Multi-model PolInSAR forest height inversion method fused with deep learning normal form

Abstract

The invention discloses a multi-model PolInSAR forest height inversion method fused with a deep learning normal form, and the method comprises the steps: building a back scattering model similar to a quadratic function based on an RVoG inversion frame, and defining a volume decoherence expression of back scattering; defining respective backscattering contours by using different parameters; in consideration of model heterogeneity, a plurality of backscattering models are obtained through a linear combination quadratic function model and a GVB model; different models are classified according to application ranges, a model pool is constructed, a CNN classifier is designed to complete classification work, a plurality of models are solved through a nonlinear iterative algorithm, CHM data of LIDAR is used for distributing model labels for pixels, and a data set is obtained; according to the method, a CNN (Convolutional Neural Network) is trained, a data set is balanced through random downsampling, proper hyper-parameters are set, after training is completed, data of a research area is input into the network for classification, an optimal backscattering model is solved by using a nonlinear iterative algorithm according to a classification result, and finally a forest height inversion result is obtained.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing inversion, and in particular relates to a multi-model PolInSAR forest height inversion method integrating a deep learning paradigm. Background Art

[0002] Polarimetric interferometric synthetic aperture radar is a technology that combines radar polarimetry and interferometry. It has been widely used in the field of forest parameter and surface parameter inversion. The complex coherence coefficient under different polarizations can be obtained by taking advantage of the different sensitivities of surface scatterers to different polarizations. This includes the phase echo information and coherence amplitude information of the ground objects. By using the random volume over ground (RVoG) two-layer scattering model, the radar scattering information and forest physical parameters can be correlated to extract the canopy height model (CHM).

[0003] The RVoG model assumes that the forest-covered area is a two-layer model. The upper layer is the volume scattering layer, which contains forest scattering information, and the lower layer is the impenetrable surface layer, where surface scattering mainly occurs. The scattering process of microwaves in the volume layer is uncertain. The traditional model simplifies this process into a structure function similar to an exponential function. This simplification is more effective when the radar penetration capability is weak. However, under the long-wave SAR field of view, the backscattering effect of the thin branches in the upper layer is relatively weakened, while the surface scattering and higher-order scattering (secondary and above scattering) caused by the trunks and the ground in the lower layer can no longer be ignored. Although later, researchers proposed the Gaussian Vertical Backscatter (GVB) model, in the case of complex backscattering, a single vertical structure function cannot accurately describe the distribution of scatterers in the vertical direction of the forest. In addition, the structure function has the characteristic of being independent of polarization, and the structure function in the volume layer can be arbitrary, so multiple backscattering is proposed to better estimate the forest height unbiased.

[0004] Deep learning can fit the complex nonlinear relationship between data and labels very well. In the absence of strong correlation features between the backscattering function and the complex coherence coefficient, a convolutional neural network (CNN) is used to build a classifier to classify the optimal backscattering function. Summary of the invention

[0005] In order to solve the technical problems existing in the background technology, the present invention aims to provide a multi-model PolInSAR forest height inversion method integrating the deep learning paradigm, by constructing a data set, dividing the sample area, performing multi-model inversion, combining LIDAR data to determine the label, and selecting features to obtain a data set through normalization. After that, the classifier in the classification framework is trained to obtain the classification result for height inversion.

[0006] In order to solve the technical problem, the technical solution of the present invention is:

[0007] A multi-model PolInSAR forest height inversion method integrating a deep learning paradigm, the method comprising:

[0008] Step S1: Based on the RVoG inversion framework, a backscattering model similar to a quadratic function is constructed to define a volume decorrelation expression of backscattering; the model includes: an exponential model, a GVB model and a quadratic function model, and different parameters are used to define respective backscattering profiles; considering the heterogeneity of the model, a plurality of backscattering models are obtained by linearly combining the quadratic function model and the GVB model;

[0009] Step S2: Classify different models according to their applicable scope, build a model pool, and design a CNN classifier to complete the classification work. Solve multiple models through a nonlinear iterative algorithm, use the CHM data of LIDAR to assign model labels to pixels, and obtain a data set;

[0010] Step S3: CNN network model training and model selection. Based on the classifier and data set obtained in step S2, the Adam optimization algorithm is used to train the CNN. The data set is balanced by random downsampling and appropriate hyperparameters are set. After training, the data of the study area are input into the network for classification. According to the classification results, a nonlinear iterative algorithm is used to solve the optimal backscattering model, and finally the forest height inversion result is obtained.

[0011] Further, in step S1, based on the RVoG inversion framework, the pure volume decorrelation coefficient expression is as follows:

[0012]

[0013] where h is the forest height; z represents the volume layer thickness; i is an imaginary unit; k z is the vertical effective wave number, representing the phase change per unit height; in the traditional RVoG model, f(z) is assumed to be an exponential function curve, and the complex coherence coefficient expression of pure volume decoherence is:

[0014]

[0015]

[0016] Where σ is the extinction coefficient, representing the attenuation of microwaves in the volume layer; θ is the radar incident angle;

[0017] In the GVB model, f(z) is assumed to be a Gaussian curve, and the pure volume decorrelation coefficient is expressed as follows:

[0018]

[0019] Where ε is the Gaussian mean, representing the vertical height of the maximum backscattering intensity; χ is the Gaussian standard deviation, indicating the degree of fluctuation of backscattering; erf is the error function;

[0020] Analogous to the GVB model, a quadratic function form is proposed, and its basic expression and pure volume decorrelation expression are as follows:

[0021]

[0022] Among them: ε represents the vertical height with the minimum backscattering intensity; χ represents the fluctuation degree of backscattering; erfi is the imaginary error function:

[0023] erfi(z)=-ierf(iz) (8)

[0024] Further, in step S1, the quadratic function model and the GVB model are linearly combined to realize more contour models;

[0025] The backscattering function f(z) is expressed as:

[0026]

[0027] The pure body decorrelation expression after the combination is:

[0028]

[0029] Among them, erf and erfi are the error function and imaginary error function respectively;

[0030] By fixing the two parameters ε and χ, the parameters c and d are solved, and three representative backscattering models are selected. The ε parameter is fixed to h / 2, 0, and h, and the χ parameter is fixed to h / 4, h / 8, and h / 4, respectively.

[0031] Further, the step S2 comprises:

[0032] A classification framework was constructed to divide the backscattering models into two model pools: model pool 1 includes exponential model, GVB model and quadratic function model, and model pool 2 includes three backscattering models; CNN classifier was used to classify pixels. First, classifier 1 was used to classify pixels into model pool 1 or model pool 2, and then classifiers 2 and 3 were used to subdivide specific models; the constructed CNN classification network structure included multi-layer convolution, maximum pooling, batch normalization layer and activation layer, and a deep neural network was used to learn nonlinear features; then a classification data set was constructed to extract features from SAR data, including phase information, amplitude information and complex plane geometry information; all features were linearly normalized and outliers were removed to ensure data balance to improve the stability of the training process, and sample area classification labels were constructed. The CHMR data of LIDAR was used for auxiliary decision-making to select the model label with the smallest error;

[0033] Among them, firstly, features are selected through SAR complex coherence coefficient, model pool classification is decided by classifier 1, and secondary classification is subdivided into optimal models by classifiers 2 and 3.

[0034] Further, in step S2:

[0035] Phase information:

[0036] angle diff =angle(PDH)-angle(PDL) (11)

[0037] Where angle(i) represents the phase corresponding to polarization i, and eight features including coherent optimization channels PDH-PDL, PDH-Opt1, PDL-Opt1, HV-HH, HV-VV, and HH-VV are selected in turn;

[0038] Amplitude information:

[0039] abs diff =abs(PDH)-abs(PDL) (12)

[0040] Where abs(i) represents the amplitude corresponding to polarization i, and eight features including coherent optimization channels PDH-PDL, PDH-Opt1, PDL-Opt1, HV-HH, HV-VV, and HH-VV are selected in turn;

[0041] Complex plane geometry information:

[0042]

[0043] Where abs(i) represents the amplitude of the corresponding value, Representing the complex value under unit amplitude corresponding to the surface phase, the coherent optimization channels PDH, PDL, Opt1, Opt2, Opt3 and the distance from the surface phase point and two distances such as PDH-PDL and Opt1-PDL were selected in turn, totaling 7 features;

[0044] In addition to the above features, the initial values ​​of tree heights at three stages, the incident angle, and the ground amplitude ratios of all channels were selected as feature inputs.

[0045] Further, constructing the sample area classification label includes:

[0046] We selected small areas within the study area, calculated the forest heights of each area under the six models, used the lidar values ​​as the true values ​​to select training samples and construct a labeled dataset;

[0047] Classifier 1 label construction: The function of classifier 1 is to classify pixels into model pool 1 or model pool 2. Therefore, in the training area, the top three models with the smallest error are selected for decision making. If the number of models from model pool 1 is large, they are classified as label 1, otherwise they are classified as label 2.

[0048] Classifier 2 and 3 label construction: The two secondary classifiers are used to subdivide the three models in the model pool. First, the size of the model error is determined:

[0049] error i =(h i -h lidar ) 2 (14)

[0050] where h lidar As the true value of this pixel, h i is the value calculated by the three models in the corresponding model pool, error i It can be used as an indicator to determine the optimal backscattering model for decision making;

[0051] In multi-model classification, attention is paid to the heterogeneity of each model. If the difference in performance between models is not obvious enough and there is no feature that is strongly correlated with backscatter power, it is necessary to maintain the discrimination between models.

[0052] Further, the step S3 comprises:

[0053] S301: According to the classification framework and data set established in step 2, the classification framework is trained, the data set is divided into a training set and a test set, all data are flattened into 1 dimension, and are trained in three models respectively. All data sets are balanced by random downsampling. The Adam gradient descent optimization algorithm is used, the number of iterations is set to 10 times, and overfitting will not occur while the convergence is the fastest. The initial learning rate is 0.001, and the learning rate reduction factor is set to 0.1. It is reduced after 8 trainings. The data set is shuffled for each training, the verification frequency is 500 times, and the batch size is set to 128;

[0054] S302: Model selection. After step S301, the entire classification framework is obtained. The study area is input into the network according to the network input layer format, and it is classified. After the classification result of each pixel is obtained, it is solved under nonlinear iteration.

[0055] Compared with the prior art, the advantages of the present invention are:

[0056] The present invention provides a multi-model PolInSAR forest height inversion method integrating deep learning paradigm. A new backscattering profile is proposed to better fit the trend of the scattering power of each pixel changing with height, and then an inversion framework of different models is established. Then, a convolutional neural network (CNN) is used to classify pixels to select the optimal backscattering profile.

[0057] In the traditional RVoG model, the backscattering function is assumed to be an exponential function, and the extinction coefficient is used as a parameter to adjust the exponential function, which can achieve an unbiased estimation of forest height. However, in the P band or other long-wave fields of view, this assumption is no longer reasonable. Therefore, a new backscattering profile is proposed to complement the existing exponential model and GVB model, providing a more reasonable backscattering profile to fit each pixel and improve the inversion accuracy of forest height; the classification of the backscattering function based on the complex coherence coefficient is a highly nonlinear classification problem. Through the CNN classifier, a classification framework between models with different applicability is established to prevent the serious interference of the unbalanced number of samples on the classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 , the process framework of the embodiment of the present invention;

[0059] Figure 2 , 6 types of profile schematic diagrams used in the embodiments of the present invention: (a) exponential model; (b) GVB model; (c) quadratic function model; (d), (e), (f) are profile schematic diagrams when ε1=h / 2, χ1=h / 4, ε2=0, χ2=h / 8, ε3=h, χ3=h / 4 respectively;

[0060] Figure 3 , Changes of backscattering function when parameters of GVB model and quadratic model are changed: (a) GVB model; (b) quadratic function model;

[0061] Figure 4 ,Schematic diagram of the model classification framework;

[0062] Figure 5 , Schematic diagram of the CNN classifier network structure. DETAILED DESCRIPTION

[0063] The specific implementation mode of the present invention is described below in conjunction with embodiments:

[0064] It should be noted that the structures, proportions, sizes, etc. shown in this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them, and are not used to limit the conditions under which the present invention can be implemented. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0065] At the same time, the terms such as "upper", "lower", "left", "right", "middle" and "one" cited in this specification are only for the convenience of description and are not used to limit the scope of implementation of the present invention. Changes or adjustments to their relative relationships should be regarded as the scope of implementation of the present invention without substantially changing the technical content.

[0066] Embodiment 1:

[0067] The present invention provides a multi-backscattering function classification framework based on deep learning and a PolInSAR forest height inversion method; by proposing a new backscattering profile to better fit the variation of the scattering power of each pixel with height, an inversion framework of different models is established, and then a convolutional neural network (CNN) is used to classify the pixels to select the optimal backscattering profile. Figure 1 As shown, the present invention comprises the following steps:

[0068] Step 1: Construct various backscattering functions.

[0069] Step 1.1: Propose a quadratic function-like backscattering model.

[0070] In the RVoG inversion framework, the overall decoherence is defined as follows:

[0071]

[0072] where h is the forest height; z represents the volume layer thickness; i is an imaginary unit; k zis the vertical effective wave number, representing the phase change caused by unit height. In the traditional RVoG model, f(z) is assumed to be an exponential function curve, such as Figure 2 (a), the complex coherence coefficient expression of pure volume decorrelation is:

[0073]

[0074] Where σ is the extinction coefficient, representing the attenuation of microwaves in the volume layer; θ is the radar incident angle, and other parameters are consistent with the above formula.

[0075] In the GVB model, f(z) is assumed to be a Gaussian curve, such as Figure 2 (b), the pure volume decorrelation coefficient is expressed as follows:

[0076]

[0077] Where ε is the Gaussian mean, representing the vertical height of the maximum backscattering intensity; χ is the Gaussian standard deviation, indicating the degree of fluctuation of backscattering; erf is the error function, and the other parameters are consistent with the above formula. The backscattering profile shapes corresponding to different parameters are as follows Figure 3 (a) shown.

[0078] Different backscatter functions provide different ways of retrieving height in the inversion, similar to different k in multi-baseline inversion. z effect. A well-applicable backscattering function model needs to meet two conditions: First, for a single backscattering function framework, the structural parameters (such as extinction coefficient σ, Gaussian mean ε, Gaussian standard deviation χ) need to be flexible enough to realize the flexible change of the scattering function, so that the scattering function has enough space to fit the backscattering of each pixel, otherwise it will cause serious biased estimation. Second, the number of solution parameters cannot be greater than 2. In the case of a single baseline, too many unknowns in the model will cause underdetermination. Even a 6-dimensional nonlinear iterative algorithm that can solve multiple unknowns can only solve two unknowns at the same time. Therefore, the proposed function model has a maximum of 2 unknowns. The GVB model and the exponential model are both models under this condition.

[0079] Analogous to the GVB model, a quadratic function form is proposed, whose basic function form is as follows Figure 2 (c), its basic expression and pure volume decorrelation expression are as follows:

[0080]

[0081]

[0082] Among them: ε represents the vertical height with the minimum backscattering intensity; χ represents the fluctuation degree of backscattering; erfi is the imaginary error function:

[0083] erfi(z)=-ierf(iz) (8)

[0084] The backscattering profiles corresponding to different parameters are as follows: Figure 3 (b) as shown.

[0085] Step 1.2: Linearly combine the quadratic function model and the Gaussian backscattering model to achieve more contour models.

[0086] like Figure 3 As shown, the quadratic function form and the GVB model are combined to achieve more backscattering profiles. At this time, the backscattering function f(z) is expressed as:

[0087]

[0088] The pure body decorrelation expression after the combination is:

[0089]

[0090] Among them, erf and erfi are the error function and imaginary error function respectively;

[0091] c, d, χ, and ε are four parameters involved in the inversion function to describe the backscattering. The function profile can be adjusted to fit the backscattering in different situations. In this case, the model contains 4 unknowns that cannot be solved, so two of the parameters are fixed for solution.

[0092] In the entire multi-model inversion framework, more attention is paid to the heterogeneity of each model. If the two parameters ε and χ are inverted, the combined model will lead to unclear outlines of the backscattering function, increase the ambiguity of the inversion, and make subsequent classification difficult and cause excessive computational redundancy in the inversion. For the above reasons, the two parameters ε and χ are fixed, and the parameters c and d are solved. Theoretically, different values ​​of ε and χ will produce a series of different backscattering functions. However, in order to ensure the differences between the backscattering functions and the feasibility of classifying different models, the three most representative models are selected. The ε parameter is fixed to h / 2, 0, and h, respectively, and the χ parameter is fixed to h / 4, h / 8, and h / 4, respectively. The backscattering is as follows: Figure 2 As shown in (d), (e) and (f).

[0093] Step 2: Implementation of model classification.

[0094] The classifier is composed of a model pool classifier and two secondary model classifiers, using the same classification features. CNN classifiers are good at mining the nonlinear relationship between features (SAR amplitude and phase information in this example) and targets (backscatter function in this example), so CNN is used for classification.

[0095] Step 2.1: Construct a classification framework. As shown in the above step 1.1, the exponential model, GVB model and quadratic function model have different performances in inversion than the latter three models. The first three models are subsequently referred to as model pool 1 and the latter three as model pool 2. The parameters in model pool 1 can satisfy the adjustment of the backscattering function within a larger range, thus having more universal applicability, while the three models in model pool 2 do not have this feature and are only applicable to a small number of pixels. Using a single classifier will increase the number of misclassified pixels in model pool 2. In order to solve this "unbalanced" classification problem, a new classification framework was established. Its classification process is as follows: Figure 4 As shown, first, features are selected through SAR observation information, and the model pool is classified in the first classifier to determine which model pool the pixel should belong to. Subsequently, a single model is subdivided in the 1 or 2 model pool, and finally the most suitable backscatter is obtained by integration.

[0096] The depth and width of the neural network show different characteristics in classification performance. A deeper neural network means that it has stronger nonlinear expression capabilities, while a wider network means that each layer can learn more feature information. Figure 5 The constructed CNN network structure is shown. The two convolutional layers in the output layer also each contain a batch normalization layer and an activation layer. In the hidden layer, the number of channels gradually increases from 124 to 1024, and then gradually decreases to 256 in the output layer after the maximum pooling filter. After the two convolutional layers in the output layer, there is also a batch normalization layer (BatchNormalization) and an activation layer (ReLU). Figure 4 The structures of classifiers 2 and 3 in the network are shown in the figure. Since the main task of classifier 1 is to classify the pixel model pool, its nonlinearity is lower than that of classifiers 2 and 3, so two hidden layers are reduced. Classifiers 2 and 3 contain 34 layers in total, while classifier 1 contains 26 layers. In order to prevent overfitting, a random inactivation layer (Dropout) is set in the network to avoid excessive weights of some neurons by randomly discarding some hidden neurons.

[0097] Step 2.2: Build a classification dataset;

[0098] Step 2.2.1: Feature selection. The features selected for the training area must be representative of the entire study area. At the same time, since the classification in this example is relatively special, in order to ensure the accuracy of the classification network training, the selected features cannot have trends, that is, to ensure that the feature distribution of the sample area is consistent with the feature distribution of the entire test area, such as k z (At close and long distances, there is an obvious trend from large to small.) In addition, in this example, the phase and amplitude information of the pixel are subtracted to better characterize the difference in the response of each pixel to the scattering information (the change of the backscattering function) under the changed backscattering. As shown below, it is generally divided into three parts: phase information, amplitude information, and complex plane geometric information.

[0099] Phase information:

[0100] angle diff =angle(PDH)-angle(PDL) (11)

[0101] Among them, angle(i) represents the phase corresponding to polarization i, and eight features including coherent optimization channels PDH-PDL, PDH-Opt1, PDL-Opt1, HV-HH, HV-VV, and HH-VV are selected in turn.

[0102] Amplitude information:

[0103] abs diff =abs(PDH)-abs(PDL) (12)

[0104] Where abs(i) represents the amplitude corresponding to polarization i, and eight features including coherent optimization channels PDH-PDL, PDH-Opt1, PDL-Opt1, HV-HH, HV-VV, and HH-VV are selected in turn.

[0105] Complex plane geometry information:

[0106]

[0107] Where abs(i) represents the amplitude of the corresponding value, Represents the complex value of the unit amplitude corresponding to the surface phase. The coherent optimization channels PDH, PDL, Opt1, Opt2, Opt3 are selected in turn. The distances from the surface phase point and the distances PDH-PDL and Opt1-PDL are a total of 7 features.

[0108] In addition to the above features, the initial values ​​of the three-stage tree height, the incident angle, and the ground amplitude ratio of all channels (the relative ground amplitude ratio when the ground amplitude ratio of the pure body decorrelation channel is 0) are also selected as feature inputs.

[0109] Step 2.2.2: Normalization. Normalizing features is a very important step before training a neural network, which allows the network to converge more quickly while improving the generalization ability of the network. All features are processed using linear normalization: each feature is flattened into 1 dimension, and then the minimum and maximum values ​​of each row are mapped to [0,1] for normalization. Before performing this step, it is necessary to process possible outliers in the data, remove some excessively large values ​​or null values, and replace them with the average value of the surrounding pixels within a certain range, so as to prevent erroneous values ​​from affecting the entire normalization process. After normalization, the selected training area is cropped out to participate in the subsequent training process.

[0110] Step 2.2.3: Classification label. Select a small area within the study area, calculate the forest height of each area under the six models, use the lidar value as the true value to select training samples and construct a label dataset.

[0111] Classifier 1 label construction: The function of classifier 1 is to classify pixels into model pool 1 or model pool 2. Therefore, in the training area, the top three models with the smallest error are selected for decision making. If the number of models from model pool 1 is large, they are classified as label 1, otherwise they are classified as label 2.

[0112] Classifier 2 and 3 label construction: The two secondary classifiers are used to subdivide the three models in the model pool. First, the size of the model error is determined:

[0113] error i =(h i -h lidar ) 2 (14)

[0114] where h lidar As the true value of this pixel, h i is the value calculated by the three models in the corresponding model pool, error i It can be used as an indicator to determine the optimal backscattering model for decision making.

[0115] In addition, there are a few pixels where the performance of the three models in the model pool is not much different. For such pixels, useless noise information will be introduced, so they are not involved in training. In multi-model classification, attention is paid to the heterogeneity of each model. If the performance difference between the models is not obvious enough and there is no feature that is strongly correlated with the backscatter power, it is necessary to maintain the distinction between the models, otherwise the classification accuracy will be greatly reduced due to label confusion.

[0116] Step 3: CNN network model training and model selection

[0117] Step 3.1: Network training. The data set is divided into training, test and validation sets in a ratio of 8:1:1. Since normalization was performed before training, it is no longer normalized. All data are flattened into 1 dimension and trained in 3 models respectively. All data sets are balanced by random downsampling to prevent the training results from being biased towards the side with a larger number of samples. The Adam gradient descent optimization algorithm is used, and the number of iterations is set to 10 times. It converges fastest without overfitting. The initial learning rate is 0.001, and the learning rate reduction factor is set to 0.1. It is reduced after 8 trainings. The data set is shuffled for each training, the verification frequency is 500 times, and the batch size is set to 128.

[0118] Step 3.2: Model selection. After step 3.1, the entire classification framework is obtained. The study area is input into the network according to the network input layer format and classified. After obtaining the classification result of each pixel, it is solved under nonlinear iteration.

[0119] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0121] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0122] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0123] The preferred embodiments of the present invention are described in detail above, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

[0124] Many other changes and modifications may be made without departing from the concept and scope of the present invention.It should be understood that the present invention is not limited to the specific embodiments, and the scope of the present invention is defined by the appended claims.

Claims

1. A multi-model PolInSAR forest height inversion method integrating deep learning paradigm, characterized in that: The method comprises: Step S1: Based on the RVoG inversion framework, a backscattering model similar to a quadratic function is constructed to define a volume decorrelation expression of backscattering; the model includes: an exponential model, a GVB model and a quadratic function model, and different parameters are used to define respective backscattering profiles; considering the heterogeneity of the model, a plurality of backscattering models are obtained by linearly combining the quadratic function model and the GVB model; Step S2: Classify the above different models according to their applicable scope, build a model pool, and design a CNN classifier to complete the classification work. Solve multiple models through a nonlinear iterative algorithm, use the CHM data of LIDAR to assign model labels to pixels, and obtain a data set; Step S3: CNN network model training and model selection. Based on the classifier and data set obtained in step S2, the Adam optimization algorithm is used to train the CNN. The data set is balanced by random downsampling and appropriate hyperparameters are set. After training, the data of the study area are input into the network for classification. According to the classification results, a nonlinear iterative algorithm is used to solve the optimal backscattering model, and finally the forest height inversion result is obtained.

2. A multi-model PolInSAR forest height inversion method integrating a deep learning paradigm according to claim 1, characterized in that: In step S1, based on the RVoG inversion framework, the pure volume decorrelation coefficient expression is as follows: where h is the forest height; z represents the volume layer thickness; i is an imaginary unit; k z is the vertical effective wave number, representing the phase change per unit height; in the traditional RVoG model, f(z) is assumed to be an exponential function curve, and the complex coherence coefficient expression of pure volume decoherence is: Where σ is the extinction coefficient, representing the attenuation of microwaves in the volume layer; θ is the radar incident angle; In the GVB model, f(z) is assumed to be a Gaussian curve, and the pure volume decorrelation coefficient is expressed as follows: Where ε is the Gaussian mean, representing the vertical height of the maximum backscattering intensity; χ is the Gaussian standard deviation, indicating the degree of fluctuation of backscattering; erf is the error function; Analogous to the GVB model, a quadratic function form is proposed, and its basic expression and pure volume decorrelation expression are as follows: Among them: ε represents the vertical height with the minimum backscattering intensity; χ represents the fluctuation degree of backscattering; erfi is the imaginary error function: erfi(z)=-ierf(iz) (8).

3. The multi-model PolInSAR forest height inversion method integrating deep learning paradigm according to claim 1, characterized in that: In step S1, the quadratic function model and the GVB model are linearly combined to realize more contour models; The backscattering function f(z) is expressed as: The pure body decorrelation expression after the combination is: Among them, erf and erfi are the error function and imaginary error function respectively; By fixing the two parameters ε and χ, the parameters c and d are solved, and three representative backscattering models are selected. The ε parameter is fixed to h / 2, 0, and h, and the χ parameter is fixed to h / 4, h / 8, and h / 4, respectively.

4. The multi-model PolInSAR forest height inversion method integrating deep learning paradigm according to claim 3 is characterized in that: The step S2 comprises: A classification framework is constructed to divide the backscattering model into two model pools: model pool 1 includes: exponential model, GVB model and quadratic function model, and model pool 2 includes: the three backscattering models; a CNN classifier is used to classify pixels, firstly, the pixels are classified into model pool 1 or model pool 2 through classifier 1, and then the specific models are subdivided through classifiers 2 and 3; the constructed CNN classification network structure includes multi-layer convolution, maximum pooling, batch normalization layer and activation layer, and a neural network is used to learn nonlinear features; then a classification data set is constructed to extract features from SAR data, including phase information, amplitude information and complex plane geometric information; all features are linearly normalized, and outliers are removed to ensure data balance to improve the stability of the training process and accelerate the convergence speed, and a sample area classification label is constructed, and auxiliary decision-making is performed through the CHMR data of LIDAR to select the model label with the smallest error; Among them, firstly, features are selected through SAR complex coherence coefficient, model pool classification is decided by classifier 1, and secondary classification is subdivided into optimal models by classifiers 2 and 3.

5. The multi-model PolInSAR forest height inversion method integrating deep learning paradigm according to claim 4, characterized in that: In step S2: Phase information: angle diff =angle(PDH)-angle(PDL) (11) Where angle(i) represents the phase corresponding to polarization i, and the coherent optimization channels selected in sequence include: PDH-PDL, PDH-Opt1, PDL-Opt1 and HV-HH, HV-VV, HH-VV features; Amplitude information: abs diff =abs(PDH)-abs(PDL) (12) Where abs(i) represents the amplitude corresponding to polarization i, and the coherent optimization channels selected in turn include: PDH-PDL, PDH-Opt1, PDL-Opt1 and HV-HH, HV-VV, HH-VV features; Complex plane geometry information: where abs(i) represents the amplitude corresponding to polarization i, Representing the complex value under unit amplitude corresponding to the surface phase, the coherent optimization channels PDH, PDL, Opt1, Opt2, Opt3 and the distance from the surface phase point and the two distances PDH-PDL and Opt1-PDL were selected in turn, totaling 7 features; In addition to the above features, the initial values ​​of tree heights at three stages, the incident angle, and the ground amplitude ratios of all channels were selected as output features.

6. The multi-model PolInSAR forest height inversion method integrating deep learning paradigm according to claim 4, characterized in that: The step of constructing a sample region classification label comprises: We selected small areas within the study area, calculated the forest heights of each area under the six models, used the lidar values ​​as the true values ​​to select training samples and construct a labeled dataset; Classifier 1 label construction: The function of classifier 1 is to classify pixels into model pool 1 or model pool 2. Therefore, in the training area, the top three models with the smallest error are selected for decision making. If the number of models from model pool 1 is large, they are classified as label 1, otherwise they are classified as label 2. Classifier 2 and 3 label construction: The two secondary classifiers are used to subdivide the three models in the model pool. First, the size of the model error is determined: error i =(h i -h lidar ) 2 (14) where h lidar As the true value of this pixel, h i is the value calculated by the three models in the corresponding model pool, error i It can be used as an indicator to determine the optimal backscattering model for decision making; In multi-model classification, attention is paid to the heterogeneity of each model. If the difference in performance between models is not obvious enough and there is no feature that is strongly correlated with backscatter power, it is necessary to maintain the discrimination between models.

7. The multi-model PolInSAR forest height inversion method integrating deep learning paradigm according to claim 1, characterized in that: The step S3 comprises: S301: According to the classification framework and data set established in step 2, the classification framework is trained, the data set is divided into a training set and a test set, all data are flattened into 1 dimension, and are trained in three CNN models respectively. All data sets are balanced by random downsampling, and the Adam gradient descent optimization algorithm is used. The number of iterations is set to 10 times, and the convergence is fastest without overfitting. The initial learning rate is 0.001, and the learning rate reduction factor is set to 0.

1. It is reduced after 8 trainings. The data set is shuffled for each training, the verification frequency is 500 times, and the batch size is set to 128; S302: Model selection. After step S301, the entire classification framework is obtained. The study area is input into the network according to the network input layer format, and it is classified. After the classification result of each pixel is obtained, it is solved under nonlinear iteration.

Images