) when transmitting the raw data to a ground station on the Earth's surface. This is important because the amount of data on the moving object should be kept as small as possible, as typical space-based SAR missions would otherwise experience bottlenecks due to limited onboard storage capacity and data download capabilities.

[0004] The state of the art for quantizing SAR raw data is known as BAQ quantization (BAQ = Block Adaptive Quantization). Generally, BAQ quantization is characterized by dividing the digital data samples of the SAR raw data into data blocks. Within each data block, the corresponding quantization levels are determined based on one or more parameters derived from the values ​​of the digital data samples in that block. These parameters can include, for example, the total or average power of the data samples within a data block, allowing the quantization levels to be placed within a range where most of the data sample values ​​lie.

[0005] In publications [1] and [2], conventional BAQ quantization methods are described in which the same bit rate (i.e., number of bits per quantized data sample) is used for quantizing all data blocks.

[0006] From publication [3], a dynamic FDBA quantization (FDBA = Flexible Dynamic Block Adaptive Quantization) for SAR raw data is known. In this method, the bit rate of the quantization of the respective data blocks is dynamically determined depending on the signal-to-noise ratio of the data samples within a data block.

[0007] From publication [4], an adaptive BAQ quantization for SAR raw data is known, in which the bit rate of the quantization of the respective data blocks (i.e., the BAQ blocks with typically about 100 samples) is optimized with respect to parameters obtained from SAR images of the raw data or from further processed SAR products (e.g., SAR interferograms). This method requires a priori information about the local distribution of backscatter signals for the currently acquired section of the Earth's surface.

[0008] BAQ quantization with adaptive bitrate according to documents [3] and [4] has the disadvantage that larger amounts of data are required for the rules governing the bitrates (e.g., in the form of lookup tables) or for corresponding a priori information. This is particularly problematic when performing quantization on the moving object that captures the raw SAR data, as larger amounts of data must be transferred to or stored on the object. Furthermore, the use of lookup tables and a priori information leads to increased computational complexity for the quantization.

[0009] The publication [5] discloses a method for the computer-aided generation of a data-driven model in the form of a neural network that outputs the bit rates used for quantization of the data blocks of a raw SAR dataset. The neural network is trained using training data based on a rule that depends on the values ​​of at least one parameter, which are derived from previously known raw data sets of the training data. The values ​​of the at least one parameter are, for example, quality values ​​for a SAR product obtained from SAR processing. The neural network does not consider any criteria regarding the bit rate requirements of the quantized raw SAR data sets.

[0010] The object of the invention is to extend a BAQ quantization based on a data-driven model with adaptive bit rates of the data blocks compared to the prior art.

[0011] This problem is solved by the method for computer-aided generation of a data-driven model according to claim 1, the method for computer-aided processing of digital SAR raw data according to claim 11, the device for computer-aided processing of digital SAR raw data according to claim 15, the computer program product according to claim 16, and the computer program according to claim 17. Further developments of the invention are defined in the dependent claims.

[0012] The invention according to the method of claim 1 is characterized in that a data-driven model in the form of a neural network is generated, which can then be used to dynamically determine appropriate bit rates for quantizing data blocks in raw digital SAR data. In other words, the data-driven model is provided for processing raw digital SAR data. Such raw SAR data comprises, in a manner known per se, the backscattered radar echoes of radar pulses, the backscattering typically occurring at the Earth's surface. The radar pulses were emitted by one or more transmitters of at least one object moving in a first direction. This first direction is also referred to as the azimuth direction. Preferably, the at least one object is an aircraft, in particular one or more satellites.

[0013] The radar echoes contained in the SAR raw data were received by one or more receivers of the at least one object. For each receiver, the SAR raw data contains a raw data set consisting of numerous sequential data samples, each associated with a position in the first direction (azimuth direction) and a position in a second direction, also known as the range direction. This range direction extends perpendicular to the azimuth direction. During the processing of the SAR raw data, data blocks (also called BAQ raw data blocks) from sequential data samples of the raw data set are quantized separately using bit rate information assigned to each data block. This bit rate determines the number of bits per quantized data sample in the respective data block.The bitrate information specifies the bitrate to be used for quantizing each data block. The bitrate can be an integer or, optionally, a non-integer value. A non-integer value represents the arithmetic mean of the bitrates of the quantized data samples contained within a given data block. The data-driven model generated using the method described here is intended for determining the bitrate information for the raw data set.

[0014] The quantization of the raw data described above is not part of the process for generating the data-driven model. Rather, the data-driven model, generated in the following steps, is intended for use in the quantization described above. However, the above description of the quantization is necessary for understanding the subsequent steps for generating the data-driven model.

[0015] In the process of generating the data-driven model, training data is first created by generating a training vector with one or more training entries for a multitude of different, pre-existing raw datasets. These raw datasets are structured similarly to the raw datasets described above and are divided into data blocks in the same way. Each training entry contains bitrate information, specifying the bitrates for data blocks within the respective pre-existing raw dataset. The bitrate information of each training entry corresponds to the bitrate information defined above; that is, it specifies an associated bitrate for quantization for each data block of the respective pre-existing raw dataset. The bitrate can be an integer or, if necessary, a non-integer.The bitrate information is generated based on a rule assigned to the respective training entry from a number of different rules, wherein the assigned rule is linked to at least one parameter and depends on values ​​of the at least one parameter obtained from at least the respective previously known raw data set. Preferably, the values ​​of the at least one parameter are quality values ​​for a respective previously known raw data set, in particular quality values ​​for the respective data blocks of the respective previously known raw data set, or quality values ​​for a SAR product (e.g., SAR image, SAR interferogram, and the like) determined from at least the respective previously known raw data set, in particular quality values ​​for pixels or sections of the SAR product. In a variant of the invention, it may optionally be possible to...The training vector should contain only one training entry, so that only one rule is considered in the training data.

[0016] Known rules can be used to generate the bitrate information for each training entry. In particular, the rules described in publications [3] and [4] can be used. However, other suitable rules for generating the bitrate information can also be defined. The values ​​of at least one parameter are preferably defined for each data block of the corresponding known raw data set and thereby determine the bitrate of the quantization of the respective data block.

[0017] The training data generated from the previously known raw data sets, in the form of corresponding training vectors, are used in a next step to train a neural network. The trained neural network represents the data-driven model generated according to the invention. Neural networks are known per se and are therefore not described in detail. A neural network contains a plurality of layers of neurons that serve to determine an output at an output as a function of an input at an input. At least one input layer of neurons represents the input of the neural network, and at least one output layer of neurons represents the output of the neural network.Between the at least one input layer and the at least one output layer are one or more hidden layers of neurons, which convert the input at the input of the neural network into a corresponding output at the output of the neural network.

[0018] According to the invention, the neural network comprises an input for a raw data set and an output for a bitrate vector for the input raw data set, wherein the bitrate vector comprises one or more bitrate entries, each bitrate entry containing bitrate information that specifies bitrates for data blocks of the input raw data set. Each bitrate entry is associated with a feature set from a number of feature sets (i.e., from one or more feature sets), wherein each feature set from the number of feature sets includes a first weight, a second weight, a rule from the number of rules, and a bitrate criterion in the form of a predetermined bitrate value or a predetermined bitrate interval.

[0019] According to the invention, the neural network is trained using the training data based on supervised learning with an optimization goal that comprises a first sub-goal and a second sub-goal for each feature set, wherein the first sub-goal is weighted with the first weight of the respective feature set and the second sub-goal is weighted with the second weight of the respective feature set. Corresponding methods for training a neural network based on supervised learning (e.g., using error backpropagation) are known to those skilled in the art. The corresponding sub-goals can, for example, be modeled as components of a cost function, which is to be minimized during the neural network's learning process.

[0020] According to the invention, in accordance with the first sub-goal, for the respective previously known raw data sets that are input into the neural network, the bit rate information of the bit rate entry, to which the respective feature set in the output bit rate vector is assigned, output by the neural network deviates as little as possible from the bit rate information of the training entry, to which the rule of the respective feature set in the training vector for the respective previously known raw data set is assigned.

[0021] Furthermore, in the inventive method according to the second sub-goal, for the respective previously known raw data sets that are entered into the neural network, the respective bitrates of the bitrate information of the bitrate entry to which the respective feature set in the output bitrate vector is assigned deviate as little as possible from the predetermined bitrate value or the predetermined bitrate interval of the respective feature set.

[0022] In the inventive method, during quantization performed with the generated data-driven model, in addition to ensuring good agreement between the outputs of the neural network and the corresponding training data, a requirement for the bit rates of the quantized data blocks is also taken into account. In this way, limitations regarding the data rate of the transmission path of the quantized SAR raw data, e.g., to a ground station, can also be considered during quantization.

[0023] In one embodiment of the invention, only one set of features consisting of a first weight, a second weight, a rule, and a bitrate criterion is considered when training the neural network. In this case, the neural network can only generate bitrates for one possible quantization. In another embodiment, several sets of features are considered when training the neural network, which differ alternately in at least one of the parameters from the first weight, the second weight, a rule, and a bitrate criterion. This has the advantage that the neural network outputs bitrates for several different quantizations, and the appropriate quantization can be used depending on the requirements.

[0024] In a preferred embodiment, the optimization goal defined in the inventive method for training the neural network is represented by finding an extremum, and preferably the minimum, of a cost function. The cost function is a vector consisting of a number of function entries, which may optionally contain only a single function entry. Each function entry is correlated with a feature set of a certain number of feature sets and represents a weighted sum comprising a first term weighted with the first weight to represent the first sub-goal and a second term weighted with the second weight to represent the second sub-goal.

[0025] In each function entry, the first term is a difference measure that, for a given pre-existing raw data set, measures the difference between the bitrate information of the training entry (to which the rule of the feature set correlated with that function entry is assigned and which is contained in the training vector for that pre-existing raw data set) and the bitrate information of the bitrate entry (to which the feature set correlated with that function entry is assigned and which is contained in the bitrate vector that the neural network outputs for that pre-existing raw data set). The difference measure of the first term can, for example, be the mean error or the mean squared error between the bitrates of the bitrate information of the training entry and the bitrates of the bitrate information output by the neural network.

[0026] In contrast, the second term in each function entry is a difference measure. For a given predefined raw data set, this measure represents the difference between the specified bitrate value or bitrate interval and the respective bitrates of the bitrate information in the entry to which the feature set correlated with that function entry is assigned. This entry is contained in the bitrate vector that the neural network outputs for that predefined raw data set. The difference measure of the second term can, for example, be the mean error or the mean squared error between the specified bitrate value or bitrate interval and the respective bitrates of the bitrate information in the corresponding entry. If a predefined bitrate interval is used, the error is zero if the respective bitrate lies within that interval.If this is not the case, the error represents the deviation of the respective bit rate from the bit rate of the nearest interval boundary.

[0027] The cost function above is a multidimensional vector if multiple sets of features (and thus function entries) are considered. Furthermore, the cost function varies depending on the raw data set used. The search for the extremum of the cost function can be performed iteratively across the known raw data sets, with the weights of the neural network being varied in each iteration step so that the cost function reaches an extremum for the respective known raw data set.

[0028] In a particularly preferred embodiment of the above, the optimization goal is represented by finding the minimum of the following cost function (L): L=[α1⋮αm]⋅[(y1−y^1)2⋮(ym−y^m)2]+[β1⋮βm]⋅[(R1−y^1)2⋮(Rm−y^m)2] where [β1⋮βm]=([1⋮1]−[α1⋮αm]) where L is the cost function; where m is the number of function entries; where α r (r = 1, ..., m) is the first weight for the feature set correlated with the respective function entry; where y r (r = 1, ..., m) is the bit rate information of the training entry to which the rule of the feature set correlated with the respective function entry is assigned and which is contained in the training vector for the respective previously known raw data set; where ŷ r (r = 1, ..., M) is the bit rate information of the bit rate entry to which the feature set correlated with the respective function entry is assigned and which is contained in the bit rate vector that the neural network outputs at its output for the respective previously known raw data set at its input; where β r(r = 1, ..., M) is the second weight for the feature set correlated with the respective function entry; where R r (r = 1, ..., M) is the specified bit rate value for the feature set correlated with the respective function entry, where the specified bit rate value is identical for all bit rates of the bit rate information of the bit rate entry to which the feature set correlated with the respective function entry is assigned and which is contained in the bit rate vector that the neural network outputs at its output for the respective pre-known raw data set at its input.

[0029] In a preferred embodiment of the inventive method for generating a data-driven model, SAR processing is performed for each previously known raw data set during the generation of the training data for at least one of the number of requirements. This processing involves, in a manner known per se, the determination of one or more SAR images from at least the respective previously known raw data set (i.e., from the respective previously known raw data set and optionally one or more further raw data sets). The values ​​of the at least one parameter associated with the at least one requirement are derived from the information provided by the SAR processing. Preferably, the values ​​of the at least one parameter are quality values ​​for a SAR product obtained from the SAR processing (e.g.,SAR interferogram, elevation model, classification map and the like), in particular quality values ​​for pixels or sections from the SAR product.

[0030] A SAR image is understood to be two-dimensional image information that specifies the amplitude and phase for a large number of pixels in the captured scene (preferably a section of the Earth's surface). If only one known raw data set is used in SAR processing, only one SAR image is obtained for that specific raw data set. The resulting SAR image(s) can represent the outcome of SAR processing and thus a SAR product, but they can also be an intermediate step in this SAR processing. For example, SAR interferograms can be generated from corresponding SAR images as SAR products in a known manner, whereby in this case, in addition to the respective known raw data set, at least one other known raw data set is also used in the SAR processing of the respective known raw data set.

[0031] The values ​​of the at least one parameter associated with the at least one rule are preferably linked to corresponding pixels of a SAR image or a correspondingly processed image (e.g., a SAR intensity image or a SAR interferogram), wherein the pixels can each be linked in a known manner to a corresponding data block in the respective previously known raw data set, the data samples of the linked data block significantly determine the amplitude and phase of the associated pixel. The values ​​of the at least one parameter at a pixel thus determine the bit rate of the quantization of a linked data block in the respective previously known raw data set.

[0032] In one variant of the embodiment of the inventive method described above, the at least one parameter associated with the at least one rule comprises the standard deviation of the distribution of backscattered signals in a section around a pixel of a SAR image, which is determined for the respective previously known raw data set without quantization or after quantization with a constant bit rate (i.e., a fixed bit rate for all data blocks). Standard deviation values ​​are determined for a plurality of pixels and, in particular, for all pixels in the SAR image. As mentioned above, the corresponding pixels can be linked to data blocks in the respective previously known raw data set, so that bit rates for the linked data blocks can be determined depending on the standard deviation values.The parameter of the standard deviation of the distribution of backscattered signals described above is also used in publication [4] to determine the bit rate information.

[0033] In a further preferred embodiment, the at least one parameter associated with the at least one rule comprises the signal-to-quantization noise ratio in a section around a pixel of a SAR image (i.e., the signal-to-quantization noise ratio averaged over the section), wherein the SAR image is determined for the respective previously known raw data set after quantization at a constant bit rate (i.e., a fixed bit rate for all data blocks), and wherein values ​​of the signal-to-quantization noise ratio are determined for a plurality of pixels in several SAR images obtained for the respective previously known raw data set, the several SAR images differing in the bit rate of the quantization of the respective previously known raw data set. The signal-to-quantization noise ratio is a quantity well known to those skilled in the art and is defined, for example, in publication [4].For the signal-to-quantization noise ratio values ​​at different quantizations, a requirement, e.g., in the form of a specified value or range of values, may be defined in the procedure for generating the bit rate information. A corresponding requirement is defined in more detail, for example, in publication [4].

[0034] In a further preferred embodiment, the at least one parameter associated with the at least one rule comprises the coherence loss caused by quantization in a section around a pixel of a SAR interferogram (i.e., the coherence loss averaged over the section), wherein the SAR interferogram is determined from the respective previously known raw data set and at least one further previously known raw data set after quantization at a constant bit rate (i.e., a fixed bit rate for all data blocks), and wherein values ​​of the coherence loss caused by quantization are determined for a plurality of pixels for several SAR interferograms, wherein the several SAR interferograms differ in the bit rate of the quantization of the respective previously known raw data set and the at least one further previously known raw data set.The determination of corresponding SAR interferograms, which describe phase differences between several SAR images, is well known to those skilled in the art. Furthermore, the coherence loss mentioned above is also a well-known quantity, which is described, for example, in publication [4]. For the values ​​of the coherence loss at different quantizations, a requirement, e.g., in the form of a specified value or range of values, can be defined in the procedure for generating the bit rate information. Such a requirement is defined in more detail, for example, in publication [4].

[0035] In a further embodiment, the at least one parameter associated with the at least one rule comprises the standard deviation of the distribution of the phase error caused by quantization in a section around a pixel of a SAR interferogram, which is determined from the respective previously known raw data set and at least one further previously known raw data set after quantization with a constant bit rate (i.e., a fixed bit rate for all data blocks), wherein values ​​of the standard deviation of the distribution of the phase error caused by quantization are determined for a plurality of pixels for several SAR interferograms, wherein the several SAR interferograms differ in the bit rate of the quantization of the respective previously known raw data set and the at least one further previously known raw data set.

[0036] The phase error caused by quantization is a well-known quantity and is defined in more detail, for example, in publication [4]. For the standard deviation values ​​of the distribution of the phase error caused by quantization at different quantizations, a requirement, e.g., in the form of a specified value or range of values, may be defined in the procedure for generating the bit rate information. Such a requirement is defined in more detail, for example, in publication [4].

[0037] In a further embodiment of the method according to the invention, for at least one of the number of rules, the values ​​of the at least one parameter associated with the at least one rule are obtained directly from the data samples of the data blocks of the respective previously known raw data set, without SAR processing being performed to obtain the at least one parameter, wherein such SAR processing includes the determination of SAR images from at least the respective previously known raw data set. For example, the at least one parameter can be the mean signal-to-noise ratio of the data samples in a data block in the respective previously known raw data set, wherein signal-to-noise ratio values ​​are determined for each data block in the respective previously known raw data set. This parameter is also used in the method of publication [3].

[0038] According to the invention, any neural network can be used to determine the bit rate information. The use of a so-called convolutional neural network is particularly practical, as it is very well suited for processing image information. The structure of convolutional neural networks is well known to those skilled in the art and is therefore not described in detail. A convolutional neural network contains so-called convolutional layers, which are provided at the input of the neural network, and usually also include so-called polling layers for data reduction. The convolutional layers perform a convolution of the input data using a filter kernel. By means of this convolution, structures of corresponding SAR images can be extracted from the raw data set in advance.

[0039] In another preferred embodiment, the neural network is a deep neural network. The concept of a deep neural network is well known to those skilled in the art. Such a neural network is characterized by a large number of hidden layers and enables particularly good modeling of bit rate information. In a particularly preferred embodiment, the deep neural network is a convolutional neural network.

[0040] In another preferred embodiment, the trained neural network is implemented on one or more FPGAs (FPGA = Field Programmable Gate Array). FPGAs enable very efficient execution of the neural network.

[0041] In addition to the method described above for generating a data-driven model, the invention relates to a method for the computer-aided processing of digital SAR raw data, which comprises the backscattered radar echoes of radar pulses, wherein the radar pulses were emitted by one or more transmitters of at least one object moving in a first direction, and the radar echoes were received by one or more receivers of the at least one object, wherein the SAR raw data for each receiver contains a raw data set consisting of a plurality of temporally successive data samples, each of which is assigned a position in the first direction and a position in a second direction perpendicular to the first direction.During the processing of the SAR raw data, data blocks from successive data samples of the raw data set are quantized separately using bitrate information assigned to each data block. This bitrate determines the number of bits per quantized data sample in the respective data block. The bitrate information specifies the assigned bitrate for each data block.

[0042] In this process, the bitrate information corresponds to a bitrate entry generated by a data-driven model created using the computer-aided generation method described above. This means that the generation of the data-driven model is not part of the inventive method for computer-aided processing of the SAR raw data (i.e., not part of the quantization). If the data-driven model outputs or generates a bitrate vector with multiple bitrate entries, the bitrate information of one of these entries is used to quantize the corresponding raw data set.If the bitrate information used to quantize the raw data set contains a non-integer bitrate for one or more data blocks, the bitrates for the individual quantized data samples of the corresponding data block are determined such that the arithmetic mean of these bitrates corresponds to the non-integer bitrate of the corresponding data block. Such a determination of non-integer bitrates is known, for example, from publication [6].

[0043] Alternatively, the generation of the data-driven model used to generate the bit rate information in the SAR raw data processing procedure is also part of the SAR raw data processing procedure itself. In other words, the bit rate information is generated using a data-driven model that is created during the processing of the SAR raw data using the procedure described above.

[0044] In a preferred embodiment of the inventive method for processing SAR raw data, the quantized data samples obtained by the method are determined on the at least one object, i.e., directly downstream of the acquisition of the SAR raw data. The quantized data samples are then transmitted to a ground station, i.e., a station on the Earth's surface. In this case, the at least one object is preferably at least one aircraft, in particular at least one satellite.

[0045] The invention further relates to a device for processing digital SAR raw data, wherein the device is configured to carry out the inventive method for processing SAR raw data or one or more preferred variants of this method.For this purpose, the device includes a quantizer configured to quantize data blocks from successive data samples of the raw data set separately during computer-aided processing of the SAR raw data. This quantization is performed using bit rate information, which specifies the number of bits per quantized data sample in the respective data block. The bit rate information indicates the assigned bit rate for each data block. The quantizer contains a data-driven model for generating the bit rate information, which was pre-generated using the data-driven model generation method described above. Alternatively, the data-driven model can also be generated by the quantizer itself; that is, the model is not pre-existing in the quantizer but is generated during the SAR data processing.Preferably, the device is provided on the at least one object that includes the transmitter(s) or receiver(s) for capturing the SAR raw data.

[0046] The invention further relates to a computer program product with a program code stored on a machine-readable medium for carrying out the inventive method for the computer-aided generation of a data-driven model or for carrying out the inventive method for the computer-aided processing of digital SAR raw data or for carrying out one or more preferred variants of these methods when the program code is executed on a computer.

[0047] The invention further relates to a computer program with program code for carrying out the inventive method for the computer-aided generation of a data-driven model or for carrying out the inventive method for the computer-aided processing of digital SAR raw data or for carrying out one or more preferred variants of these methods when the program code is executed on a computer.

[0048] Exemplary embodiments of the invention are described in detail below with reference to the accompanying figures.

[0049] They show: Fig. 1 a schematic representation that explains the SAR principle used in the invention; Fig. 2 a schematic representation to illustrate an embodiment of the inventive method for computer-aided processing of digital SAR raw data; and Fig. 3 a schematic representation of an embodiment of a method according to the invention for generating a data-driven model for the method of Fig. 2.

[0050] For better understanding, we will first use the following as an example: Fig. 1. The principle of SAR measurement is explained using radar radiation emitted by a satellite. Instead of a satellite, the radar radiation can also be emitted by any other moving platform, such as an aircraft or a platform moving on the Earth's surface, e.g., on rails. The in Fig. The satellite shown is indicated by the reference symbol SA and moves in the direction of the arrow M shown, which corresponds to the x-direction of the Fig. The x-direction is represented by the Cartesian coordinate system shown in Figure 1. This x-direction is also commonly referred to as the azimuth. The satellite SA, whose altitude above the Earth's surface is represented by the z-coordinate of the coordinate system, continuously transmits radar pulses towards the Earth's surface at a predetermined pulse repetition frequency during its movement. The contour of a currently transmitted radar pulse on the Earth's surface is denoted by RP. The radar echo of each transmitted radar pulse is detected by scanning the radar radiation reflected from the Earth's surface in the y-direction of the Cartesian coordinate system. The y-direction is also commonly referred to as the ground range.

[0051] In the illustrated embodiment, the evaluation only captures information from the radar radiation between the two lines L and L', which are several kilometers apart (e.g., 30 km). This is also indicated by the hatched area within the radar pulse RP. The SAR measurement yields a multitude of analog raw data for a large number of radar pulses and thus for a large number of azimuth positions, each corresponding to a range position. A corresponding sample is represented by a complex signal with an amplitude and a phase.

[0052] The principle of SAR measurement is based on the fact that specific points on the Earth's surface are recorded multiple times from different angles due to the movement of the SA satellite. This is done in Fig. Figure 1 illustrates point P. In the depicted scenario, this point P has the smallest possible distance to satellite SA. As the satellite moves in the direction of M, this distance increases. Before reaching the point shown in Fig. At the position shown in Figure 1, this distance became progressively smaller. Due to the Doppler effect, a frequency shift occurs during the detection of the radar echo. This shift can be appropriately evaluated, ultimately yielding amplitude and phase information for the points on the Earth's surface where the radar pulses are reflected, and thus a pixel of the Earth's surface. The corresponding calculation of pixels of the Earth's surface from raw SAR data is well known to those skilled in the art and is therefore not explained in further detail. The SAR measurement method simulates a larger synthetic aperture, corresponding to the extent of the radar pulse on the Earth's surface, using a multitude of radar pulses from a small-aperture radar transmitter.

[0053] In the scenario of Fig. In configuration 1, both the radar transmitter and the radar echo receiver are located within the satellite SA. With this configuration, two-dimensional SAR images of the Earth's surface can be generated using a known SAR processing technique. In another configuration, the SAR measurement is performed by two synchronously moving satellites. Radar pulses are emitted from a transmitter on one of the satellites, and the corresponding radar echo is received by a receiver on both satellites. In this way, the same scene is captured from different perspectives. The resulting two-dimensional SAR images are then computer-aided interferometrically superimposed to calculate a three-dimensional image of the Earth's surface with corresponding elevation information.Alternatively, the same SAR satellite can fly over the same area on the Earth's surface twice to obtain two SAR images, which are then processed interferometrically.

[0054] During a SAR measurement, the raw SAR data received by the corresponding satellite is first digitized using an analog-to-digital converter (ADC). After quantization, as described below, the data is then transmitted to a ground station for further processing. Often, an ADC that generates 8-bit values ​​from the raw data is used. However, other types of converters can also be employed, generating digital values ​​with longer or shorter bit lengths. The SAR raw data digitized by the ADC constitutes the digital SAR raw data as defined in the patent claims. A raw data set, as defined in the patent claims, corresponds to the (digitized) samples of the digital SAR raw data received by a respective receiver for a given data acquisition, i.e., a predefined area on the Earth's surface traversed by the satellite(s).

[0055] During further processing of the raw digital SAR data at the ground station, the two-dimensional and three-dimensional images of the Earth's surface described above are calculated. However, since the storage resources in the satellite and the data rate for transmitting the raw data to the ground are limited, the raw digital SAR data is quantized before transmission. During quantization, quantization levels based on shorter bit values ​​are assigned to the digital signals. For example, quantization can reduce the original bit rate of 8 bits per data sample to a rate of 2 or 3 bits.

[0056] According to the invention, a data-driven model trained with suitable training data is used to quantize the raw digital SAR data. The following describes how this is achieved using the following examples: Fig. 2. An embodiment of quantization using a data-driven model in the form of a neural network is described.

[0057] Fig. Figure 2 schematically shows a raw data set RD to be quantized in the form of a two-dimensional matrix. This matrix contains a multitude of digital data samples, which are only indicated by hatched squares SP. The entire matrix is ​​filled with such samples. The samples were determined for a multitude of azimuth positions and a multitude of range positions, where in Fig. 2 the azimuth direction is indicated by the x-axis and the range direction by the y-axis of a two-dimensional Cartesian coordinate system.

[0058] As part of the quantization process, the raw data set RD is first divided into a large number of equally sized data blocks (BAQ blocks) DB, whereby in Fig. Figure 2 merely illustrates a data block represented by a hatched rectangle. The data blocks contain several temporally successive samples (SP) in the range direction. BAQ quantization is performed separately for each individual data block, as explained further below. The bit rate (i.e., the number of bits per quantized data sample) is variably determined for each data block depending on features extracted from the data samples of the respective data block. Typically, the assignment of individual data blocks to a corresponding quantization bit rate is explicitly implemented by a rule, for example, in the form of lookup tables that may be stored in the satellite SA and may also contain prior information, such as the local distribution of backscatter signals for the currently acquired section of the Earth's surface.Therefore, it is necessary to upload a large amount of data to the SA satellite, which requires sufficient bandwidth and sufficient storage space on the satellite.

[0059] Within the scope of the invention, the corresponding regulation or rule is no longer implemented explicitly, but rather via a learned, data-driven model, which in the embodiment of the Fig. 2. A neural network (NN) is used. Several different rules are implemented in the neural network (NN), and the user can select which rule should be used for the corresponding quantization. The neural network (NN) was generated beforehand through appropriate training, which is explained below and also forms part of the invention. A cost function (L) (loss function) was used during the training of the neural network (NN), which is also explained below. The goal of training the neural network is to minimize the cost function. Unlike the prior art, the cost function includes a first sub-goal and a second sub-goal. The first sub-goal concerns a good match between the output of the neural network (NN) and the corresponding training data. The second sub-goal is a requirement for the respective bit rates used for quantization.The bit rates should be as close as possible to a predetermined bit rate value, so that the average bit rate of the quantization can also be controlled via the cost function.

[0060] Any known network structure consisting of an input layer, a multitude of hidden layers in between, and an output layer can be used as a neural network. However, it has been shown that a so-called convolutional neural network is particularly suitable for processing raw SAR data. Such networks use convolutional layers, which first subject the input data to convolution with a filter kernel.

[0061] The input layer feeds the raw data set RD to be quantized to the neural network NN, from which the neural network calculates a bitrate vector Ŷ at its output layer using its hidden layers. The bitrate vector comprises several bitrate entries B1, B2, ..., Bm, each representing a bitrate value ŷ1, ŷ2, ..., ŷ m in the form of a bitrate map. Each bitrate map specifies a corresponding bitrate for the quantization of each data block in the raw data set RD. Fig. Figure 2 schematically depicts only the bit rate map for the bit entries B1, B2, and Bm, with only a few entries of the map indicated by hatched squares. Each entry of the map represents a corresponding bit rate BR. The position of each entry of the map is linked, in a manner known per se, to the position of a data block in the raw data RD.

[0062] Each bitrate entry Br has a corresponding feature set FSr (r = 1, ..., m). Each feature set FSr has four quantities in the form of a first weight α. r , of a second weight β r , a corresponding regulation or rule RU r and a given bitrate value R r assigned, whereby the sets of characteristics differ in at least one of these quantities. However, different sets of characteristics can agree with each other for some of these quantities. For example, the corresponding regulation RU r The weights may be identical for some of the feature sets and possibly for all feature sets, provided the feature sets differ in another dimension. Among other things, the first weights α r and the second weights β r included in the cost function L, as described in more detail below. The first weight α rweights the first sub-goal of the cost function L and the second weight β r weights the second sub-goal of the cost function L.

[0063] For the subsequent quantization Q of the corresponding raw data set RD, one of the bitrate entries B1 to Bm is predefined or selected by a user. In the scenario of Fig. 2 is used to quantize the bitrate entry B1, as indicated by a dashed arrow.

[0064] Based on the bitrate information or bitrate map ŷ1 of the bitrate entry B1, the raw data set RD is then quantized Q. During this quantization, each data block DB of the raw data set RD is quantized separately with the corresponding bitrate BR from the bitrate map ŷ1 of the bitrate entry B1. Any known BAQ quantization can be used for quantizing the individual data blocks, such as the BAQ quantization from document [1] or [2].

[0065] After performing the quantization Q, the quantized raw data set QRD is obtained, in which the individual data samples are quantized and thus compressed. Examples are shown in Fig. 2 Some of the quantized data samples are represented by hatched squares and labelled with reference symbol QSP.

[0066] The following will be based on Fig. 3. Training the neural network NN Fig. 2 explained, whereby the training is usually carried out separately from the quantization beforehand on the ground once and then only the trained neural network is uploaded to the satellite for use in quantization.

[0067] According to Fig. Step 3 of the procedure for generating the trained neural network NN first involves generating suitable training data TD in the form of corresponding training vectors Y, each containing training entries T1 to Tn. Subsequently, the actual training or learning of the neural network is carried out (step S4). To generate the corresponding training data, a large number of previously known raw data sets RD' are processed, analogous to the raw data set RD of the Fig. 2. That is, a corresponding pre-existing raw data set RD' contains a multitude of data acquisitions SP' and is divided into data blocks (BAQ blocks) DB' analogous to the structure of the raw data set RD. The combination of a raw data set RD' and a training vector Y for this raw data set constitutes a training data set. The training data TD thus consists of a multitude of training data sets corresponding to the number of pre-existing raw data sets.

[0068] For each of the previously known raw data sets RD', for each of the regulations RU T1 to RU Tn , which are contained in the feature sets FS1 to FSm, training entries T1 to Tn with respective bitrate information y T1 up to y TnST1 to STn are generated in corresponding steps, with only steps ST1, ST2, and STn being shown. The individual bitrate information of the training entries T1 to Tn are in turn bitrate maps, analogous to the bitrate maps of the Fig. 2 are structured and thus define assigned bitrates BR' for the individual data blocks DB' of a respective raw data record RD'. In the scenario of Fig. 3. In steps ST1 to STn, a total of n training entries T1 to Tn are created using appropriate regulations RU. T1 to RU Tn generated, whereby a respective rule of values ​​of at least one parameter P is generated. Ti (i = 1, ..., n) depends on the respective previously known raw data set. Examples of corresponding regulations were explained above and are also defined in the attached claims.

[0069] To illustrate, in Fig. Figure 3 shows, as an example for step ST1, the generation of the training entry T1 based on the procedure or rule described in publication [4], which depends on parameter values ​​in the form of standard deviation values ​​of the distribution of backscattered signals in a section around respective pixels of a SAR image, which is obtained from a corresponding raw digital data set without quantization. In substep S101, a known SAR processing SPR is first carried out, in which the raw data set RD' is subjected to known azimuth and range focusing, resulting in a complex two-dimensional SAR image IM containing a multitude of pixels PI, which are only indicated by a few hatched squares. Each pixel corresponds to a small section of the Earth's surface with a corresponding amplitude and phase.Each pixel can be spatially linked to a data block DB' in the raw data RD' using a method that is known in itself.

[0070] In sub-step S102, values ​​of the parameter P are extracted from the acquired SAR image IM. T1 The standard deviation of the distribution of the backscattered signal in a section around a pixel of the SAR image IM is extracted, with values ​​of this parameter being determined for all pixels in the SAR image. Depending on these parameter values, the rule RU is then applied in substep S103. T1 , which is taken from the printed publication [4], the bitrate information y T1 determined. In other words, the rule specifies the bit rate BR' to be used for a corresponding parameter value of a pixel PI and thus for a linked data block DB'.

[0071] Instead of or in addition to the rule or procedure described in publication [4], any other procedures can be used in the respective steps ST1 to STn, which depend on other or further parameter values ​​of the SAR image or, if applicable, on parameter values ​​resulting from further SAR processing. Examples of such parameter values ​​(signal-to-quantization noise ratio, loss of coherence, standard deviation of the distribution of the phase error caused by quantization) were described above. In a modified embodiment, it is also possible for a corresponding rule to use parameters that are derived directly from the data samples of a respective data block without the interposition of SAR processing (SPR). Such a rule is described, for example, in the aforementioned publication [3].

[0072] As a result of each step ST1 to STn, a training dataset is obtained for each raw data set RD', consisting of a raw data set RD' and a training vector Y with training elements T1 to Tn. The entirety of the training datasets represents, according to... Fig. 3. The training data TD. Subsequently, this training data is used to perform supervised learning in step S2 of the Fig. 3. To perform a training TR of the corresponding, initially untrained neural network using a cost function L. The result is the trained neural network NN, which is then uploaded to the satellite and quantized there according to the procedure of Fig. 2 is used.

[0073] A key aspect of the embodiment described here is that a special cost function is used to train the neural network NN. This cost function models an optimization goal based on a first sub-goal weighted with a first weight and a second sub-goal weighted with a second weight. The optimization goal is to minimize the cost function based on the variables defined above. Fig. 2 and Fig. 3 reads as follows: L=[α1⋮αm]⋅[(y1−y^1)2⋮(ym−y^m)2]+[β1⋮βm]⋅[(R1−y^1)2⋮(Rm−y^m)2] where [β1⋮βm]=([1⋮1]−[α1⋮αm]) where L is the cost function in the form of an m-dimensional vector; where m is the number of function entries (entries of the m-dimensional vector), where each function entry is correlated with one of the feature sets FS1 to FSm; where α r(r = 1, ..., m) is the first weight for the feature set FSr correlated with the respective function entry; where y r (r = 1, ..., m) is the bitrate information of the training entry to which the rule of the feature set FSr correlated with the respective function entry belongs and which is contained in the training vector for the respective raw data set RD'; where ŷ r (r = 1, ..., M) is the bitrate information of the bitrate entry Br, to which the feature set FSr correlated with the respective function entry is assigned and which is contained in the bitrate vector Ŷ that the neural network (NN) outputs for the respective raw data set RD' at its input; where β r (r = 1, ..., M) is the second weight for the feature set FSr correlated with the respective function entry; where R r(r = 1, ..., M) is a predefined bit rate value for the feature set FSr correlated with the respective function entry, where the predefined bit rate value is identical for all bit rates of the bit rate information ŷ r of the bitrate entry Br, to which the feature set FSr correlated with the respective function entry belongs and which is contained in the bitrate vector Ŷ, which the neural network NN outputs at its output for the respective previously known raw data set RD' at its input.

[0074] To find the (global) minimum of the above cost function L within the framework of learning the neural network NN, known numerical methods, such as gradient descent, can be used.

[0075] The embodiments of the invention described above offer several advantages. In particular, for the first time, adaptive BAQ quantization, in which data blocks of the SAR raw data can be quantized at different bit rates, is combined with a neural network that, in addition to the goal of minimizing the error of the bit rates output by the neural network, also takes into account the goal of achieving a predetermined bit rate value. Using the predetermined bit rate value, an average bit rate of the quantized data blocks can be ensured in accordance with the requirements for the data transmission of the quantized raw data. The aforementioned goals can be appropriately weighted using a first and second weight. Furthermore, the neural network can output bit rates for several possible quantizations according to the feature sets described above, so that a quantization suitable for the current circumstances can be selected.It is no longer necessary to pre-store large amounts of data in the form of lookup tables or other prior information in the quantizer. Only the learned neural network needs to be stored in the quantizer. This neural network can be implemented very efficiently using familiar FPGAs (Field Programmable Gate Arrays). Bibliography: [1] El Boustani, A.; Brunham, K.; Kinsner, W.: “A Review of Current Raw SAR Data Compression Techniques”, Electrical and Computer Engineering, Canadian Conference on, 2001, Vol. 2, pages 925 - 930 [2] US 6,255,987 B1 [3] E. Attema, C. Cafforio, M. Gottwald, P. Guccione, AM Guarnieri, F. Rocca, P. Snoeij: “Flexible Dynamic Block Adaptive Quantization for Sentinel-1 SAR Missions,” IEEE Geoscience and Remote Sensing Letters, Vol. 7, No. 4, pages 766-770, October 2010 [4] M. Martone, N. Gollin, P. Rizzoli, G. Krieger: „Performance-Optimized Quantization for SAR and InSAR Applications“, IEEE Transactions on Geoscience and Remote Sensing, Vol. 60, Seiten 1-22, 2022 [5] EP 4 340 227 A1 [6] DE 10 2012 209 113 B4

Claims

Method for the computer-aided generation of a data-driven model for the computer-aided processing of digital SAR raw data comprising the backscattered radar echoes of radar pulses (RP), wherein the radar pulses (RP) were emitted by one or more transmitters of at least one object (SA) moving in a first direction (x), and the radar echoes were received by one or more receivers of the at least one object (SA), wherein the SAR raw data for each receiver contains a raw data set (RD) consisting of a plurality of temporally successive data samples (SP), each of which is assigned a position in the first direction (x) and a position in a second direction (y) perpendicular to the first direction (x), wherein, in the course of processing the SAR raw data, data blocks (DB) of temporally successive data samples (SP) of the raw data set (RD) are processed separately with a bit rate (BR) assigned to the respective data block (DB).which defines the number of bits per quantized data sample (QSP) in the respective data block (DB), are quantized using bit rate information (ŷ1, ..., ŷm), wherein the bit rate information (ŷ1, ..., ŷm) specifies the assigned bit rate (BR) for each data block (DB), and the data-driven model is provided for determining the bit rate information (ŷ1, ..., ŷm), wherein the generation of the data-driven model comprises the following steps: - Generation of training data (TD) by generating a training vector (Y) with one or more training entries (T1, ..., Tn) for each of a variety of different previously known raw data sets (RD'), wherein each training entry (T1, ..., Tn) contains bit rate information (yT1, ..., yTn) which specifies the bit rates (BR') for data blocks (DB') of a respective previously known raw data set. (RD') specifies, where the bitrate information (yT1, ..., yTn) is based on the respective training entry (T1,..., Tn) assigned rule (RUT1, ..., RUTn) is generated from a number of rules (RUT1, ..., RUTn), wherein the assigned rule (RUT1, ..., RUTn) is associated with at least one parameter (PT1, ..., PTn) and depends on values ​​of the at least one parameter (PT1, ..., PTn) obtained from at least the respective previously known raw data set (RD'); - Training a neural network (NN) with the training data (TD), thereby obtaining the generated data-driven model, wherein the neural network (NN) has an input for inputting a raw data set (RD) and an output for outputting a bitrate vector (Y) for the input raw data set (RD), wherein the bitrate vector (Y) comprises one or more bitrate entries (B1, ..., Bm), each bitrate entry (B1, ..., Bm) containing bitrate information (ŷ1, ..., ŷm) contains which bit rates (BR) for data blocks (DB) of the entered raw data set (RD) are specified,wherein each bitrate entry (B1, ..., Bm) is assigned a feature set (FS1, ..., FSm) from a number of feature sets (FS1, ..., FSm), wherein each feature set (FS1, ..., FSm) of the number of feature sets (FS1, ..., FSm) includes a first weight (α1, ..., αm), a second weight (β1, ..., βm), a rule (RUT1, ..., RUTn) from the number of rules (RUT1, ..., RUTn), and a bitrate criterion in the form of a predefined bitrate value (R1, ..., Rm) or a predefined bitrate interval;- wherein the training of the neural network (NN) using the training data (TD) is based on supervised learning with an optimization goal that includes, for each feature set (FS1, ..., FSm), a first subgoal and a second subgoal, wherein the first subgoal with the first weight (α1,..., αm) of the respective feature set and the second sub-goal with the second weight (β1, ..., βm) of the respective feature set (FS1, ..., FSm),where i) according to the first sub-goal, for the respective pre-known raw data sets (RD') that are input into the neural network (NN), the output bitrate information (ŷ1, ..., ŷm) of the bitrate entry (B1, ..., Bm) to which the respective feature set (FS1, ..., FSm) in the output bitrate vector (Y) is assigned, deviates as little as possible from the bitrate information (yT1, ..., yTn) of the training entry (T1, ..., Tn) to which the rule (RUT1, ..., RUTn) of the respective feature set (FS1, ..., FSm) in the training vector (Y) for the respective pre-known raw data set (RD') is assigned; ii) according to the second sub-goal, for the respective pre-known raw data sets (RD') that are input into the neural network (NN), the respective bitrates (BR) of the output bitrate information (ŷ1, ..., ŷm) of the bitrate entry (B1, ..., Bm) to which the respective feature set (FS1, ..., FSm) is assigned in the output bitrate vector (Y),deviate as little as possible from the specified bitrate value (R1, ..., Rm) or the specified bitrate interval of the respective feature set (FS1, ..., FSm). The method according to claim 1, characterized in that the optimization goal is represented by finding an extremum of a cost function (L), wherein the cost function (L) is a vector of a number of function entries, each function entry being correlated with a feature set (FS1, ..., FSm) of the number of feature sets (FS1, ..., FSm) and representing a weighted sum comprising a first term weighted with the first weight (α1, ..., ) to represent the first sub-goal and a second term weighted with the second weight (β1, ..., βm) to represent the second sub-goal, wherein in each function entry the first term is a difference measure which, for a respective previously known raw data set (RD'), is a measure of the difference between the bit rate information (yT1, ..., yTn) of the training entry (T1, ..., Tn) to which the rule (RUT1, ..., RUTn) of the with the respective function entry correlated feature set (FS1, ..., FSm) is assigned and is contained in the training vector (Y) for the respective pre-known raw data set (RD'), and the bit rate information (ŷ1, ..., ŷm) of the bit rate entry (B1, ..., Bm) to which the feature set (FS1, ..., FSm) correlated with the respective function entry is assigned and which is contained in the bit rate vector (Ŷ) that the neural network (NN) outputs at its output for the respective pre-known raw data set (RD) at its input;- in each function entry the second term is a difference measure that, for a respective pre-known raw data set (RD'), is a measure of the difference between the specified bit rate value (R1, ..., Bm) or the specified bit rate interval and the respective bit rates (BR) of the bit rate information (ŷ1, ..., ŷm) of the bit rate entry (B1, ..., Bm) to which the feature set correlated with the respective function entry is assigned; Feature set (FS1, ..., FSm) is assigned and is contained in the bitrate vector (Ŷ) that the neural network (NN) outputs for the respective pre-known raw data set (RD') at its input. The method according to claim 2, characterized in that the optimization goal is represented by finding the minimum of the following cost function (L): L = [ α 1 ⋮ α m ] ⋅ [ ( y 1 − y ^ 1 ) 2 ⋮ ( ym − y ^ m ) 2 ] + [ β 1 ⋮ β m ] ⋅ [ ( R 1 − y ^ 1 ) 2 ⋮ ( R m − y ^ m ) 2 ] where [ β 1 ⋮ β m ] = ( [ 1 ⋮ 1 ] − [ α 1 ⋮ α m ] ) where L is the cost function (L); where m is the number of function entries; where α r (r = 1, ..., m) the first weight (α r ) for the feature set correlated with the respective function entry; where y r (r = 1, ..., m) the bit rate information (y T1 , ..., y Tn ) of the training entry (T1, ..., Tn) to which the regulation (RU) T1 , ..., RU Tn ) of the feature set (FS1, ..., FSm) correlated with the respective function entry and which is contained in the training vector (Y) for the respective previously known raw data set (RD'); where ŷ r (r = 1, ..., M) the bitrate information (ŷ 1 , ..., ŷ m ) of the bitrate entry (B1, ..., Bm) to which the feature set (FS1, ..., FSm) correlated with the respective function entry is assigned and which is contained in the bitrate vector (Y) that the neural network (NN) outputs at its output for the respective previously known raw data set (RD') at its input; where β r (r = 1, ..., M) is the second weight for the feature set correlated with the respective function entry; where R r (r = 1, ..., M) is the specified bit rate value for the feature set (FS1, ..., FSm) correlated with the respective function entry, where the specified bit rate value is identical for all bit rates of the bit rate information (ŷ r ) of the bitrate entry (B1, ..., Bm) to which the feature set (FS1, ..., FSm) correlated with the respective function entry is assigned and which is contained in the bitrate vector (Y) that the neural network (NN) outputs at its output for the respective previously known raw data set (RD') at its input. A method according to one of the preceding claims, characterized in that, within the framework of the generation of the training data (TD) for at least one procedure (RUT1, ..., RUTn) of the number of procedures (RUT1, ..., RUTn), SAR processing (SPR) is carried out for a respective previously known raw data set (RD'), which includes the determination of one or more SAR images (IM) from at least the respective previously known raw data set (RD'), wherein the values ​​of the at least one parameter (PT1, ..., PTn) associated with the at least one procedure (RUT1, ..., RUTn) are obtained from information provided by the SAR processing (SPR), wherein the values ​​of the at least one parameter (PT1, ..., PTn) are preferably quality values ​​for a SAR product obtained from the SAR processing. Method according to claim 4, characterized in that the at least one parameter (PT1, ..., PTn) associated with the at least one rule (RUT1, ..., RUTn) comprises the standard deviation of the distribution of the backscattered signals in a section around a pixel (PI) of a SAR image (IM) determined for the respective previously known raw data set (RD') without quantization or after quantization with constant bit rate, wherein values ​​of the standard deviation are determined for a plurality of pixels in the SAR image (IM). A method according to claim 4 or 5, characterized in that the at least one parameter (PT1, ..., PTn) associated with the at least one rule (RUT1, ..., RUTn) comprises the signal-to-quantization noise ratio in a section around a pixel (PI) of a SAR image (IM) determined for the respective previously known raw data set (RD') after quantization with a constant bit rate, wherein values ​​of the signal-to-quantization noise ratio for a plurality of pixels (PI) in several SAR images (IM) determined for the respective previously known raw data set (RD') are determined, wherein the several SAR images (IM) differ in the bit rate of the quantization of the respective previously known raw data set (RD'). A method according to one of claims 4 to 6, characterized in that the at least one parameter (PT1, ..., PTn) associated with the at least one rule (RUT1, ..., RUTn) comprises the loss of coherence caused by quantization in a section around a pixel of a SAR interferogram, which is determined from the respective previously known raw data set (RD') and at least one further previously known raw data set after quantization with a constant bit rate, wherein values ​​of the loss of coherence caused by quantization are determined for a plurality of pixels for several SAR interferograms, wherein the several SAR interferograms differ in the bit rate of the quantization of the respective previously known raw data set (RD') and of the at least one further previously known raw data set. A method according to any one of claims 4 to 7, characterized in that the at least one parameter (PT1, ..., PTn) associated with the at least one rule (RUT1, ..., RUTn) comprises the standard deviation of the distribution of the phase error caused by quantization in a section around a pixel of a SAR interferogram, which is determined from the respective previously known raw data set (RD') and at least one further previously known raw data set after quantization with a constant bit rate, wherein values ​​of the standard deviation of the distribution of the phase error caused by quantization are determined for a plurality of pixels for several SAR interferograms, wherein the several SAR interferograms differ in the bit rate of the quantization of the respective previously known raw data set (RD') and of the at least one further previously known raw data set. Method according to one of the preceding claims, characterized in that for at least one rule (RUT1, ..., RUTn) of the number of rules (RUT1, ..., RUTn) the values ​​of the at least one parameter (PT1, ..., PTn) which is linked to the at least one rule (RUT1, ..., RUTn) are obtained directly from the data samples (SP') of the data blocks (DB') of the respective previously known raw data set (RD'), without SAR processing (SPR) being carried out for the acquisition of the at least one parameter (P), which includes the determination of SAR images (IM) from at least the respective previously known raw data set. Method according to claim 9, characterized in that the at least one parameter (PT1, ..., PTn) associated with the at least one rule (RUT1, ..., RUTn) is the mean signal-to-noise ratio of the data samples (SP') in a data block (DB') in the respective previously known raw data set (RD'), wherein values ​​of the signal-to-noise ratio are determined for each data block (DB') in the respective previously known raw data set (RD'). Method according to one of the preceding claims, characterized in that the neural network (NN) is a Convolutional Neural Network. Method according to one of the preceding claims, characterized in that the neural network (NN) is a deep neural network. Method for computer-aided processing of digital SAR raw data comprising the backscattered radar echoes of radar pulses (RP), wherein the radar pulses (RP) were emitted by one or more transmitters of at least one object (SA) moving in a first direction (x), and the radar echoes were received by one or more receivers of the at least one object (SA), wherein the SAR raw data for each receiver contains a raw data set (RD) consisting of a plurality of temporally successive data samples (SP), each of which is assigned a position in the first direction (x) and a position in a second direction (y) perpendicular to the first direction (x), wherein, in the context of computer-aided processing of the SAR raw data, data blocks (DB) of temporally successive data samples (SP) of the raw data set (RD) are processed separately with a bit rate (BR) assigned to the respective data block (DB).which defines the number of bits per quantized data sample (QSP) in the respective data block (DB), are quantized using bit rate information (ŷ1, ..., ŷm), wherein the bit rate information f(ŷ1, ..., ŷm) specifies the assigned bit rate (BR) for each data block (DB), wherein the bit rate information (ŷ1, ..., ŷm) corresponds to a bit rate entry (B1, ..., Bm) generated with a data-driven model created by a method according to any one of claims 1 to 12. Method according to claim 13, characterized in that the quantized data samples (QSP) are determined on the at least one object (SA) and subsequently transmitted to a ground station. Device for computer-aided processing of digital SAR raw data comprising the backscattered radar echoes of radar pulses (RP), wherein the radar pulses (RP) were emitted by one or more transmitters of at least one object (SA) moving in a first direction (x), and the radar echoes were received by one or more receivers of the at least one object (SA), wherein the SAR raw data for each receiver contains a raw data set (RD) consisting of a plurality of temporally successive data samples (SP), each of which is assigned a position in the first direction (x) and a position in a second direction (y) perpendicular to the first direction (x), wherein the device comprises a quantizer configured toWithin the framework of computer-aided processing of the SAR raw data, data blocks (DB) of temporally successive data samples (SP) of the raw data set (RD) are quantized separately with a bit rate (BR) assigned to the respective data block (DB), which determines the number of bits per quantized data sample (QSP) in the respective data block (DB), using bit rate information (ŷ1, ..., ŷm), wherein the bit rate information (ŷ1, ..., ŷm) specifies the assigned bit rate (BR) for each data block (DB), wherein the quantizer for generating the bit rate information (ŷ1, ..., ŷm) includes a data-driven model that was generated using a method according to any one of claims 1 to 12. Computer program product comprising program code stored on a machine-readable medium for carrying out a method according to any one of claims 1 to 14, when the program code is executed on a computer. Computer program comprising program code for carrying out a method according to any one of claims 1 to 14, when the program code is executed on a computer.

Images