A farmland remote sensing image target detection method based on a DTNE-NTZNN model

Abstract

This invention discloses a method for target detection in farmland remote sensing images based on the DTNE-NTZNN model, comprising the following steps: acquiring farmland remote sensing image data streams; constructing a farmland remote sensing CEM detection model using constrained energy minimization; iteratively solving the farmland remote sensing CEM detection model in real time based on the constructed DTNE-NTZNN solver until convergence, outputting the system state solution vector at the current moment; extracting optimal filter coefficients based on the system state solution vector; filtering and reconstructing the farmland remote sensing image using the optimal filter coefficients to obtain a reconstructed image; calculating the optimal extraction threshold using the Otsu algorithm; segmenting the reconstructed image based on the optimal extraction threshold to obtain a target distribution map. This invention can accurately detect single and multiple targets in complex backgrounds.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image detection technology, and in particular to a method for target detection in farmland remote sensing images based on the DTNE-NTZNN model. Background Technology

[0002] With the development of precision agriculture, remote sensing technology has become a core method for crop classification and target detection. The Constrained Energy Minimization (CEM) algorithm, due to its characteristic of requiring only prior spectral information of the target without needing background statistical information, is widely used in hyperspectral and multispectral image processing of farmland. The mathematical essence of the CEM algorithm is solving a convex quadratic programming problem with linear constraints: that is, finding an optimal filter vector. This makes the filtered output energy Minimum, while simultaneously satisfying the response constraints of multiple target spectra as a constant 1 (i.e. ).in, It is the background autocorrelation matrix of the image. This is the prior spectral feature matrix of the crop to be detected. To solve this optimization problem, the Lagrange multiplier method is usually used to transform it into an equivalent system of linear equations. The solution is then required. However, in dynamic farmland scenarios such as real-time drone monitoring or video stream processing, the amount of background data is enormous and changes drastically over time or spatially. The static equations described above evolve into a time-varying dynamic linear equation system that requires real-time solution. Faced with such dynamic problems, traditional static matrix inversion methods, such as pseudo-inverse methods, have high computational complexity, reaching... This makes it difficult to meet real-time requirements.

[0003] To address the real-time solution problem, existing approximate implementation schemes mainly include: 1) using Gradient Neural Networks (GNNs) based on the gradient descent principle. However, GNNs cannot converge the error to zero when solving time-varying problems, always exhibiting a fundamental lag error, which limits detection accuracy. 2) using the standard Noise-Tolerant Zeroing Neural Network (NTZNN), introducing an integral term to suppress noise; its discretization form is usually based on Euler forward difference. Although NTZNN solves the noise sensitivity problem of standard ZNNs, its convergence speed is mainly limited by fixed scaling parameters, lacking an active acceleration mechanism to cope with rapidly changing dynamic environments. When processing large-scale farmland remote sensing data, the bottleneck of convergence speed limits the system's response frame rate. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a target detection method for farmland remote sensing images based on the DTNE-NTZNN model. In this method, the proposed Discrete-Time Nesterov-Enhanced Noise-Tolerant Zeroing Neural Network (DTNE-NTZNN) solver introduces Nesterov momentum acceleration technology into the traditional noise-resistant ZNN, utilizing "predicted location" to evaluate errors, thereby significantly improving convergence speed while maintaining noise resistance.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A method for target detection in farmland remote sensing images based on the DTNE-NTZNN model includes the following steps:

[0007] The acquired farmland remote sensing image data stream was used to construct a farmland remote sensing CEM detection model by constrained energy minimization.

[0008] The farmland remote sensing CEM detection model is solved iteratively in real time based on the constructed DTNE-NTZNN solver until convergence, and the system state solution vector at the current moment is output.

[0009] The optimal filter coefficients are extracted based on the system state solution vector. The optimal filter coefficients are then used to filter and reconstruct the farmland remote sensing image to obtain the reconstructed image. The optimal extraction threshold is calculated using the Otsu algorithm. Based on the optimal extraction threshold, the reconstructed image is segmented to obtain the target distribution map.

[0010] Preferably, the farmland remote sensing CEM detection model is constructed using constrained energy minimization, including the following steps:

[0011] Assuming the acquired farmland remote sensing image data stream is at time... The autocorrelation matrix is The prior spectral feature matrix of the target object to be detected is Based on the constraint energy minimization, a convex quadratic programming problem with linear constraints is established, and the filter coefficients are solved. , making Minimum, ;

[0012] Transform the problem into a dynamic system of linear equations:

[0013]

[0014] In the formula, the subscript Represents the iteration round; coefficient matrix Unknown variables , including Lagrange multipliers Filter coefficients to be solved ; target vector ; superscript This represents the matrix transpose operation.

[0015] Preferably, the farmland remote sensing CEM detection model is solved iteratively in real time based on the constructed DTNE-NTZNN solver, including the following steps:

[0016] Based on the unknown variables obtained from the solution of the previous time step and speed Determine the predicted position at the current moment. The formula is as follows:

[0017]

[0018] in, These are the Nesterov acceleration parameters;

[0019] Calculate the predicted location Linear equation error at [location] and integral terms The formula is as follows:

[0020]

[0021]

[0022] in, For the accumulated integrals from the past, This is the sampling step size;

[0023] Error of linear equations based on feedback and integral terms Determine intermediate variables Update velocity with momentum based on intermediate variables The formula is as follows:

[0024]

[0025]

[0026] in, for The false reversal; and ; For proportional gain parameters, This is the integral gain parameter; This is a momentum parameter; Represents the damping coefficient;

[0027] The solution is updated based on the evolution rate, and the formula is as follows:

[0028]

[0029] in, For the current solution, To update the solution.

[0030] Preferably, the optimal filter coefficients are extracted based on the system state solution vector, and the farmland remote sensing image is filtered and reconstructed using the optimal filter coefficients to obtain the reconstructed image, including the following steps:

[0031] Extracting the middle and front from the system state solution vector The elements are used as optimal filter coefficients. , This represents the number of spectral bands in the remotely sensed image.

[0032] The extracted optimal filter coefficients are used to perform pixel-by-pixel linear filtering on the current frame of the remote sensing image, and the output response value of each pixel is calculated. The formula is as follows:

[0033]

[0034] in, Represents any one in the remote sensing image The spectral feature vector of each pixel.

[0035] Based on the above technical solution, the beneficial effects of the present invention are:

[0036] 1) This invention achieves extremely fast convergence: by using the Nesterov acceleration mechanism and introducing "predicted position" and momentum term, the number of iteration steps is significantly reduced without increasing time complexity, thus greatly improving the solution speed of CEM filter coefficients;

[0037] 2) This invention maintains high robustness: it inherits the integral feedback mechanism of NTZNN, effectively suppressing common sensor thermal noise and atmospheric interference in farmland remote sensing imaging;

[0038] 3) This invention enhances dynamic tracking capabilities: by "forward-looking" assessment of error trends, it eliminates lag errors and enables real-time and accurate detection of dynamically changing farmland targets. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of a farmland remote sensing image target detection method based on the DTNE-NTZNN model in one embodiment;

[0040] Figure 2This is a schematic diagram of single-target detection results on the Indian pines dataset in one embodiment, where (a) is the ground truth map of each type of crop; (b) is the original image (drawn using the 50th band); (c) is the ground truth map of oats; (d) is the detection result map; (e) is the target region; and (f) is a schematic diagram of the real-time residual of the solution system.

[0041] Figure 3 This is a schematic diagram of single-target detection results on the Salinas dataset in one embodiment, where (a) is the original image (drawn using the 50th band); (b) is the ground truth image of celery; (c) is the detection result; and (d) is the target region.

[0042] Figure 4 This is a schematic diagram of multi-target detection results on the Indian pines dataset in one embodiment, where (a) is the ground truth map of all types; (b) is the original image (drawn in the 50th band); (c) is the ground truth map of grassland, wheat, and buildings; (d) is the detection result; (e) is the target region; and (f) is the real-time residual of the solution system.

[0043] Figure 5 This is a schematic diagram of the multi-target detection results of the Salinas dataset in one embodiment, where (a) is the original image (drawn using the 50th band); (b) is the ground truth image of weeds and celery; (c) is the detection result; and (d) is the target region.

[0044] Figure 6 This is a schematic diagram of the multi-target detection results of the WHU-Hi-LongKou dataset in one embodiment, where (a) is the original image (drawn in the 200th band); (b) is the ground truth image of sesame and rice; (c) is the detection result; and (d) is the target region. Detailed Implementation

[0045] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0046] like Figure 1 As shown, this embodiment provides a target detection method for farmland remote sensing images based on the DTNE-NTZNN model, including the following steps:

[0047] Step 1: Obtain the farmland remote sensing image data stream and construct a farmland remote sensing CEM detection model using constrained energy minimization.

[0048] In this embodiment, it is assumed that the acquired farmland remote sensing image data stream is at time... The autocorrelation matrix is The prior spectral feature matrix of the crop to be tested is The CEM optimization problem is transformed into solving the following system of time-varying linear equations:

[0049]

[0050] Among them, with The relevant subscripts represent discrete time steps in the iteration process. It represents the previous moment, signifying the past state. It represents the current moment and the current state. It represents the next time step, indicating the updated state; coefficient matrix The matrix updates dynamically as the image background changes; unknown variables It contains the filter coefficients to be solved. Lagrange multipliers; target vector ; superscript This represents the matrix transpose operation.

[0051] Step 2: Based on the constructed DTNE-NTZNN solver, the farmland remote sensing CEM detection model is solved iteratively in real time until convergence, and the system state solution vector at the current moment is output.

[0052] In this embodiment, in order to solve the above equations in real time and quickly... A discrete-time Nesterov accelerated noise suppression nullification neural network (DTNE-NTZNN) solver was designed. This DTNE-NTZNN solver introduces the prediction location... The error is evaluated at the predicted location. The specific iterative steps of the solver are as follows:

[0053] 1) Calculate the predicted position (Lookahead). Based on the unknown variables solved in the previous time step. and speed Calculate the current predicted position The formula is as follows:

[0054]

[0055] in, These are the Nesterov acceleration parameters.

[0056] 2) Calculate the prediction error and integral feedback. Calculate the error at the predicted location. Linear equation error at [location] Instead of the current point error, the formula is as follows:

[0057]

[0058] And calculate the integral term. To suppress noise, the formula is as follows:

[0059]

[0060] in, For the accumulated integrals from the past, This is the sampling step size.

[0061] 3) Update the evolution direction. Calculate the evolution rate at the current moment based on the predicted linear equation error and integral term. First, calculate the intermediate variables. The formula is as follows:

[0062]

[0063] Then update the velocity with momentum. The formula is as follows:

[0064]

[0065] in, for The false reversal; and ; For proportional gain parameters, This is the integral gain parameter; This is a momentum parameter; This represents the damping coefficient.

[0066] 4) Update the solution.

[0067]

[0068] in, For the current solution, To update the solution.

[0069] Step 3: Extract the optimal filter coefficients based on the system state solution vector, and use the optimal filter coefficients to filter and reconstruct the farmland remote sensing image to obtain the reconstructed image; calculate the optimal extraction threshold using the Otsu algorithm, and segment the reconstructed image based on the optimal extraction threshold to obtain the target distribution map.

[0070] In this embodiment, after the DTNE-NTZNN solver converges iteratively, the system state solution vector at the current moment will be obtained. The subsequent image processing and detection process is as follows:

[0071] 1) Extract the optimal filter coefficients. From the solution vector Extract the first part as the optimal filter coefficients. When constructing a system of linear equations, unknown variables... Defined by the filter coefficients and Lagrange multipliers The augmented vector formed by concatenation, i.e. Therefore, the "preliminary quantity" here specifically refers to the quantity. The front of the middle elements ( (Represents the number of spectral bands in the remote sensing image).

[0072] 2) Image filtering and reconstruction. Using the extracted coefficients... Perform pixel-by-pixel linear filtering on the current frame of the remote sensing image and calculate the output response value for each pixel. :

[0073]

[0074] in, Represents any one in the remote sensing image The spectral feature vector of each pixel. Using the above formula, multidimensional spectral information is compressed into a single-dimensional response intensity. The energy of non-target background pixels is suppressed, i.e. ; while target pixels that match prior spectral characteristics are significantly enhanced, i.e. .

[0075] 3) Otsu adaptive segmentation. This involves segmenting the filtered and reconstructed grayscale image. The Otsu's maximum inter-class variance algorithm is employed. This algorithm automatically calculates the optimal segmentation threshold, binarizes the image, and generates a target distribution mask, thereby enabling precise localization and detection of specific crops or facilities.

[0076] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0077] experiment

[0078] To verify the effectiveness of the method of this invention in remote sensing target detection in farmland, a series of simulation experiments were conducted using publicly available standard farmland hyperspectral datasets such as Indian Pines, Salinas, and WHU-Hi-LongKou. The DTNE-NTZNN solver used in this invention, through discretization design directly adapting to the digital controller and combined with Nesterov's unique prediction and correction mechanism, represents the optimal solution balancing speed, accuracy, and engineering implementation complexity. The experimental results intuitively demonstrate the accurate detection capability of this invention for single and multiple targets in complex backgrounds.

[0079] See Figure 2 This paper demonstrates the effectiveness of oat detection on the Indian Pines dataset. (a) shows the ground truth map for each type of crop; (b) shows the original image (drawn using the 50th band); (c) shows the ground truth map for oats; (d) shows the detection result; (e) shows the target region; and (f) shows a schematic diagram of the real-time residual of the solution system. Comparing the ground truth map in (c) and the detection result in (d), it can be seen that the method of this invention can accurately extract the oat planting area from the complex farmland background. Furthermore, in the detection result (d), the background is effectively suppressed to black, retaining only the white highlight signal of the target region, and the target region (e) obtained after threshold segmentation highly overlaps with the ground truth.

[0080] See Figure 3 The image demonstrates the effectiveness of celery detection in the Salinas Valley agricultural region. (a) is the original image (plotted using the 50th band); (b) is the ground truth image of celery; (c) is the detection result; and (d) is the target region. The CEM filter derived in this invention successfully suppressed backgrounds from other crops with similar spectra to celery. In the detection result image (c), the target region's brightness is significantly higher than the background, demonstrating the superior resolution of DTNE-NTZNN when processing crops with similar spectral characteristics.

[0081] See Figure 4 This paper demonstrates the simultaneous detection of three different target types—grassland, wheat, and buildings—in the Indian Pines dataset. (a) shows the ground truth map for all types; (b) shows the original image (plotted at band 50); (c) shows the ground truth maps for grassland, wheat, and buildings; (d) shows the detection results; (e) shows the target regions; and (f) shows the real-time residuals of the solution system. Figure (d) shows that the method of this invention successfully highlights three discretely distributed target regions in the same image without interference, indicating that the DTNE-NTZNN solver can quickly converge to meet multiple constraints. .

[0082] See Figure 5This paper demonstrates the simultaneous detection of three widely planted crops in the Salinas dataset. (a) is the original image (drawn using the 50th band); (b) is the ground truth image of weeds and celery; (c) is the detection result; and (d) is the target region. Results (c) and (d) show that even with large and contiguous target regions, the method of this invention maintains the uniformity and integrity of the detection results, without any "holes," indicating that the solved filter has high stability across different pixels.

[0083] See Figure 6 The paper demonstrates the application of this method on the WHU-Hi-LongKou UAV high-resolution hyperspectral dataset, where (a) is the original image (plotted at band 200); (b) is the ground truth image of sesame and rice; (c) is the detection result; and (d) is the target region. For the detection of sesame and rice, the method of this invention not only distinguishes the crop species but also, thanks to the high-precision solution of DTNE-NTZNN, achieves very sharp detection edges (as shown in Figure d), meeting the requirements for fine crop boundary delineation in precision farmland.

[0084] The above are merely preferred embodiments of the present application and are not intended to limit the embodiments of the present application. For those skilled in the art, the embodiments of the present application can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of the present application should be included within the protection scope of the embodiments of the present application.

Claims

1. A method for target detection in farmland remote sensing images based on the DTNE-NTZNN model, characterized in that, Includes the following steps: The acquired farmland remote sensing image data stream was used to construct a farmland remote sensing CEM detection model by constrained energy minimization. The farmland remote sensing CEM detection model is solved iteratively in real time based on the constructed DTNE-NTZNN solver until convergence, and the system state solution vector at the current moment is output. The optimal filter coefficients are extracted based on the system state solution vector. The optimal filter coefficients are then used to filter and reconstruct the farmland remote sensing image to obtain the reconstructed image. The optimal extraction threshold is calculated using the Otsu algorithm. Based on the optimal extraction threshold, the reconstructed image is segmented to obtain the target distribution map.

2. The method for target detection in farmland remote sensing images based on the DTNE-NTZNN model according to claim 1, characterized in that, The farmland remote sensing CEM detection model is constructed using constrained energy minimization, including the following steps: Assuming the acquired farmland remote sensing image data stream is at time... The autocorrelation matrix is The prior spectral feature matrix of the target object to be detected is Based on the constraint energy minimization, a convex quadratic programming problem with linear constraints is established, and the filter coefficients are solved. , making Minimum, ; Transform the problem into a dynamic system of linear equations: ； In the formula, the subscript Represents the iteration round; coefficient matrix ; Unknown variables , including Lagrange multipliers Filter coefficients to be solved ; target vector ; superscript This represents the matrix transpose operation.

3. The method for target detection in farmland remote sensing images based on the DTNE-NTZNN model according to claim 2, characterized in that, The farmland remote sensing CEM detection model is solved in real time using the constructed DTNE-NTZNN solver, including the following steps: Based on the unknown variables obtained from the solution of the previous time step and speed Determine the predicted position at the current moment. The formula is as follows: ； in, These are the Nesterov acceleration parameters; Calculate the predicted location Linear equation error at [location] and integral terms The formula is as follows: ； ； in, For the accumulated integrals from the past, This is the sampling step size; Error of linear equations based on feedback and integral terms Determine intermediate variables Update velocity with momentum based on intermediate variables The formula is as follows: ； ； in, for The false reversal; and ; For proportional gain parameters, This is the integral gain parameter; This is a momentum parameter; Represents the damping coefficient; The solution is updated based on the evolution rate, and the formula is as follows: ； in, For the current solution, To update the solution.

4. The method for target detection in farmland remote sensing images based on the DTNE-NTZNN model according to claim 3, characterized in that, Based on the system state solution vector, the optimal filter coefficients are extracted. The optimal filter coefficients are then used to filter and reconstruct the farmland remote sensing image to obtain the reconstructed image. The process includes the following steps: Extracting the middle and front from the system state solution vector The elements are used as optimal filter coefficients. , This represents the number of spectral bands in the remotely sensed image. The extracted optimal filter coefficients are used to perform pixel-by-pixel linear filtering on the current frame of the remote sensing image, and the output response value of each pixel is calculated. The formula is as follows: ； in, Represents any one in the remote sensing image The spectral feature vector of each pixel.

Images