A method for constructing optimal profile levels and eliminating obstacles for adapting to dry farming in northeast black soil

Abstract

The application provides a method for constructing optimal profile levels and eliminating obstacle layers for the adaptability of dry farming in northeast black land, and relates to the technical field of application of computer vision in agricultural soil analysis. The method comprises the following steps: S1, constructing a standardized profile sampling surface in a target northeast black land dry farming field, and collecting soil profile image data by using a multi-modal imaging device, wherein the multi-modal imaging device comprises at least one of an RGB camera, a near-infrared imaging module or a multispectral imaging module, and is used for acquiring soil profile image information at different wave bands. Through multi-source fusion of the soil profile image and application of a deep learning model, the accuracy of soil structure identification is effectively improved. Through fusion processing of the multi-modal image, different wave band information can be accurately supplemented, and the identification limitation caused by single source of the image in the traditional method is solved.

Description

Technical Field

[0001] This invention relates to the field of computer vision application technology in agricultural soil analysis, specifically to a method for constructing optimal profile layers and eliminating barrier layers to achieve adaptability to dryland farming in Northeast China's black soil. Background Technology

[0002] The application of computer vision in agricultural soil analysis represents a specialized application of computer vision technology in specific agricultural scenarios. It focuses on the comprehensive application of technologies such as image acquisition, image preprocessing, feature extraction, pattern recognition, object detection, semantic segmentation, 3D reconstruction, and intelligent decision-making in soil information analysis. One such method, aimed at achieving optimal profile hierarchy construction and obstacle layer elimination for dryland crop adaptation in Northeast China's black soil, involves intelligently identifying, hierarchically analyzing, and locating obstacle layers in the soil profile using computer vision technology. This is combined with data modeling and optimization algorithms to construct an optimal profile structure parameter model suitable for dryland crop growth, while simultaneously generating targeted obstacle layer elimination schemes. Its purpose is to improve the rationality of the black soil's topsoil structure and root penetration capacity, enhance soil water and fertilizer retention, improve the stability of dryland agricultural production and crop yield, and achieve the sustainable utilization of black soil resources.

[0003] Current technologies for soil profile analysis primarily rely on single image acquisition and processing methods, leading to limitations in soil feature extraction and hierarchical segmentation. First, traditional methods typically limit soil image acquisition to a single wavelength, lacking comprehensive information on different soil layers and easily missing accurate identification of deep barrier layers. Second, image preprocessing and fusion techniques are relatively simple, making it difficult to handle noise or illumination variations in images, resulting in inaccurate identification of soil layer boundaries, especially in complex soil environments, where erroneous segmentation results are prone to occur. Furthermore, existing methods often analyze soil layers based on shallow feature extraction, neglecting the combined analysis of multi-dimensional features such as color, texture, and pore structure, failing to comprehensively reflect the physical properties of the soil and affecting the accuracy of subsequent operational plans. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an optimal method for constructing profile layers and eliminating barrier layers to achieve adaptability to dryland farming in Northeast China's black soil. This method solves the problem that existing technologies mainly rely on a single image acquisition and processing method in soil profile analysis, which leads to certain limitations in soil feature extraction and layer segmentation. Traditional methods often limit the acquisition of soil images to a single band, lack comprehensive information feedback on different soil layers, and are prone to missing the accurate identification of deep barrier layers.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing optimal profile layers and eliminating barrier layers to achieve adaptability to dryland farming in Northeast China's black soil, comprising the following steps:

[0006] S1: Construct a standardized profile sampling surface in the target dryland black soil field in Northeast China, and collect soil profile image data using a multimodal imaging device. The multimodal imaging device includes at least one of an RGB camera, a near-infrared imaging module, or a multispectral imaging module, used to acquire soil profile image information under different wavebands.

[0007] S2: Perform distortion correction, illumination normalization, noise removal and scale alignment on the acquired multimodal images, and fuse the multi-source images based on feature-level fusion algorithm or convolutional neural network fusion model to generate an enhanced comprehensive soil profile image.

[0008] S3: The enhanced soil profile composite image is segmented at the pixel level based on a deep learning semantic segmentation model to identify the boundary positions of the topsoil, plow pan, subsoil and potential barrier layers; at the same time, the color features, texture features and structural features of each soil layer are extracted. The texture features include gray-level co-occurrence matrix features or local binary pattern features, and the structural features include porosity characterization parameters or particle distribution parameters.

[0009] S4: Based on the trained convolutional neural network classification model or object detection model, analyze the segmentation results, identify the physical barrier layer in the soil profile, and output the spatial location, thickness, and compaction level parameters of the barrier layer.

[0010] S5: Construct a standard database containing soil profile data of historical high-yield dryland fields, and use transfer learning or Bayesian optimization algorithms to establish a mapping model between soil profile structure and dryland crop growth indicators, and generate an optimal soil profile hierarchical structure parameter model suitable for the target area. The optimal soil profile hierarchical structure parameter model includes the target thickness range, porosity range and structural level of each soil layer.

[0011] S6: Perform feature matching calculations between the current profile feature parameters extracted in S3 and the optimal soil profile hierarchical structure parameter model constructed in S5, and output the dryland adaptability index and the deviation value of each soil layer.

[0012] S7: Based on the dryland adaptability index and obstacle layer deviation parameters, generate a corresponding obstacle layer elimination operation parameter scheme based on a rule-based reasoning model or a reinforcement learning decision model. The operation parameter scheme includes recommended deep tillage depth, operation frequency or mechanical disturbance intensity.

[0013] S8: After the obstacle layer removal operation is carried out, soil profile images are collected again, and repeated analysis is performed using S2 to S6. The parameters of the deep learning model are updated through the model error backpropagation mechanism to achieve adaptive optimization of the model.

[0014] Preferably, the multimodal imaging device in S1 further includes a depth information acquisition module for acquiring three-dimensional point cloud data of the soil profile; the three-dimensional point cloud data is generated by a stereo vision reconstruction algorithm or a structured light reconstruction algorithm to assist in determining the true spatial morphology of the soil layer and the volume parameters of the barrier layer.

[0015] Preferably, the image fusion processing in S2 adopts a multi-scale feature fusion structure based on convolutional neural networks. By constructing a multi-layer feature pyramid network, the low-level texture features of different modal images are weighted and fused with the high-level semantic features to improve the accuracy of soil boundary recognition.

[0016] Preferably, the deep learning semantic segmentation model in S3 is U-Net, DeepLab, or an improved model thereof, and the ability to identify ambiguous areas of soil layer boundaries is enhanced by introducing an attention mechanism module, so as to improve the segmentation accuracy of the plow pan and dense layer.

[0017] Preferably, the structural feature extraction in S3 further includes calculating a soil compaction index based on image grayscale distribution. The compaction index is calculated by combining a texture direction consistency parameter and pore visibility, and is used to quantitatively characterize the degree of soil compaction.

[0018] Preferably, the obstacle layer identification model in S4 adopts a combination structure of target detection network and classification network, wherein the target detection network is used to locate the obstacle layer region, and the classification network is used to determine the obstacle layer type, which includes plow layer, heavy viscous deposit layer and sandy layer.

[0019] Preferably, the optimal profile hierarchical model construction in S5 uses a Bayesian optimization algorithm to jointly model soil structure parameters and crop yield data from the historical database, and determines the target structure interval of each soil layer by establishing a multivariate response function.

[0020] Preferably, the dryland adaptability index in S6 is obtained by weighted calculation, and the weights include soil layer thickness deviation weight, porosity deviation weight and barrier layer influence weight, wherein the barrier layer influence weight is dynamically adjusted according to the barrier layer depth and continuity.

[0021] Preferably, the reinforcement learning decision model in S7 adopts a policy gradient-based optimization algorithm, which realizes dynamic optimization of obstacle layer removal operation parameters by establishing a reward function between operation parameters and soil structure improvement effect.

[0022] Preferably, the model adaptive optimization in S8 includes constructing a profile change database and updating the weights of the deep learning model through online incremental learning, so that the system can adapt to different soil types and annual farming conditions.

[0023] This invention provides an optimal profile layer construction and barrier layer removal method for achieving dryland farming adaptability in Northeast China's black soil region. It has the following beneficial effects:

[0024] This invention effectively improves the accuracy of soil structure recognition by applying multi-source fusion of soil profile images and deep learning models. Through multimodal image fusion, it accurately supplements information from different wavebands, overcoming the limitations of traditional methods caused by single-source images. The precision of image preprocessing and feature fusion helps eliminate noise and inconsistencies, making soil feature extraction more reliable. The application of a deep learning semantic segmentation model achieves pixel-level accuracy in identifying soil layer boundaries, improving the precise division between cultivated and barrier layers and enabling more multi-dimensional feature extraction, encompassing features such as color, texture, and structure. Automatic identification and quantitative analysis of barrier layers provide reliable data support for subsequent operational parameter optimization. Modeling historical high-yield area data further optimizes the soil profile model, making it more adaptable and predictive, allowing for personalized adjustments based on different soil characteristics. This improves the adaptability of dryland crops and the stability of agricultural production. The introduction of an adaptive optimization mechanism ensures the system can effectively adjust in long-term applications and cope with the challenges of soil structure and environmental changes, achieving continuous optimization. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example:

[0028] like Figure 1 As shown, this embodiment of the invention provides a method for constructing optimal profile layers and eliminating barrier layers to achieve adaptability to dryland farming in Northeast China's black soil region, comprising the following steps:

[0029] S1: Construct a standardized profile sampling surface in the target dryland black soil field in Northeast China, and collect soil profile image data using a multimodal imaging device. The multimodal imaging device includes at least one of an RGB camera, a near-infrared imaging module, or a multispectral imaging module, used to acquire soil profile image information under different wavebands.

[0030] A standardized vertical profile sampling interface was established around dryland fields in the black soil region of Northeast China. Representative field areas were selected, and sampling units were divided according to plot area and cultivation history. For example, several sampling points were arranged in a typical dryland field according to an equidistant grid. Each sampling point was excavated to form a flat profile panel with width and height within the conventional observation range of agricultural soil profiles. Verticality and scale were calibrated using a level and ruler. Scale markings were set on the profile panel for subsequent pixel-to-actual depth conversion. During the imaging stage, an RGB camera was used to acquire visible light images. Near-infrared or multispectral acquisition units are installed, and the shooting distance is controlled within the lens focal length range to stabilize the imaging ratio. By setting the imaging resolution to the number of rows (M) and columns (N), each pixel is mapped to an actual length unit. The mapping relationship can be expressed as Lactual = Lpixel × k, where k is obtained by converting from the scale. For example, scale calibration can be completed when the ratio of the actual length of the scale to the length of the corresponding pixel in the image is used as the scale coefficient. After acquiring multi-band images, the data from different bands are matched according to the timestamp to form a matrix of corresponding images of the same profile in different bands. and ,in Represents the visible light matrix. This represents a near-infrared matrix, where each matrix element... i,j represent the grayscale or reflectance value of the pixel in the i-th row and j-th column. After the band data acquisition is completed, a multi-source soil profile image dataset is formed as the input result for subsequent processing.

[0031] S2: Perform distortion correction, illumination normalization, noise removal and scale alignment on the acquired multimodal images, and fuse the multi-source images based on feature-level fusion algorithm or convolutional neural network fusion model to generate an enhanced comprehensive soil profile image.

[0032] For the acquired multi-source image matrix, lens distortion correction is first performed by establishing a radial distortion model rcorrected = roriginal. Where r represents the distance from the pixel to the principal point, and The pre-calibrated distortion coefficients are solved using a calibration board. Substituting these values ​​into the actual measured pixel radius completes pixel remapping. Then, illumination normalization is performed to calculate the average brightness of the entire image. j, sort the pixel values ​​according to , Target The original line was adjusted linearly. The target is to select the median of the mean range of natural light in the field as the benchmark, and then remove salt-and-pepper noise through median filtering. The window size is taken as an appropriate value within an odd number range. Each pixel is replaced with the median of its sorted neighborhood. In the scale alignment stage, images of different bands are resampled according to the scaling factor k to make the pixel spacing consistent. Then, feature-level fusion is performed, and the visible light feature vector is set. With near-infrared feature vectors A fusion vector is constructed from the pixel grayscale mean g and the texture contrast c. ,in and These are weighting coefficients, and the weights are calculated based on the signal-to-noise ratio (SNR) of each band. For noise, the normalized signal-to-noise ratio is used as the weight allocation ratio, for example, when and When they are on the same order of magnitude, and Values ​​within a relatively balanced range are selected, and after weighting, a comprehensive image matrix is ​​generated. The fusion is used as an enhanced profile data result.

[0033] S3: The enhanced soil profile composite image is segmented at the pixel level based on a deep learning semantic segmentation model to identify the boundary positions of the topsoil, plow pan, subsoil and potential barrier layers; at the same time, the color features, texture features and structural features of each soil layer are extracted. The texture features include gray-level co-occurrence matrix features or local binary pattern features, and the structural features include porosity characterization parameters or particle distribution parameters.

[0034] A pixel-level classification process is constructed based on the fused image matrix. Before inputting images into the semantic segmentation network, sample standardization is performed, and pixel values ​​are classified according to... Normalization is performed, where μ and σ are the mean and standard deviation of the training samples, respectively. The network outputs a probability matrix P, where Pi,j,k represents the probability that the pixel in the i-th row and j-th column belongs to the k-th soil layer. The category is selected according to the maximum probability principle. When Pi,j,k is greater than a preset threshold T, it is determined to be the corresponding soil layer. The threshold T is determined based on the accuracy curve of the validation set, selecting values ​​within the inflection point interval of the curve. After segmenting the topsoil layer from other layers, color features are extracted for each layer, and the mean and standard deviation are calculated. Texture features are analyzed by calculating the contrast using the gray-level co-occurrence matrix. Where m and n are gray levels, Pm,n is the co-occurrence probability, and the porosity in the structural features is calculated as φ = Vpore / Vtotal volume. Vpore is obtained by multiplying the number of dark area pixels in the image by the volume of a single pixel. The particle distribution can be achieved through particle size classification statistics. Various features are combined into vectors for subsequent recognition, and the boundary position and feature parameter output results are completed.

[0035] S4: Based on the trained convolutional neural network classification model or object detection model, analyze the segmentation results, identify the physical barrier layer in the soil profile, and output the spatial location, thickness, and compaction level parameters of the barrier layer.

[0036] A barrier layer identification process is constructed based on the segmentation results. The feature vectors of each layer are input into a classification network to obtain the category probability value Qk. If the compactness index S of a certain region is near the upper limit of a preset interval and the thickness d exceeds the median value of the depth interval, it is determined to be a physical barrier layer. The compactness index S can be represented by a weighted sum of the texture consistency index and the inverse of the porosity. α and β are set based on the regression results of historical samples. The relationship between compactness and measured penetration resistance is fitted in the sample set using the least squares method. The regression coefficient is selected as the proportional coefficient of α and β. The thickness d is obtained by multiplying the pixel difference between the upper and lower boundaries by the proportional coefficient k. The spatial position is represented by the profile depth coordinate. The compactness level is divided into intervals based on the S value. For example, the distribution range of S is divided into several continuous intervals corresponding to different levels. The output results of the barrier layer position, thickness and level parameters are completed.

[0037] S5: Construct a standard database containing soil profile data of historical high-yield dryland fields, and use transfer learning or Bayesian optimization algorithms to establish a mapping model between soil profile structure and dryland crop growth indicators, and generate an optimal soil profile hierarchical structure parameter model suitable for the target area. The optimal soil profile hierarchical structure parameter model includes the target thickness range, porosity range and structural level of each soil layer.

[0038] Establish a database containing historical high-yield field profile data, record soil layer thickness h, porosity φ, and crop yield Y for each sample, and construct a mapping model. Structural index + b, where The regression coefficients are obtained using the least squares method. Let the sample size be n, and construct the matrix equation. X contains the feature data, β is the coefficient vector, and so on. inverse matrix After obtaining the parameters, the current feature values ​​are input into the target region for prediction. During the transfer learning phase, the existing model parameters are fine-tuned, and the loss function is calculated using mini-batch samples. predict The true square is used to update parameters through gradient descent. In the Bayesian optimization stage, the objective function is set to minimize the prediction error. The parameter space range is defined, and the posterior probability is calculated among several candidate parameter combinations. The combination with higher probability is selected to obtain the hierarchical parameter range result applicable to the target region.

[0039] S6: Perform feature matching calculations between the current profile feature parameters extracted in S3 and the optimal soil profile hierarchical structure parameter model constructed in S5, and output the dryland adaptability index and the deviation value of each soil layer.

[0040] The current profile feature vector is matched with the target structural parameter model. Let the current thickness be h, and the target range be from the lower limit of h to the upper limit of h. The deviation is then calculated. now The target median, where h is the average of the upper and lower limits of the interval, and the porosity deviation Δφ is calculated similarly, to construct a comprehensive adaptability index. ,in These are the normalized values ​​for thickness deviation, porosity deviation, and structural grade difference, respectively. The normalization method is xstandard = x - xminimum / xmaximum - xminimum, with weights... Based on the correlation coefficients of historical samples, the index values ​​are calculated by substituting each deviation into the formula, and the deviation results for each layer are output.

[0041] S7: Based on the dryland adaptability index and obstacle layer deviation parameters, generate a corresponding obstacle layer elimination operation parameter scheme based on a rule-based reasoning model or a reinforcement learning decision model. The operation parameter scheme includes recommended deep tillage depth, operation frequency or mechanical disturbance intensity.

[0042] Based on the adaptability index and obstacle layer deviation parameters, a reasoning process for operational parameters is constructed. It is assumed that the deep loosening depth D is correlated with the bottom boundary depth d_bottom of the obstacle layer, and D = d_bottom + Δ_safety, where Δ_safety is an empirical coefficient obtained from the statistical average of historical operational records. The operational frequency F can be calculated as F = γ × deviation amplitude, where γ is a proportionality coefficient obtained through sample regression. The mechanical disturbance intensity I can be mapped to a corresponding interval value based on the compaction level S. The index A is then correlated with the deviation Δh... Substitute into the rule table or reinforcement learning policy function The strategy function is updated based on the reward function R = output change - energy consumption coefficient × energy consumption, and the stable output operation parameter scheme is selected after multiple iterations.

[0043] S8: After the obstacle layer removal operation is carried out, soil profile images are collected again, and repeated analysis is performed using S2 to S6. The parameters of the deep learning model are updated through the model error backpropagation mechanism to achieve adaptive optimization of the model.

[0044] After the operation is completed, profile images are acquired again, and the preprocessing and feature extraction processes are repeated. The error e = Y predicted - Y actual or parameter difference between the new feature vector and the target model is calculated, and the loss function is constructed. Square the gradients of each parameter and calculate them using the backpropagation algorithm. ,according to The model parameters are updated by performing multiple iterations, where η is the learning rate, which is selected within a set range. The iterations continue until the error change is less than a preset convergence threshold, which is determined based on the historical training error distribution.

[0045] The multimodal imaging device in S1 also includes a depth information acquisition module, which is used to acquire three-dimensional point cloud data of the soil profile; the three-dimensional point cloud data is generated by a stereo vision reconstruction algorithm or a structured light reconstruction algorithm, and is used to assist in determining the true spatial morphology of the soil layer and the volume parameters of the barrier layer.

[0046] Based on the addition of spatial information acquisition units to the multimodal imaging device, when deploying binocular camera components or structured light projection components in actual dryland fields in Northeast China's black soil region, the extrinsic and intrinsic parameters of the left and right imaging units are first calibrated. The intrinsic parameter matrix K includes the focal length. and principal point coordinates The extrinsic parameter matrices R and T represent the relationship between rotation and translation. Several sets of images are acquired under different postures using a calibration plate, and the reprojection error is used. Perform calculations, where These are the actual corner coordinates. Using projected coordinates, the calibration is considered valid when the average error is within the lower limit of a preset error range. Subsequently, synchronous acquisition is performed at a baseline distance B in front of the section face to obtain the grayscale matrices of the left and right images. and Through the parallax calculation formula The pixel disparity value is obtained, where and These are the x-coordinates of the corresponding points, and the depth Z is calculated as follows: The calculation is performed, where f is the equivalent focal length, B is the baseline distance, and d is the disparity value. This calculation is applied to each pixel to obtain its three-dimensional coordinates (X, Y, Z). u and v are pixel coordinates, forming a point cloud set. Outlier filtering is performed on the point cloud, and the average distance between each point and its neighborhood is calculated. ,like Values ​​exceeding the upper limit of the interval equal to twice the standard deviation of the neighborhood mean are discarded. The neighborhood radius is set based on the profile size ratio coefficient. The soil surface and lower layer boundary are defined by statistically analyzing the distribution intervals of all point cloud Z values. When the Z gradient change rate of a certain continuous region is... When the value exceeds the upper limit of the set threshold range, it is judged as a morphological abrupt change location, and the volume parameter V is adjusted accordingly. Calculation, where The area of ​​the grid cells is formed by adjacent points. To correspond to the depth difference, the volume of the suspected obstacle area was calculated by dividing it into grids, thus obtaining the true spatial morphology of the soil layer and the volume parameters of the obstacle layer.

[0047] The image fusion processing in S2 adopts a multi-scale feature fusion structure based on convolutional neural networks. By constructing a multi-layer feature pyramid network, the low-level texture features and high-level semantic features of different modal images are weighted and fused to improve the accuracy of soil boundary recognition.

[0048] The construction process of the multi-scale feature fusion structure involves inputting images of different modalities into convolutional layers to obtain multi-layer feature maps before fusion. Where l represents the hierarchical index and m represents the modality category, texture response values ​​are extracted from low-level features. N is the number of pixels in the feature map, representing the semantic response value extracted from high-level features. Similarly, when constructing the feature pyramid, features at different scales are upsampled or downsampled. The scale transformation coefficient r is determined based on the ratio of the original resolution to the target resolution, and the new pixel value is calculated using linear interpolation. Where α is the interpolation weight, then weighted fusion is performed. , Based on the characteristic variance of each layer calculate, When the variance of a certain layer is in the upper limit of the overall variance distribution interval, a larger weight interval value is assigned, and the weight satisfies the following conditions: The specific numerical values ​​are obtained through normalization. In actual calculations, the mean values ​​of lower-level features and higher-level features are selected and substituted into the formula to calculate the fused response value. Boundary gradients are then calculated on the fused result. The gradient is obtained by the difference between adjacent pixels. When the G value falls into the upper segment of the preset boundary judgment interval, it is marked as a candidate boundary region. The interval division is based on the statistical results of the true boundary gradient distribution in historical samples. The fused multi-scale features are output as a comprehensive feature map for subsequent boundary recognition results.

[0049] The deep learning semantic segmentation model in S3 is U-Net, DeepLab, or an improved model thereof, and the ability to identify ambiguous areas of soil layer boundaries is enhanced by introducing an attention mechanism module, so as to improve the segmentation accuracy of the plow pan and dense layer.

[0050] Regarding the specific execution flow of the semantic segmentation network model, image size unification is performed during the sample input stage. Let the original size be H×W, and the target size be... Through the proportionality coefficient Scaling is performed, and pixel resampling uses a bilinear interpolation formula. , Distance weights are used to extract features at the encoding end after entering the U-shaped network. Features generated at the decoding end By skipping connections and splicing Output class probability for each pixel ,in For the k-th linear output value, the cross-entropy loss is applied. Perform calculations. As the true label, during training, iteration stops when the rate of change ΔL of L is less than the lower limit of the set convergence threshold after several iterations. The attention module calculates channel weights on the feature map. σ is the activation function, which weights the features. The system participates in subsequent convolution operations. For the determination of blurred boundary regions, the probability difference Δp = p maximum - p second largest is used. When Δp is in the lower segment of the preset interval, it is classified as an uncertain boundary region and the probability value is updated again by attention weighting. The interval range is obtained based on the statistical distribution of different categories in the validation set. The system completes the output of the plowshare and compact layer segmentation results.

[0051] The structural feature extraction in S3 also includes calculating a soil compaction index based on image grayscale distribution. The compaction index is calculated by combining the texture direction consistency parameter and the pore visibility rate, and is used to quantitatively characterize the degree of soil compaction.

[0052] The soil compaction index is calculated based on the image grayscale distribution and texture orientation consistency parameter. First, the soil image is converted to grayscale to obtain a grayscale matrix. By analyzing the grayscale distribution in the image, the texture orientation consistency parameter is calculated, which quantifies the texture distribution in different directions within the image. The method for calculating texture orientation consistency typically uses a gray-level co-occurrence matrix (GLCM) based on statistical analysis of grayscale values ​​in local image regions. Within each small region, the GLCM calculates the frequency of occurrence of each pair of grayscale values ​​in a specific direction, thus reflecting the image's texture information. For the calculation of porosity visibility, porosity regions are obtained through image segmentation, and the visibility rate is obtained by calculating the ratio of porosity to the total soil area. Subsequently, the texture orientation consistency parameter and porosity visibility are combined using a weighted average to obtain the soil compaction index. Specifically, the texture orientation consistency parameter t is multiplied by the porosity visibility p, and its importance is controlled by setting weights w1 and w2. The formula is expressed as:

[0053] ,

[0054] Where w1 and w2 are the weighting coefficients for texture and porosity, respectively, t is the texture consistency value, and p is the porosity visibility rate. Assuming that in practical applications, weighting coefficients w1=0.6 and w2=0.4 are set, the soil compaction value can be obtained by weighted calculation when t=0.85 and p=0.3. This index can effectively assess the degree of soil compaction.

[0055] The obstacle layer identification model in S4 adopts a combination structure of target detection network and classification network. The target detection network is used to locate the obstacle layer region, and the classification network is used to determine the obstacle layer type, which includes plow layer, heavy clay deposit layer and sandy ginger layer.

[0056] The execution flow of the obstacle layer recognition model first involves the object detection network locating the obstacle layer region. Object detection networks often employ architectures such as Faster R-CNN or YOLO. This network extracts image features using a convolutional neural network (CNN) and then uses a region proposal network (RPN) to locate the obstacle layer region in the image. The key to this process is comparing the image with obstacle layer templates in the training set using a feature matching algorithm to determine the possible obstacle layer location. Specifically, the input image is processed through a convolutional network to generate candidate boxes, which are then filtered using a non-maximum suppression (NMS) algorithm to finally output the precise obstacle layer region. After obstacle layer region localization, a classification network categorizes the identified regions. Classification networks typically use deep neural networks such as ResNet or DenseNet. Through detailed feature learning of local regions, they determine which type of obstacle layer the region belongs to. This determination is based on the unique texture, structure, and other physical characteristics of the obstacle layer. Through model training, the classification network provides a probability value for each region belonging to the plowshare layer, viscous sedimentary layer, or sandy lobe layer, selecting the category with the highest probability value as the region's type. In practical applications, if a target detection network identifies a region, and its classification network labels it as the plow pan, the specific obstacle layer type for that region can be determined. This combination of target detection and classification allows for the accurate identification of different types of obstacle layer regions, aiding in soil assessment and tillage depth analysis.

[0057] The optimal profile hierarchical model in S5 uses a Bayesian optimization algorithm to jointly model soil structure parameters and crop yield data from the historical database, and determines the target structure interval of each soil layer by establishing a multivariate response function.

[0058] The process of constructing the optimal profile hierarchical model first involves the joint modeling of soil structure parameters and crop yield data from a historical database. Bayesian optimization is used to optimize the relationship between these data. Bayesian optimization is a global optimization method based on a probabilistic model. It establishes a surrogate model to represent the unknown parts of the objective function and selects the next parameter combination most likely to improve the objective function value through continuous trial and error. In this application, after combining historical soil data (such as soil clay content, organic matter content, soil layer thickness, etc.) with crop yield data, the Bayesian optimization algorithm generates a multivariate response function based on the input parameter space. This function represents the relationship between soil parameters and crop yield. The specific optimization process involves analyzing historical soil data, calculating the crop yield variation within each soil layer interval, and adjusting the values ​​of each soil layer parameter according to the Bayesian model to find the soil structure parameters that maximize crop yield. The Bayesian optimization process selects the next set of experimental points by calculating the posterior probability of the sampling points, gradually approaching the optimal solution until the set convergence condition is met. In practice, Bayesian optimization algorithms may use Gaussian process regression (GPR) models to predict and optimize soil parameters, and gradually determine the optimal profile hierarchy by evaluating the benefits of each optimization result.

[0059] The dryland adaptability index in S6 is obtained by weighted calculation, and the weights include soil layer thickness deviation weight, porosity deviation weight and barrier layer influence weight, wherein the barrier layer influence weight is dynamically adjusted according to the barrier layer depth and continuity.

[0060] First, measured soil layer thickness data, standard reference thickness range, and barrier layer burial depth and continuity data were collected for the target site. Then, a soil profiler was used to obtain the thickness sequence of each layer. Where i represents the i-th layer, and the thickness deviation is calculated. In the formula The reference thickness for the corresponding soil layer is obtained by using the median range of historical samples within the statistical region. For example, the average thickness within the middle range of the statistical sample is used as the reference thickness. Then the thickness deviation rate was calculated. and will Divided into interval levels, when The lower range is set to a small percentage range, the middle range to a medium percentage range, and the higher range to a large percentage range. Each range corresponds to a different scoring coefficient. Porosity deviation was handled in the same way, and the volumetric mass was measured on-site using the ring cutter method. Then according to the formula Calculate the porosity n, where As a reference value for particle density, the mean of the regional samples is set to be [value missing]. The porosity deviation rate was obtained. It is divided into three levels—low, medium, and high—according to a preset range and assigned accordingly. The determination of the coefficients and the influence weights of the obstacle layer begins with reading the obstacle layer depth. With continuous length Constructing continuity coefficients ,in The length of the quadrat is used, and a larger weighting coefficient is applied when C is in a higher region. When it is in a lower segment, a smaller weighting coefficient is used. Meanwhile, depth coefficient ,in To determine the average safe depth for the region, Multiplying by C yields the comprehensive influence factor of the barrier layer. Ultimately, the thickness deviation weight will be determined. Porosity deviation weight and the influence weight of the barrier layer Perform normalization processing to meet the requirements. Through formula Complete the index calculation, where This represents the dryland adaptability index, and the corresponding index range value is obtained by substituting example plot data into the calculation.

[0061] The reinforcement learning decision model in S7 adopts a policy gradient-based optimization algorithm. By establishing a reward function between the operation parameters and the soil structure improvement effect, the dynamic optimization of the obstacle layer removal operation parameters is achieved.

[0062] First, construct the job parameter set. ,in Indicates the depth of the operation. Indicates the speed of operation. The tool spacing is represented by the soil structure change data Δs recorded under each set of parameters through field experiments, and a reward function is defined. Where Δn represents the change in porosity, Δh represents the change in barrier layer thickness, and E represents the energy consumption per unit area. These are the weighting coefficients, obtained through historical sample regression analysis. For example, the coefficient matrix is ​​solved using the least squares method based on thirty sets of samples to minimize the sum of squared differences between the predicted reward and the actual return. Subsequently, the strategy probability is calculated in each round of decision-making. Where θ is the policy parameter vector, obtained through the gradient formula The update process is performed, where η is the learning step size. Its value is determined by dividing the reward fluctuation range into small step size segments and large step size segments. When the reward fluctuation range is in the high segment, a smaller η is selected, and when the fluctuation range is in the low segment, a larger η is selected. In actual execution, the average value of Δn and Δh is obtained by recording three operations on a certain plot of land and then substituting it into the reward function to calculate the R value. The R value is then substituted into the gradient update formula to complete the iterative adjustment of the parameter θ. After multiple rounds of iteration, a new combination of operation parameters is obtained.

[0063] The model adaptive optimization in S8 includes constructing a profile change database and updating the weights of the deep learning model through online incremental learning, so that the system can adapt to different soil types and annual farming conditions.

[0064] The adaptive optimization process first establishes a profile change database, whose fields include the year identifier 't' and the soil layer thickness sequence. Porosity and barrier layer depth The superscript t represents year t, calculated by differencing data from consecutive years. Obtain the change vector Will As incremental samples, they are input into the deep learning model, denoted as W, and the model weight matrix is ​​updated online using the incremental gradient formula. Where λ is the update coefficient, and its value is divided into intervals based on the sample size. A smaller λ is selected when the proportion of newly added samples in the current year is in a higher interval, and a larger λ is selected when it is in a lower interval. The loss function L is obtained by using the actual predicted values. Compared with measured values Constructing the squared difference between them, i.e. Σ represents the summation over all samples. In the actual execution, several groups of newly added samples in a certain year are selected, their prediction errors are calculated and summed to obtain the L value. Then, the partial derivatives of each element of the weight matrix are calculated and the weights are updated to complete one online iteration. The above process is repeated until the error change rate is within a preset small range, and finally the updated model parameters are obtained.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing optimal profile layers and eliminating barrier layers to achieve adaptability to dryland farming in Northeast China's black soil region, characterized in that, Includes the following steps: S1: Construct a standardized profile sampling surface in the target dryland black soil field in Northeast China, and collect soil profile image data using a multimodal imaging device. The multimodal imaging device includes at least one of an RGB camera, a near-infrared imaging module, or a multispectral imaging module, used to acquire soil profile image information under different wavebands. S2: Perform distortion correction, illumination normalization, noise removal and scale alignment on the acquired multimodal images, and fuse the multi-source images based on feature-level fusion algorithm or convolutional neural network fusion model to generate an enhanced comprehensive soil profile image. S3: The enhanced soil profile composite image is segmented at the pixel level based on a deep learning semantic segmentation model to identify the boundary positions of the topsoil, plow pan, subsoil and potential barrier layers; at the same time, the color features, texture features and structural features of each soil layer are extracted. The texture features include gray-level co-occurrence matrix features or local binary pattern features, and the structural features include porosity characterization parameters or particle distribution parameters. S4: Based on the trained convolutional neural network classification model or object detection model, analyze the segmentation results, identify the physical barrier layer in the soil profile, and output the spatial location, thickness, and compaction level parameters of the barrier layer. S5: Construct a standard database containing soil profile data of historical high-yield dryland fields, and use transfer learning or Bayesian optimization algorithms to establish a mapping model between soil profile structure and dryland crop growth indicators, and generate an optimal soil profile hierarchical structure parameter model suitable for the target area. The optimal soil profile hierarchical structure parameter model includes the target thickness range, porosity range and structural level of each soil layer. S6: Perform feature matching calculations between the current profile feature parameters extracted in S3 and the optimal soil profile hierarchical structure parameter model constructed in S5, and output the dryland adaptability index and the deviation value of each soil layer. S7: Based on the dryland adaptability index and obstacle layer deviation parameters, generate a corresponding obstacle layer elimination operation parameter scheme based on a rule-based reasoning model or a reinforcement learning decision model. The operation parameter scheme includes recommended deep tillage depth, operation frequency or mechanical disturbance intensity. S8: After the obstacle layer removal operation is carried out, soil profile images are collected again, and repeated analysis is performed using S2 to S6. The parameters of the deep learning model are updated through the model error backpropagation mechanism to achieve adaptive optimization of the model.

2. The method for constructing optimal profile layers and eliminating barrier layers to achieve adaptability to dryland farming in Northeast China's black soil, as described in claim 1, is characterized in that: The multimodal imaging device in S1 also includes a depth information acquisition module, which is used to acquire three-dimensional point cloud data of the soil profile; the three-dimensional point cloud data is generated by a stereo vision reconstruction algorithm or a structured light reconstruction algorithm, and is used to assist in determining the true spatial morphology of the soil layer and the volume parameters of the barrier layer.

3. The method for constructing optimal profile layers and eliminating barrier layers to achieve adaptability to dryland farming in Northeast China's black soil, as described in claim 1, is characterized in that: The image fusion processing in S2 adopts a multi-scale feature fusion structure based on convolutional neural networks. By constructing a multi-layer feature pyramid network, the low-level texture features and high-level semantic features of different modal images are weighted and fused to improve the accuracy of soil boundary recognition.

4. The method for constructing optimal profile layers and eliminating barrier layers to achieve adaptability to dryland farming in Northeast China's black soil, as described in claim 1, is characterized in that: The deep learning semantic segmentation model in S3 is U-Net, DeepLab, or an improved model thereof, and the ability to identify ambiguous areas of soil layer boundaries is enhanced by introducing an attention mechanism module, so as to improve the segmentation accuracy of the plow pan and dense layer.

5. The method for constructing optimal profile layers and eliminating barrier layers to achieve adaptability to dryland farming in Northeast China's black soil, as described in claim 1, is characterized in that: The structural feature extraction in S3 also includes calculating a soil compaction index based on image grayscale distribution. The compaction index is calculated by combining the texture direction consistency parameter and the pore visibility rate, and is used to quantitatively characterize the degree of soil compaction.

6. The method for constructing optimal profile layers and eliminating barrier layers to achieve adaptability to dryland farming in Northeast China's black soil, as described in claim 1, is characterized in that: The obstacle layer identification model in S4 adopts a combination structure of target detection network and classification network. The target detection network is used to locate the obstacle layer region, and the classification network is used to determine the obstacle layer type, which includes plow layer, heavy clay deposit layer and sandy ginger layer.

7. The method for constructing optimal profile layers and eliminating barrier layers to achieve dryland farming adaptability in Northeast China's black soil, as described in claim 1, is characterized in that: The optimal profile hierarchical model in S5 uses a Bayesian optimization algorithm to jointly model soil structure parameters and crop yield data from the historical database, and determines the target structure interval of each soil layer by establishing a multivariate response function.

8. The method for constructing optimal profile layers and eliminating barrier layers to achieve dryland farming adaptability in Northeast China's black soil, as described in claim 1, is characterized in that: The dryland adaptability index in S6 is obtained by weighted calculation, and the weights include soil layer thickness deviation weight, porosity deviation weight and barrier layer influence weight, wherein the barrier layer influence weight is dynamically adjusted according to the barrier layer depth and continuity.

9. The method for constructing optimal profile layers and eliminating barrier layers to achieve adaptability to dryland farming in Northeast China's black soil, as described in claim 1, is characterized in that: The reinforcement learning decision model in S7 adopts a policy gradient-based optimization algorithm. By establishing a reward function between the operation parameters and the soil structure improvement effect, the dynamic optimization of the obstacle layer removal operation parameters is achieved.

10. The method for constructing optimal profile layers and eliminating barrier layers to achieve adaptability to dryland farming in Northeast China's black soil, as described in claim 1, is characterized in that: The model adaptive optimization in S8 includes constructing a profile change database and updating the weights of the deep learning model through online incremental learning, so that the system can adapt to different soil types and annual farming conditions.

Images