A method for identifying interplanted trees on cultivated land combining mixed pixel decomposition and deep learning

Abstract

The present invention discloses a method for identifying interplanted trees on cultivated land that combines hybrid pixel decomposition and deep learning, comprising the following steps: generating basic classification data by processing remote sensing image data; establishing sample labels for interplanted trees on cultivated land and constructing standardized sample sets based on historical monitoring data and classification standards; using hybrid pixel decomposition technology to calculate endmember abundance and cross-endmember contribution ratio, and completing preliminary classification in combination with regional area; designing a deep learning network architecture, using multi-scale and directional feature extraction technology to train and optimize the model; automatically extracting spectral, texture and spatial features through the model, and fusing and generating classification results; using a validation set and a test set to evaluate the performance of the model, and optimizing classification parameters. The present invention realizes the accurate classification and identification of interplanted trees on cultivated land.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, and in particular to a method for identifying interplanted trees on cultivated land by combining mixed pixel decomposition and deep learning. Background Art

[0002] With the rapid development of remote sensing technology and geographic information systems, farmland surveys have become an important means of land resource management. Especially in land change surveys, the phenomenon of intercropping of trees on farmland has gradually attracted attention. Intercropping of trees on farmland is a complex type of land feature, involving the intercropping and interplanting of fruit trees or trees with crops. This phenomenon not only affects the accurate identification of farmland, but may also lead to the potential loss of farmland resources, directly threatening the implementation of the farmland protection red line and the implementation of the strictest farmland protection system.

[0003] At present, the investigation and determination of the phenomenon of interplanting trees on cultivated land mainly rely on traditional remote sensing image interpretation methods and manual field investigations. These methods have the following obvious defects in practical applications:

[0004] 1. High labor input and low efficiency: Traditional methods require a large amount of manpower to interpret remote sensing images and conduct field surveys, which is time-consuming and labor-intensive, and its efficiency cannot meet the needs of large-scale farmland surveys.

[0005] 2. Lack of objectivity and stability: The remote sensing image interpretation process relies on human subjective judgment. The judgment results are easily affected by the experience and knowledge level of the interpreters. There is a lack of unified standards, resulting in poor stability of the judgment results.

[0006] 3. Low recognition accuracy: Traditional recognition methods for intercropping of trees on cultivated land lack quantitative analysis capabilities for key indicators such as fruit tree coverage and tree density. The recognition results have large errors and are difficult to meet the accuracy requirements of cultivated land monitoring tasks.

[0007] 4. Difficulty in dealing with complex scenarios: Traditional methods have obvious limitations when faced with the diverse scenarios and complex spectral characteristics of intercropping trees on cultivated land, and are unable to effectively distinguish between cultivated land, gardens, and forests.

[0008] Therefore, how to provide a method for identifying interplanted trees on cultivated land that combines mixed pixel decomposition and deep learning is an urgent problem that technicians in this field need to solve. Summary of the invention

[0009] One purpose of the present invention is to propose a method for identifying interplanted trees on cultivated land that combines hybrid pixel decomposition and deep learning. The present invention enhances sample data by performing hybrid pixel decomposition on remote sensing images, dynamically adjusts sample optimization model performance in deep learning model training, and provides comparative verification in recognition accuracy evaluation. It has the advantages of low labor cost, high recognition accuracy and adaptability to complex scenarios.

[0010] A method for identifying interplanted trees on cultivated land combining mixed pixel decomposition and deep learning according to an embodiment of the present invention comprises the following steps:

[0011] S1. Obtain remote sensing image data and historical monitoring results, perform orthorectification, geometric correction, radiation atmospheric correction and panchromatic multispectral fusion processing on remote sensing images, and combine them with cultivated land range vector data to generate basic classification image data;

[0012] S2. Based on the generated basic classified image data and the classification standards of historical monitoring results, establish the sample labels of interplanted trees on cultivated land, and classify the ground samples into cultivated land samples, garden samples and forest samples according to the fruit tree coverage, number of fruit trees, tree canopy density and planting density parameters to generate a standardized sample set;

[0013] S3. Based on the mixed pixel decomposition technology, the nonlinear mixed pixel decomposition model is used to invert the end-member abundance of ground objects, and the number of seedlings planted and the forest density in the area are calculated in combination with the ground object abundance. The standardized sample set is preliminarily determined and divided into cultivated land, garden land and forest land categories;

[0014] S4. Build a deep neural network remote sensing image interpretation model, design a deep learning network architecture suitable for the multi-scale and directional variation characteristics of intercropping forest objects on cultivated land, use standardized sample sets to train the model and optimize parameters, and optimize the model classification ability through fully supervised learning methods;

[0015] S5. Use the deep neural network remote sensing image interpretation model to automatically extract the spectral, texture and spatial features of remote sensing images, fuse features at different levels, generate classification results for cultivated land, gardens and forests, and use validation and test sets to evaluate the model performance;

[0016] S6. Cross-validate the classification results of the deep neural network remote sensing image interpretation model with the mixed pixel decomposition classification results, compare the classification consistency and key indicators, and optimize the standardized sample set and the deep neural network remote sensing image interpretation model.

[0017] Optionally, the S2 specifically includes:

[0018] S21. Extract classification standards from historical monitoring results, and define classification rules for cultivated land samples, garden samples and forest samples based on fruit tree coverage, number of fruit trees, tree canopy density and planting density parameters, including:

[0019] Cultivated land samples: fruit tree coverage ≤ 50% or the number of fruit trees per mu ≤ 70% of the reasonable number of trees, tree density ≤ 0.2;

[0020] Garden samples: fruit tree coverage > 50% or the number of fruit trees per mu > 70% of the reasonable number;

[0021] Forest samples: tree canopy density> 0.2;

[0022] S22, extracting preliminary classification information based on the spectral characteristics, geometric characteristics and texture characteristics of each ground object target in the basic classification image data and combining the classification standard, and judging whether the target meets the classification conditions of cultivated land samples, garden samples and forest samples;

[0023] S23. For targets whose spectral characteristics cannot be clearly classified, the fruit tree coverage, number of fruit trees per mu, tree density and planting density parameters of each ground object are calculated based on the pixel-level analysis results of remote sensing images:

[0024] ;

[0025] in, is the fruit tree coverage, The area occupied by fruit trees, is the total target area;

[0026] S24, according to the reasonable number of trees in the classification conditions, combined with the land area and planting density parameters of the land object, calculate the number of fruit trees per mu, and confirm whether the land object meets the classification conditions of cultivated land samples and garden samples;

[0027] S25. For the ground objects with tree vegetation, the tree canopy density information is inverted through the tree end members in the image, the tree canopy density is calculated, and the classification conditions are used to determine whether the target meets the classification conditions of cultivated land samples and forest samples;

[0028] S26. Assign corresponding labels to the ground objects that meet the classification conditions, classify the sample objects into cultivated land samples, garden samples and forest samples, generate a standardized sample data set containing classification labels, and associate the classification information with the image feature data for storage.

[0029] Optionally, the S3 specifically includes:

[0030] S31. Based on the mixed pixel decomposition technology, a nonlinear mixed pixel decomposition model is constructed to describe the relationship between the spectral reflectance of the observed pixel and the endmember abundance:

[0031] ;

[0032] in, represents the observed mixed spectrum, Indicates The abundance ratio of the end member, For the Spectral reflectance of the end member, For the Secondary scattering contribution ratio between end members, is the cross-endmember reflectivity product, is the error term;

[0033] S32. Based on remote sensing image data, the pure pixel index method is used to extract endmember spectral data, including illuminated canopy endmembers, illuminated surface endmembers and shadow endmembers, so that the extracted endmember spectra are representative and stable for constructing nonlinear mixed models.

[0034] S33, using the least squares method to invert the mixed spectrum and calculate the end member abundance, applying full constraints and non-negativity conditions in the inversion process, the constraints are:

[0035] ;

[0036] in, Indicates The abundance ratio of end members, is the secondary scattering contribution ratio of the cross-end members, so that the abundance ratio of all end members is non-negative and the sum is 1;

[0037] S34. Calculate the number of seedlings to be planted based on the end member abundance values ​​obtained by inversion and the target area and the area occupied by a single tree.

[0038] ;

[0039] in, is the total number of seedlings, is the total area of ​​the target region, is the area occupied by a single plant, For the The abundance ratio of end members, For the Contribution coefficient of the class end member;

[0040] S35. By using the canopy closure inversion algorithm of the geometric optical model, the end-member abundance of the illumination surface is combined with the solar zenith angle, the observation zenith angle, the solar azimuth angle and the geometric information of the observation azimuth angle, and the average tree height and crown radius of the dominant tree species are input to calculate the canopy closure distribution in the area;

[0041] S36. Based on the calculated number of saplings planted and canopy density distribution, the categories of the land objects in the standardized sample set are determined, and the objects are divided into cultivated land, garden land and forest land according to the classification rules.

[0042] Optionally, the S31 specifically includes:

[0043] S311. Based on the mixed pixel decomposition technology, a nonlinear mixed pixel decomposition model is constructed to describe the relationship between the spectral reflectance of the observed pixel and the end member abundance. The model includes a linear reflection component and a nonlinear scattering component.

[0044] S312. Define the linear reflection component, which is used to characterize the direct contribution of each end member to the observed spectrum. The linear reflection component is represented by the product of the end member abundance ratio and the end member spectral reflectance:

[0045] ;

[0046] in, represents the linear reflection component, For the The abundance ratio of end members, For the Spectral reflectance of the endmember-like structure;

[0047] S313. Define the nonlinear scattering component, which is used to describe the effect of the interaction between end members on the spectrum. The nonlinear scattering component is represented by the cross product of the cross-end member contribution ratio and the end member spectral reflectance:

[0048] ;

[0049] in, represents the effect of nonlinear interactions between end members on the mixed spectrum, For the The cross-contribution ratio between the two types of end members, is the product of the spectral reflectance of the two end elements;

[0050] S314. Combining the linear reflection component and the nonlinear scattering component, a nonlinear mixed pixel decomposition model is constructed, and the observed spectrum is described by the following expression:

[0051] ;

[0052] in, To observe the mixed spectrum, is the spectral mixing function, represents the abundance ratio of the 1st to the nth end members. The abundance of each end member reflects its relative contribution in the observed spectrum. It represents the spectral reflectance of each end member and describes the optical properties of the end member.

[0053] Optionally, the S33 specifically includes:

[0054] S331, inverting the mixed spectrum using the least squares method, and calculating the abundance ratio of each end member in the mixed pixel and the secondary scattering contribution ratio of the cross end member by minimizing the error between the observed spectrum and the model predicted spectrum;

[0055] S332. Full constraints are imposed during the inversion process so that the sum of the linear endmember abundance and the cross-endmember contribution ratio is equal to 1. The constraints are defined as:

[0056] ;

[0057] S333. At the same time, the non-negative condition is applied to restrict the values ​​of all end-member abundance ratios and cross-end-member contribution ratios to be non-negative. The non-negative condition is defined as:

[0058] ;

[0059] S334. On the basis of satisfying the above-mentioned full constraint conditions and non-negativity conditions, the endmember abundance ratio and the cross-endmember contribution ratio are updated through an iterative optimization algorithm until the error meets the convergence condition, and finally the endmember abundance ratio and the cross-endmember contribution ratio are determined.

[0060] Optionally, S4 specifically includes:

[0061] S41. Design a deep learning network architecture, select a convolutional neural network based on the scale change and directional characteristics of the intercropping of trees and ground objects on cultivated land, and develop an adaptive network structure based on remote sensing image classification tasks;

[0062] S42, multi-scale feature extraction is performed by setting multi-layer convolution operations and convolution kernels of different sizes to complete the feature extraction process of the ground object target;

[0063] S43. Configure a direction-sensitive model in the network architecture, combine direction encoding technology and rotation-invariant convolution operations, extract the direction characteristics of ground objects, and effectively capture direction information;

[0064] S44. Use the standardized sample set generated by mixed pixel decomposition to train the network model, adopt a fully supervised learning method, and calculate the cross entropy loss function through sample labels:

[0065] ;

[0066] in, is the loss value, is the total number of samples, For the The true labels of samples, The classification probability predicted by the network model;

[0067] S45. Optimize the network parameters through the back propagation algorithm, use the gradient descent method to minimize the loss function, complete multiple iterative training of the model, input the mixed pixel decomposition results to adjust the categories of the standardized sample set, and optimize the accuracy of category recognition.

[0068] Optionally, S5 specifically includes:

[0069] S51. Perform multi-channel processing on remote sensing images through a deep neural network remote sensing image interpretation model to extract a basic feature set including spectral features, texture features and spatial features;

[0070] S52. Use the convolutional layer operation of the deep neural network to extract spectral features, analyze the spectral response values ​​of each band of the remote sensing image, and generate a spectral feature map reflecting the spectral differences of cultivated land, gardens and forests. ;

[0071] S53, combine the gray level co-occurrence matrix method to calculate the texture features of the image, extract the surface roughness, directional consistency and texture uniformity of the object through window sliding operation, and generate a texture feature map ;

[0072] S54, extract spatial features based on the spatial attention mechanism of deep learning network, and generate spatial feature maps by capturing the spatial distribution pattern of ground objects and the correlation between adjacent pixels. ;

[0073] S55, will , and Perform multi-layer fusion operations to construct comprehensive feature representation , and input the fused features into the classification layer to generate the classification results of cultivated land, garden land and forest land;

[0074] S56. Evaluate the classification performance based on the validation set and calculate the precision, recall and F1 value of the classification model:

[0075] ;

[0076] ;

[0077] ;

[0078] in, is the accuracy, is the recall rate, It is the harmonic mean of precision and recall, with a maximum value of 1 and a minimum value of 0. represents the number of true positive examples, represents the number of false positive examples, represents the number of false negative examples;

[0079] S57. Adjust model parameters based on the performance evaluation results of the validation set to optimize the comprehensive feature representation Finally, the test set is used to verify the robustness and overall accuracy of the classification model.

[0080] The beneficial effects of the present invention are:

[0081] The present invention optimizes and enhances the sample data of intercropping of cultivated land with trees in remote sensing images by combining hybrid pixel decomposition technology and deep learning algorithm, realizes sub-pixel level target detection and refined analysis, significantly improves the classification accuracy and accuracy of remote sensing image interpretation, and solves the classification difficulty problem of traditional methods in complex scenes.

[0082] The expansion of sample data based on hybrid pixel decomposition technology has enhanced the stability and consistency of the judgment of interplanted forest objects on cultivated land. Compared with traditional manual interpretation and field survey methods, it has greatly improved the classification efficiency and work accuracy, and reduced the workload and the need for manual intervention.

[0083] The present invention improves the ability to recognize targets of different categories by dynamically adjusting sample data during the deep learning model training process and combining comparative verification of classification results, enabling the model to adapt to diverse remote sensing image scenes and significantly optimizing the robustness and reliability of cultivated land, garden land and woodland classification tasks.

[0084] The comprehensive land object recognition method that combines hybrid pixel decomposition and deep learning algorithm effectively reduces labor input and human interference, reduces monitoring costs, and at the same time meets the high-precision and low-cost technical requirements of farmland protection tasks, significantly improving monitoring efficiency and the application value of classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0086] Figure 1 A flowchart of a method for identifying interplanted trees on cultivated land that combines mixed pixel decomposition and deep learning proposed by the present invention;

[0087] Figure 2 This is a flow chart of remote sensing interpretation of interplanting trees on cultivated land by combining hybrid pixel decomposition with deep learning algorithm. DETAILED DESCRIPTION

[0088] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0089] refer to Figure 1-2 , a method for identifying interplanted trees on cultivated land combining mixed pixel decomposition and deep learning, comprising the following steps:

[0090] S1. Obtain remote sensing image data and historical monitoring results, perform orthorectification, geometric correction, radiation atmospheric correction and panchromatic multispectral fusion processing on remote sensing images, and combine them with cultivated land range vector data to generate basic classification image data;

[0091] S2. Based on the generated basic classified image data and the classification standards of historical monitoring results, establish the sample labels of interplanted trees on cultivated land, and classify the ground samples into cultivated land samples, garden samples and forest samples according to the fruit tree coverage, number of fruit trees, tree canopy density and planting density parameters to generate a standardized sample set;

[0092] S3. Based on the mixed pixel decomposition technology, the nonlinear mixed pixel decomposition model is used to invert the end-member abundance of ground objects, and the number of seedlings planted and the forest density in the area are calculated in combination with the ground object abundance. The standardized sample set is preliminarily determined and divided into cultivated land, garden land and forest land categories;

[0093] S4. Build a deep neural network remote sensing image interpretation model, design a deep learning network architecture suitable for the multi-scale and directional variation characteristics of intercropping forest objects on cultivated land, use standardized sample sets to train the model and optimize parameters, and optimize the model classification ability through fully supervised learning methods;

[0094] S5. Use the deep neural network remote sensing image interpretation model to automatically extract the spectral, texture and spatial features of remote sensing images, fuse features at different levels, generate classification results for cultivated land, gardens and forests, and use validation and test sets to evaluate the model performance;

[0095] S6. Cross-validate the classification results of the deep neural network remote sensing image interpretation model with the mixed pixel decomposition classification results, compare the classification consistency and key indicators, and optimize the standardized sample set and the deep neural network remote sensing image interpretation model.

[0096] In this implementation, S2 specifically includes:

[0097] S21. Extract classification standards from historical monitoring results, and define classification rules for cultivated land samples, garden samples and forest samples based on fruit tree coverage, number of fruit trees, tree canopy density and planting density parameters, including:

[0098] Cultivated land samples: fruit tree coverage ≤ 50% or the number of fruit trees per mu ≤ 70% of the reasonable number of trees, tree density ≤ 0.2;

[0099] Garden samples: fruit tree coverage > 50% or the number of fruit trees per mu > 70% of the reasonable number;

[0100] Forest samples: tree canopy density> 0.2;

[0101] S22, extracting preliminary classification information based on the spectral characteristics, geometric characteristics and texture characteristics of each ground object target in the basic classification image data and combining the classification standard, and judging whether the target meets the classification conditions of cultivated land samples, garden samples and forest samples;

[0102] S23. For targets whose spectral characteristics cannot be clearly classified, the fruit tree coverage, number of fruit trees per mu, tree density and planting density parameters of each ground object are calculated based on the pixel-level analysis results of remote sensing images:

[0103] ;

[0104] in, is the fruit tree coverage, The area occupied by fruit trees, is the total target area;

[0105] S24, according to the reasonable number of trees in the classification conditions, combined with the land area and planting density parameters of the land object, calculate the number of fruit trees per mu, and confirm whether the land object meets the classification conditions of cultivated land samples and garden samples;

[0106] S25. For the ground objects with tree vegetation, the tree canopy density information is inverted through the tree end members in the image, the tree canopy density is calculated, and the classification conditions are used to determine whether the target meets the classification conditions of cultivated land samples and forest samples;

[0107] S26. Assign corresponding labels to the ground objects that meet the classification conditions, classify the sample objects into cultivated land samples, garden samples and forest samples, generate a standardized sample data set containing classification labels, and associate the classification information with the image feature data for storage.

[0108] In this implementation, S3 specifically includes:

[0109] S31. Based on the mixed pixel decomposition technology, a nonlinear mixed pixel decomposition model is constructed to describe the relationship between the spectral reflectance of the observed pixel and the endmember abundance:

[0110] ;

[0111] in, represents the observed mixed spectrum, Indicates The abundance ratio of end members, For the Spectral reflectance of the end member, For the Secondary scattering contribution ratio between end members, is the cross-endmember reflectivity product, is the error term;

[0112] S32. Based on remote sensing image data, the pure pixel index method is used to extract endmember spectral data, including illuminated canopy endmembers, illuminated surface endmembers and shadow endmembers, so that the extracted endmember spectra are representative and stable for constructing nonlinear mixed models.

[0113] S33, using the least squares method to invert the mixed spectrum and calculate the end member abundance, applying full constraints and non-negativity conditions in the inversion process, the constraints are:

[0114] ;

[0115] in, Indicates The abundance ratio of the end member, is the secondary scattering contribution ratio of the cross-end members, so that the abundance ratio of all end members is non-negative and the sum is 1;

[0116] S34. Calculate the number of seedlings to be planted based on the end member abundance values ​​obtained by inversion and the target area and the area occupied by a single tree.

[0117] ;

[0118] in, is the total number of seedlings, is the total area of ​​the target region, is the area occupied by a single plant, For the The abundance ratio of the end member, For the Contribution coefficient of the class end member;

[0119] S35. By using the canopy closure inversion algorithm of the geometric optical model, the end-member abundance of the illumination surface is combined with the solar zenith angle, the observation zenith angle, the solar azimuth angle and the geometric information of the observation azimuth angle, and the average tree height and crown radius of the dominant tree species are input to calculate the canopy closure distribution in the area;

[0120] S36. Based on the calculated number of saplings planted and canopy density distribution, the categories of the land objects in the standardized sample set are determined, and the objects are divided into cultivated land, garden land and forest land according to the classification rules.

[0121] In this implementation manner, the S31 specifically includes:

[0122] S311. Based on the mixed pixel decomposition technology, a nonlinear mixed pixel decomposition model is constructed to describe the relationship between the spectral reflectance of the observed pixel and the end member abundance. The model includes a linear reflection component and a nonlinear scattering component.

[0123] S312. Define the linear reflection component, which is used to characterize the direct contribution of each end member to the observed spectrum. The linear reflection component is represented by the product of the end member abundance ratio and the end member spectral reflectance:

[0124] ;

[0125] in, represents the linear reflection component, For the The abundance ratio of the end member, For the Spectral reflectance of the endmember-like structure;

[0126] S313. Define the nonlinear scattering component, which is used to describe the effect of the interaction between end members on the spectrum. The nonlinear scattering component is represented by the cross product of the cross-end member contribution ratio and the end member spectral reflectance:

[0127] ;

[0128] in, represents the effect of nonlinear interactions between end members on the mixed spectrum, For the The cross-contribution ratio between the two types of end members, is the product of the spectral reflectance of the two end elements;

[0129] S314. Combining the linear reflection component and the nonlinear scattering component, a nonlinear mixed pixel decomposition model is constructed, and the observed spectrum is described by the following expression:

[0130] ;

[0131] in, To observe the mixed spectrum, is the spectral mixing function, represents the abundance ratio of the 1st to the nth end members. The abundance of each end member reflects its relative contribution in the observed spectrum. It represents the spectral reflectance of each end member and describes the optical properties of the end member.

[0132] In this implementation manner, the S33 specifically includes:

[0133] S331, inverting the mixed spectrum using the least squares method, and calculating the abundance ratio of each end member in the mixed pixel and the secondary scattering contribution ratio of the cross end member by minimizing the error between the observed spectrum and the model predicted spectrum;

[0134] S332. Full constraints are imposed during the inversion process so that the sum of the linear endmember abundance and the cross-endmember contribution ratio is equal to 1. The constraints are defined as:

[0135] ;

[0136] S333. At the same time, the non-negative condition is applied to restrict the values ​​of all end-member abundance ratios and cross-end-member contribution ratios to be non-negative. The non-negative condition is defined as:

[0137] ;

[0138] S334. On the basis of satisfying the above-mentioned full constraint conditions and non-negativity conditions, the endmember abundance ratio and the cross-endmember contribution ratio are updated through an iterative optimization algorithm until the error meets the convergence condition, and finally the endmember abundance ratio and the cross-endmember contribution ratio are determined.

[0139] In this implementation, S4 specifically includes:

[0140] S41. Design a deep learning network architecture, select a convolutional neural network based on the scale change and directional characteristics of the intercropping of trees and ground objects on cultivated land, and develop an adaptive network structure based on remote sensing image classification tasks;

[0141] S42, multi-scale feature extraction is performed by setting multi-layer convolution operations and convolution kernels of different sizes to complete the feature extraction process of the ground object target;

[0142] S43. Configure a direction-sensitive model in the network architecture, combine direction encoding technology and rotation-invariant convolution operations, extract the direction characteristics of ground objects, and effectively capture direction information;

[0143] S44. Use the standardized sample set generated by mixed pixel decomposition to train the network model, adopt a fully supervised learning method, and calculate the cross entropy loss function through sample labels:

[0144] ;

[0145] in, is the loss value, is the total number of samples, For the The true labels of samples, The classification probability predicted by the network model;

[0146] S45. Optimize the network parameters through the back propagation algorithm, use the gradient descent method to minimize the loss function, complete multiple iterative training of the model, input the mixed pixel decomposition results to adjust the categories of the standardized sample set, and optimize the accuracy of category recognition.

[0147] In this implementation, S5 specifically includes:

[0148] S51. Perform multi-channel processing on remote sensing images through a deep neural network remote sensing image interpretation model to extract a basic feature set including spectral features, texture features and spatial features;

[0149] S52. Use the convolutional layer operation of the deep neural network to extract spectral features, analyze the spectral response values ​​of each band of the remote sensing image, and generate a spectral feature map reflecting the spectral differences of cultivated land, gardens and forests. ;

[0150] S53, combine the gray level co-occurrence matrix method to calculate the texture features of the image, extract the surface roughness, directional consistency and texture uniformity of the object through window sliding operation, and generate a texture feature map ;

[0151] S54, extract spatial features based on the spatial attention mechanism of deep learning network, and generate spatial feature maps by capturing the spatial distribution pattern of ground objects and the correlation between adjacent pixels. ;

[0152] S55, will , and Perform multi-layer fusion operations to construct comprehensive feature representation , and input the fused features into the classification layer to generate the classification results of cultivated land, garden land and forest land;

[0153] S56. Evaluate the classification performance based on the validation set and calculate the precision, recall and F1 value of the classification model:

[0154] ;

[0155] ;

[0156] ;

[0157] in, is the accuracy, is the recall rate, It is the harmonic mean of precision and recall, with a maximum value of 1 and a minimum value of 0. represents the number of true positive examples, represents the number of false positive examples, represents the number of false negative examples;

[0158] S57. Adjust model parameters based on the performance evaluation results of the validation set to optimize the comprehensive feature representation Finally, the test set is used to verify the robustness and overall accuracy of the classification model.

[0159] Embodiment 1:

[0160] In order to verify the actual application effect of the present invention in the identification of interplanted trees on cultivated land, the present invention was applied to a cultivated land protection and investigation project carried out by a provincial land and resources management department. The project covers a total area of ​​about 10,000 square kilometers, and the main task is to accurately identify and classify the phenomenon of interplanted trees on cultivated land. The terrain in this area is complex, and the interplanting of trees and crops is diverse. Traditional manual field investigations and visual interpretation are inefficient and cannot meet the needs of large-scale investigations.

[0161] In this project, the present invention first preprocesses the multi-source remote sensing images acquired in the region, including geometric correction, radiation correction and multispectral fusion processing, to generate basic classified image data for subsequent analysis. Then, based on the hybrid pixel decomposition technology, the end member spectral information and abundance distribution of the image are extracted to enhance the sample data and form a standardized sample set. This process significantly improves the representativeness and classification accuracy of the sample data, and provides a reliable data foundation for the subsequent training of the deep learning model.

[0162] The present invention designs a classification model in combination with a deep learning algorithm. The model uses a multi-scale convolution kernel to extract the multi-scale features of the interplanted tree targets on cultivated land, and uses a direction-sensitive module to capture the directional distribution information of the targets. The network weights are optimized through the cross-entropy loss function to complete the parameter tuning of the model, thereby generating accurate classification results for cultivated land, gardens, and forests.

[0163] The classification results were cross-validated by mixed pixel decomposition and the performance of the validation set was evaluated. The following are the specific experimental data and validation results:

[0164] Table 1 Comparison of remote sensing image classification results

[0165]

[0166] The classification results show that the method of the present invention has a high accuracy rate in the classification of cultivated land, garden land and woodland, especially in the classification of cultivated land, reaching a recognition accuracy rate of 99%.

[0167] The present invention also verifies the performance of the model through multiple indicators. The model performance is comprehensively evaluated through precision, recall and F1 value:

[0168] Table 2 Model performance evaluation

[0169]

[0170] It has been verified that the present invention can effectively complete the accurate classification of cultivated land, gardens and woodlands in cultivated land protection projects, and has significantly improved the classification efficiency and accuracy compared with traditional methods.

[0171] In summary, the present invention solves the problems of low efficiency, large errors and high labor costs in traditional methods of identifying interplanted trees on cultivated land. In practical applications, it not only improves the classification accuracy, but also significantly shortens the investigation time and labor costs, providing an efficient and reliable technical means for large-scale cultivated land protection tasks.

[0172] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for identifying interplanted trees on cultivated land combining mixed pixel decomposition and deep learning, characterized in that: The steps include: S1. Obtain remote sensing image data and historical monitoring results, pre-process remote sensing images, and generate basic classification image data; S2. Based on the classification standards of basic classified image data and historical monitoring results, establish sample labels for interplanted trees on cultivated land and generate standardized sample sets; S3. Based on the mixed pixel decomposition technology, the nonlinear mixed pixel decomposition model is used to invert the end-member abundance of ground objects, and the number of seedlings planted and the forest density in the area are calculated in combination with the abundance of ground objects; S4. Build a deep neural network remote sensing image interpretation model; S5. Use the deep neural network remote sensing image interpretation model to automatically extract the spectral, texture and spatial features of remote sensing images and generate classification results for cultivated land, gardens and forests; S6. Cross-validate the classification results of the deep neural network remote sensing image interpretation model with the mixed pixel decomposition classification results to optimize the standardized sample set and the deep neural network remote sensing image interpretation model; The S2 specifically includes: S21. Extract classification standards from historical monitoring results, and define classification rules for cultivated land samples, garden samples and forest samples based on fruit tree coverage, number of fruit trees, tree canopy density and planting density parameters, including: Cultivated land samples: fruit tree coverage ≤ 50% or the number of fruit trees per mu ≤ 70% of the reasonable number of trees, tree density ≤ 0.2; Garden samples: fruit tree coverage > 50% or the number of fruit trees per mu > 70% of the reasonable number; Forest samples: tree density> 0.2; S22, extracting preliminary classification information based on the spectral characteristics, geometric characteristics and texture characteristics of each ground object target in the basic classification image data and combining the classification standard, and judging whether the target meets the classification conditions of cultivated land samples, garden samples and forest samples; S23. For targets whose spectral characteristics cannot be clearly classified, the fruit tree coverage, number of fruit trees per mu, tree density and planting density parameters of each ground object are calculated based on the pixel-level analysis results of remote sensing images: ； in, is the fruit tree coverage, The area occupied by fruit trees, is the total target area; S24, according to the reasonable number of trees in the classification conditions, combined with the land area and planting density parameters of the land object, calculate the number of fruit trees per mu, and confirm whether the land object meets the classification conditions of cultivated land samples and garden samples; S25. For the ground objects with tree vegetation, the tree canopy density information is inverted through the tree end members in the image, the tree canopy density is calculated, and the classification conditions are used to determine whether the target meets the classification conditions of cultivated land samples and forest samples; S26. Assign corresponding labels to the ground objects that meet the classification conditions, classify the sample objects into cultivated land samples, garden samples and forest samples, generate a standardized sample data set containing classification labels, and associate the classification information with the image feature data for storage.

2. The method for identifying interplanted trees on cultivated land by combining mixed pixel decomposition and deep learning according to claim 1, characterized in that: The S3 specifically includes: S31. Based on the mixed pixel decomposition technology, a nonlinear mixed pixel decomposition model is constructed to describe the relationship between the spectral reflectance of the observed pixel and the endmember abundance: ； in, represents the observed mixed spectrum, Indicates The abundance ratio of end members, For the Spectral reflectance of the end member, is the secondary scattering contribution ratio between the i-th and j-th end members, is the cross-endmember reflectivity product, is the error term; S32. Based on remote sensing image data, the pure pixel index method is used to extract endmember spectral data, including illuminated canopy endmembers, illuminated surface endmembers and shadow endmembers, so that the extracted endmember spectra are representative and stable for constructing nonlinear mixed models. S33, using the least squares method to invert the mixed spectrum and calculate the end member abundance, applying full constraints and non-negativity conditions in the inversion process, the constraints are: ； in, Indicates The abundance ratio of end members, is the secondary scattering contribution ratio of the cross-end members, so that the abundance ratio of all end members is non-negative and the sum is 1; S34. Based on the end-member abundance values ​​obtained by inversion, the number of seedlings to be planted is calculated in combination with the target area and the area occupied by a single tree: ； in, is the total number of seedlings, is the total area of ​​the target region, The area occupied by a single plant, For the The abundance ratio of end members, For the Contribution coefficient of the class end member; S35. By using the canopy closure inversion algorithm of the geometric optical model, the end-member abundance of the illumination surface is combined with the solar zenith angle, the observation zenith angle, the solar azimuth angle and the geometric information of the observation azimuth angle, and the average tree height and crown radius of the dominant tree species are input to calculate the canopy closure distribution in the area; S36. Based on the calculated number of saplings planted and canopy density distribution, the categories of the land objects in the standardized sample set are determined, and the objects are divided into cultivated land, garden land and forest land according to the classification rules.

3. The method for identifying interplanted trees on cultivated land by combining mixed pixel decomposition and deep learning according to claim 2, characterized in that: The S31 specifically includes: S311. Based on the mixed pixel decomposition technology, a nonlinear mixed pixel decomposition model is constructed to describe the relationship between the spectral reflectance of the observed pixel and the end member abundance. The model includes a linear reflection component and a nonlinear scattering component. S312. Define the linear reflection component, which is used to characterize the direct contribution of each end member to the observed spectrum. The linear reflection component is represented by the product of the end member abundance ratio and the end member spectral reflectance: ； in, represents the linear reflection component, For the The abundance ratio of end members, For the Spectral reflectance of the endmember-like structure; S313. Define the nonlinear scattering component, which is used to describe the effect of the interaction between end members on the spectrum. The nonlinear scattering component is represented by the cross product of the cross-end member contribution ratio and the end member spectral reflectance: ； in, represents the effect of nonlinear interactions between end members on the mixed spectrum, is the cross-contribution ratio between the jth and kth end members, is the product of the spectral reflectance of the two end elements; S314. Combining the linear reflection component and the nonlinear scattering component, a nonlinear mixed pixel decomposition model is constructed, and the observed spectrum is described by the following expression: ； in, To observe the mixed spectrum, is the spectral mixing function, represents the abundance ratio of the 1st to the nth end members. The abundance of each end member reflects its relative contribution in the observed spectrum. It represents the spectral reflectance of each end member and describes the optical properties of the end member.

4. The method for identifying interplanted trees in cultivated land by combining mixed pixel decomposition and deep learning according to claim 2, characterized in that: The S33 specifically includes: S331, inverting the mixed spectrum using the least squares method, and calculating the abundance ratio of each end member in the mixed pixel and the secondary scattering contribution ratio of the cross end member by minimizing the error between the observed spectrum and the model predicted spectrum; S332. Full constraints are imposed during the inversion process so that the sum of the linear endmember abundance and the cross-endmember contribution ratio is equal to 1. The constraints are defined as: ； S333. At the same time, the non-negative condition is applied to restrict the values ​​of all end-member abundance ratios and cross-end-member contribution ratios to be non-negative. The non-negative condition is defined as: ； S334. On the basis of satisfying the above-mentioned full constraint conditions and non-negativity conditions, the endmember abundance ratio and the cross-endmember contribution ratio are updated through an iterative optimization algorithm until the error meets the convergence condition, and finally the endmember abundance ratio and the cross-endmember contribution ratio are determined.

5. The method for identifying interplanted trees in cultivated land by combining mixed pixel decomposition and deep learning according to claim 1, characterized in that: S4 specifically includes: S41. Design a deep learning network architecture, select a convolutional neural network based on the scale change and directional characteristics of the intercropping of trees and ground objects on cultivated land, and develop an adaptive network structure based on remote sensing image classification tasks; S42, multi-scale feature extraction is performed by setting multi-layer convolution operations and convolution kernels of different sizes to complete the feature extraction process of the ground object target; S43. Configure a direction-sensitive model in the network architecture, combine direction encoding technology and rotation-invariant convolution operations, extract the direction characteristics of ground objects, and effectively capture direction information; S44. Use the standardized sample set generated by mixed pixel decomposition to train the network model, adopt a fully supervised learning method, and calculate the cross entropy loss function through the sample labels: ； in, is the loss value, is the total number of samples, For the The true labels of samples, The classification probability predicted by the network model; S45. Optimize the network parameters through the back propagation algorithm, use the gradient descent method to minimize the loss function, complete multiple iterative training of the model, input the mixed pixel decomposition results to adjust the categories of the standardized sample set, and optimize the accuracy of category recognition.

6. The method for identifying interplanted trees in cultivated land by combining mixed pixel decomposition and deep learning according to claim 1, characterized in that: S5 specifically includes: S51. Perform multi-channel processing on remote sensing images through a deep neural network remote sensing image interpretation model to extract a basic feature set including spectral features, texture features and spatial features; S52. Use the convolutional layer operation of the deep neural network to extract spectral features, analyze the spectral response values ​​of each band of the remote sensing image, and generate a spectral feature map reflecting the spectral differences of cultivated land, gardens and forests. ; S53, combine the gray level co-occurrence matrix method to calculate the texture features of the image, extract the surface roughness, directional consistency and texture uniformity of the object through window sliding operation, and generate a texture feature map ; S54, extract spatial features based on the spatial attention mechanism of deep learning network, and generate spatial feature maps by capturing the spatial distribution pattern of ground objects and the correlation between adjacent pixels. ; S55, will , and Perform multi-layer fusion operations to construct comprehensive feature representation , and input the fused features into the classification layer to generate the classification results of cultivated land, garden land and forest land; S56. Evaluate the classification performance based on the validation set and calculate the precision, recall and F1 value of the classification model: ； ； ； in, is the accuracy, is the recall rate, is the F1 value, represents the number of true positive examples, represents the number of false positive examples, represents the number of false negative examples; S57. Adjust model parameters based on the performance evaluation results of the validation set to optimize the comprehensive feature representation Finally, the test set is used to verify the robustness and overall accuracy of the classification model.

Images