Abandoned land identification method and system based on true value approximation principle for bidirectional network

Abstract

The present application belongs to the field of abandoned land monitoring, and particularly relates to a two-way network abandoned land identification method and system based on the true value approximation principle. The method first pre-processes the land image sequence to be identified and automatically divides the blocks using a vector marking model. Second, a two-way identification network model and a crop type identification library are configured, and a reverse identification network sub-model is used to identify potential abandoned land blocks that do not contain crops and their proportion. Third, the proportion of abandoned land blocks is confirmed through a forward identification network sub-model. At the same time, combined with historical land rotation information and forest and grass wetland block information data, the proportion of non-abandoned land blocks in different seasons is obtained. Finally, the proportion of potential abandoned land blocks, the proportion of abandoned land blocks, and the proportion of non-abandoned land blocks at the corresponding time point are input into the logical discrimination layer, and the discrimination error is reduced through training until it is zero, thereby outputting accurate abandoned land proportion and corresponding division marking blocks, and realizing efficient and accurate abandoned land identification.

Description

Technical Field

[0001] The present invention belongs to the field of abandoned land monitoring, and in particular relates to an abandoned land identification method and system based on a two-way network of a truth value approximation principle. Background Art

[0002] Most of China's arable land consists of small plots belonging to different individuals. This decentralized land ownership structure poses many challenges to traditional classification methods when identifying abandoned land. The most important of these is the unclear boundaries between abandoned land and potential abandoned land, making it difficult for traditional methods to accurately distinguish them. Especially in complex terrain conditions, the classification accuracy is low and even lower. This not only increases the difficulty of identifying abandoned land, but also affects its effective management and utilization.

[0003] For example, the Chinese patent with authorization announcement number CN103914678B discloses a remote sensing method for identifying abandoned land based on texture and vegetation index. This method utilizes the characteristics of multi-sensor, multi-resolution, multi-spectral, and multi-temporal remote sensing data, integrates the spectral, texture characteristics and vegetation index characteristics of remote sensing images, and establishes a method for identifying abandoned land based on texture and vegetation index by constructing a remote sensing image registration and fusion framework, spectral and texture feature space, and vegetation index time series.

[0004] For example, a Chinese patent with authorization announcement number CN115082803B discloses a method, device and storage medium for monitoring abandoned farmland based on seasonal changes in vegetation. The method includes: for the area to be detected, by collecting the normalized vegetation index NDVI time series curve, obtaining an abandoned land sample feature set and a non-abandoned land sample feature set; based on the abandoned land sample feature set and the non-abandoned land sample feature set, determining the abandoned land criterion for the area to be detected; and according to the abandoned land criterion, identifying the abandoned land in the area to be detected.

[0005] The above existing technologies have the following problems: when facing the identification of abandoned land, the traditional classification method has low classification accuracy and even lower accuracy for complex terrain due to the blurred boundaries between categories such as abandoned land and potential abandoned land. For this reason, the present invention provides an abandoned land identification method and system based on a bidirectional network of the truth value approximation principle. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention proposes an abandoned land identification method and system based on a bidirectional network of the truth value approximation principle. The method first pre-processes the land image sequence to be identified and uses a pre-trained vector labeling model to automatically divide the blocks; secondly, a bidirectional recognition network model and a crop type recognition library are configured, and the reverse recognition network sub-model is used to identify potential abandoned area blocks that do not contain crops and their proportions; thirdly, the proportion of abandoned area blocks is further confirmed by the forward recognition network sub-model; at the same time, combined with historical land rotation information and forest, grassland and wetland block information data, a seasonal non-abandoned land proportion prediction model is trained to obtain the proportion of non-abandoned area blocks in different seasons; finally, the potential abandoned area block proportion, the abandoned area block proportion and the non-abandoned area block proportion information at the corresponding time point are input into the logical discrimination layer of the bidirectional recognition network model, and the discrimination error is reduced to zero through training, thereby outputting the accurate abandoned area block proportion and the corresponding divided marked blocks, thereby achieving efficient and accurate abandoned land identification.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for identifying abandoned land using a bidirectional network based on the principle of true value approximation includes the following steps:

[0009] S1. Obtain a sequence of land images to be identified, input the sequence of land images to be identified into a vector marking model for automatic land block marking, and obtain the sequence of land images to be identified after block division, the number of divided blocks, and vector data of the corresponding blocks;

[0010] S2. Configure a crop type recognition library and input the crop type recognition library, the image sequence of land to be identified after block division, the vector data, and the number of divided blocks into the reverse recognition network sub-model of the configured bidirectional recognition network model to obtain a sequence of potential abandoned land to be identified that does not contain crops and the corresponding proportion of potential abandoned land blocks;

[0011] S3. Input the sequence of blocks of potential abandoned land that do not contain crops and the vector data into the forward recognition network sub-model of the bidirectional recognition network model to obtain the proportion of abandoned land blocks and the corresponding marked blocks;

[0012] S4. Obtaining historical land rotation information and forest, grassland and wetland block information data corresponding to different seasons in the area to be identified, and inputting the obtained data into the configured seasonal non-abandoned land proportion prediction model for training to obtain non-abandoned land block proportion information corresponding to different seasonal points;

[0013] S5. Input the information of the proportion of potentially abandoned blocks, the proportion of abandoned blocks, and the proportion of non-abandoned blocks at the corresponding time point into the logical discrimination layer of the bidirectional recognition network model to obtain a true value approximation error value;

[0014] S6. Feedback the true value approximation error value to the bidirectional recognition network model for training until the true value approximation error value is less than or equal to a preset value, and output the abandoned area block ratio and the corresponding partitioned marked blocks obtained by the forward recognition network sub-model.

[0015] Specifically, the steps of automatically marking land blocks include:

[0016] S101, preprocessing the acquired sequence of land images to be identified through filtering and brightness enhancement algorithms to obtain a preprocessed sequence of land images to be identified;

[0017] S102. Vector marking is performed on the corresponding blocks using the block area, block boundary and shape, and plot slope data in the vector data. The brightness and pixel color of the ridges inside the blocks and at the block boundaries are secondary marked using the YUV color space. The marked blocks are numbered to obtain a marked sequence of land images to be identified.

[0018] S103, constructing a farmland block division model, inputting the marked land image sequence to be identified into the farmland block division model to extract image content information and extract brightness and chromaticity difference features of the field ridges inside the block and at the block boundary, and constructing a comprehensive training loss function using the extracted brightness and chromaticity difference features and image content information extraction loss;

[0019] S104, setting a comprehensive loss threshold and a training cycle, embedding the constructed comprehensive training loss function into the farmland block division model for training, and obtaining a trained farmland block division model when the comprehensive training loss function meets the comprehensive loss threshold or meets the training cycle;

[0020] S105: The trained farmland block division model is configured in an image acquisition device, and the captured image is automatically divided into blocks and vector data marked to obtain a sequence of land images to be identified after the block division is completed and the number of divided blocks.

[0021] Specifically, the steps for constructing the reverse recognition network sub-model include:

[0022] S201. Input the block-divided sequence of land images to be identified and the number of divided blocks into the segmentation sublayer of the reverse recognition network submodel, perform block segmentation according to the block boundaries of each marked image, and obtain a block sequence set vij,i=1…I,j=1…J, where vij represents the jth block in the i-th land image to be identified, I represents the length of the land image sequence to be identified, and J represents the number of blocks in each land image to be identified;

[0023] S202: Mark each block with crops with the corresponding crop type, and mark the blocks without crops with 0, to obtain a sequence set of blocks with completed crop marking;

[0024] S203: Input the crop-labeled block sequence set into the polarization feature layer in the reverse recognition network sub-model to obtain the polarization texture feature space corresponding to each block;

[0025] S204: Input the polarization texture feature space into the extended morphology multi-attribute profile layer to obtain the multi-dimensional polarization profile texture features corresponding to the polarization texture feature space of each block.

[0026] Specifically, the steps of constructing the reverse recognition network sub-model also include:

[0027] S205. Input the multidimensional polarization profile texture features into the optimal classification feature extraction layer to obtain the optimal polarization profile texture features corresponding to the polarization texture feature space of each block. At the same time, input the multidimensional polarization profile texture features into the Gabor filter layer to obtain the polarization texture space channel features corresponding to the polarization texture feature space of each block.

[0028] S206, inputting the optimal polarization profile texture features and polarization texture spatial channel features corresponding to each block polarization texture feature space into the convolutional attention fusion layer to obtain the polarization texture fusion spatial features corresponding to each block;

[0029] S207, inputting the polarization texture fusion spatial features and the crop category features contained in the crop type recognition library simultaneously into the discriminant sublayer and the classification sublayer in the discriminant output layer, and obtaining the proportion of blocks containing crops and the probability of correct crop classification respectively;

[0030] S208. Obtain the percentage of blocks that actually contain crops based on the labels marked in S202, and use the percentage of blocks that actually contain crops as a training threshold for the reverse recognition network sub-model, while also setting a probability threshold for correct crop classification.

[0031] S209. The reverse recognition network sub-model is trained according to the two thresholds set in S208. When the proportion of blocks containing crops is equal to the proportion of blocks actually containing crops and the probability of correct crop classification is greater than the probability threshold of correct crop classification, a trained reverse recognition network sub-model is obtained.

[0032] Specifically, the steps for constructing the forward recognition network sub-model include:

[0033] S301. Mark the potential abandoned land blocks that do not contain crops as abandoned land, uncultivated land, and forest, grassland and wetland;

[0034] S302: Inputting the polarization texture fusion spatial features, vector data, and vegetation index features of the marked potential abandoned land blocks that do not contain crops into the MobileNet layer of the forward recognition network sub-model to extract non-crop vegetation features and obtain non-crop vegetation features;

[0035] S303: Input the non-agricultural crop vegetation characteristics, the corresponding block rotation information, and the number characteristics of the divided blocks into the abandonment discrimination layer in the forward recognition network sub-model to obtain the proportion of abandoned area blocks and the corresponding marked blocks.

[0036] Specifically, the steps for training the bidirectional recognition network model include:

[0037] S501, subtracting the proportion of abandoned land blocks from the proportion of potentially abandoned land blocks to obtain the error in the proportion of non-abandoned land identification;

[0038] S502: Subtract the percentage of non-abandoned land blocks at the abandoned land identification time point obtained in step S4 from the non-abandoned land identification percentage error to obtain a true value approximation error value;

[0039] S503. Set a true value approximation error threshold, feed the true value approximation error value back to the bidirectional recognition network model for training, and when the true value approximation error value is less than the true value approximation error threshold, obtain a trained bidirectional recognition network model.

[0040] The abandoned land recognition system based on a bidirectional network of the truth value approximation principle includes: an image marking module, a reverse recognition module, a forward recognition module, and a logical judgment output module;

[0041] The image marking module includes an image acquisition unit and a block automatic marking unit; the image acquisition unit is used to acquire a sequence of land images to be identified;

[0042] The block automatic labeling unit is used to automatically label land blocks based on the sequence of land images to be identified through a vector labeling model, and obtain the sequence of land images to be identified after block division, the number of divided blocks and the vector data corresponding to the divided blocks;

[0043] The reverse recognition module is used to obtain the sequence of potential abandoned land to be identified that does not contain crops and the corresponding proportion of potential abandoned land blocks based on the configured crop type recognition library, the sequence of land images to be identified after block division, the vector data and the number of divided blocks, through the reverse recognition network sub-model in the bidirectional recognition network model.

[0044] Specifically, the forward recognition module is used to input the sequence of blocks of potential abandoned land that do not contain crops and vector data into the forward recognition network sub-model of the bidirectional recognition network model to obtain the proportion of abandoned land blocks and the corresponding marked blocks;

[0045] The logic discrimination output module includes an experience information unit, a logic discrimination unit and a truth value approximation training unit;

[0046] The empirical information unit is used to obtain land rotation information and forest, grassland and wetland block information data corresponding to different seasons in the historical identified area, and input the obtained data into the configured seasonal non-abandoned land proportion prediction model for training to obtain the non-abandoned area block proportion information corresponding to different seasonal points;

[0047] A logical discrimination unit is used to input the proportion of potentially abandoned blocks, the proportion of abandoned blocks, and the proportion of non-abandoned blocks at the corresponding time point into the logical discrimination layer of the bidirectional recognition network model to obtain a true value approximation error value;

[0048] The true value approximation training unit is used to feed back the true value approximation error value to the bidirectional recognition network model for training until the true value approximation error value is less than or equal to the preset value, and output the proportion of abandoned area blocks and the corresponding divided marked blocks obtained by the forward recognition network sub-model.

[0049] A computer-readable storage medium stores computer instructions, which, when executed, execute an abandoned land identification method based on a bidirectional network and the principle of true value approximation.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] In response to the shortcomings of the existing technology, the present invention significantly improves the accuracy and reliability of abandoned land identification through multi-step image processing and intelligent model training. First, land blocks are automatically divided through preprocessing and vector labeling models, which improves the accuracy and consistency of block division. The configured bidirectional recognition network model, combined with the crop type recognition library and vector data, can more accurately distinguish between non-abandoned land containing crops and potential abandoned land not containing crops. The introduction of vector data also improves the recognition accuracy under complex terrain. In addition, the seasonal non-abandoned land proportion prediction model trained with historical data further corrects the non-abandoned land proportion information in different seasons, enhancing the temporal consistency and accuracy of the recognition results. The introduction of the logical discrimination layer ensures the high accuracy of the final output result, and the discrimination error is approached to zero through iterative training. In summary, the method effectively improves the accuracy and reliability of abandoned land identification and is suitable for large-scale farmland monitoring and management. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of a method for identifying abandoned land using a bidirectional network based on the principle of true value approximation according to embodiment 1 of the present invention;

[0053] Figure 2A physical map of the land image to be identified and divided into blocks according to the first embodiment of the present invention;

[0054] Figure 3 This is a diagram of the reverse identification network sub-model architecture of Example 1 of the present invention;

[0055] Figure 4 This is a module diagram of the abandoned land identification system based on a bidirectional network of the truth value approximation principle in embodiment 2 of the present invention. DETAILED DESCRIPTION

[0056] Example 1

[0057] See also Figure 1 The present invention provides an embodiment of a method for identifying abandoned land using a bidirectional network based on the principle of true value approximation, comprising the following steps:

[0058] S1. Obtain a sequence of land images to be identified, input the sequence of land images to be identified into a vector marking model for automatic land block marking, and obtain the sequence of land images to be identified after block division, the number of divided blocks, and vector data of the corresponding blocks;

[0059] Furthermore, in this embodiment, the step of automatically marking land blocks includes:

[0060] S101, preprocessing the acquired sequence of land images to be identified through filtering and brightness enhancement algorithms to obtain a preprocessed sequence of land images to be identified;

[0061] S102. Vector marking is performed on the corresponding blocks using the block area, block boundary and shape, and plot slope data in the vector data. The brightness and pixel color of the ridges inside the blocks and at the block boundaries are secondary marked using the YUV color space. The marked blocks are numbered to obtain a marked sequence of land images to be identified.

[0062] S103, constructing a farmland block division model, inputting the marked land image sequence to be identified into the farmland block division model to extract image content information and extract brightness and chromaticity difference features of the field ridges inside the block and at the block boundary, and constructing a comprehensive training loss function using the extracted brightness and chromaticity difference features and image content information extraction loss;

[0063] S104, setting a comprehensive loss threshold and a training cycle, embedding the constructed comprehensive training loss function into the farmland block division model for training, and obtaining a trained farmland block division model when the comprehensive training loss function meets the comprehensive loss threshold or meets the training cycle;

[0064] S105: The trained farmland block division model is configured in an image acquisition device, and the captured image is automatically divided into blocks and vector data marked to obtain a sequence of land images to be identified after the block division is completed and the number of divided blocks.

[0065] Further, see Figure 2 , ① marked in the figure is the block boundary marked by the block division, Figure 2 The image is from the official land database after the block division, which is used for the training of the farmland block division model. In this embodiment, the land image to be identified after the real-time block division and Figure 2 The style is the same, and the official land database is obtained through the permanent basic farmland query platform.

[0066] The process first preprocesses the image through filtering and brightness enhancement algorithms to ensure the basic quality of subsequent processing. Secondly, vector data is used to perform detailed vector marking on the blocks, and the interior of the blocks and the boundary ridges are secondary marked in combination with the YUV color space, which not only enhances the feature discrimination but also improves the accuracy of the marking. The constructed farmland block division model extracts image content information and brightness and chromaticity difference features to form a comprehensive training loss function to ensure that the model learns the most effective feature representation. The comprehensive loss threshold and training cycle are set so that the model can stop training when the optimal performance is achieved to avoid overfitting or underfitting. Finally, the trained model is configured in the image acquisition device to realize real-time, automatic block division and vector data marking. This process not only greatly reduces manual intervention and improves work efficiency, but also ensures the accuracy and consistency of block division, providing a solid foundation for subsequent abandoned land identification and other agricultural management applications.

[0067] S2. Configure a crop type recognition library and input the crop type recognition library, the image sequence of land to be identified after block division, the vector data, and the number of divided blocks into the reverse recognition network sub-model of the configured bidirectional recognition network model to obtain a sequence of potential abandoned land to be identified that does not contain crops and the corresponding proportion of potential abandoned land blocks;

[0068] Further, see Figure 3 In this embodiment, the steps of constructing the reverse recognition network sub-model include:

[0069] S201. Input the block-divided sequence of land images to be identified and the number of divided blocks into the segmentation sublayer of the reverse recognition network submodel, perform block segmentation according to the block boundaries of each marked image, and obtain a block sequence set vij,i=1…I,j=1…J, where vij represents the jth block in the i-th land image to be identified, I represents the length of the land image sequence to be identified, and J represents the number of blocks in each land image to be identified;

[0070] S202: Mark each block with crops with the corresponding crop type, and mark the blocks without crops with 0, to obtain a sequence set of blocks with completed crop marking;

[0071] S203: Input the crop-labeled block sequence set into the polarization feature layer in the reverse recognition network sub-model to obtain the polarization texture feature space corresponding to each block;

[0072] Furthermore, in this embodiment, the polarization feature layer is configured with an existing polarization attention network to extract the polarization texture feature space corresponding to each block.

[0073] S204, inputting the polarization texture feature space into the extended morphology multi-attribute profile layer to obtain the multi-dimensional polarization profile texture features corresponding to the polarization texture feature space of each block;

[0074] Furthermore, the extended morphological multi-attribute profile layer in this embodiment is configured with an extended morphological multi-attribute profile algorithm. The extended morphological multi-attribute profile algorithm can solve the problem that a single attribute profile is difficult to accurately describe the texture characteristics of high-resolution PolSAR images by integrating different extended attribute profile filters.

[0075] S205. Input the multidimensional polarization profile texture features into the optimal classification feature extraction layer to obtain the optimal polarization profile texture features corresponding to the polarization texture feature space of each block. At the same time, input the multidimensional polarization profile texture features into the Gabor filter layer to obtain the polarization texture space channel features corresponding to the polarization texture feature space of each block.

[0076] Furthermore, in this embodiment, the optimal classification feature extraction layer is configured with a random forest algorithm to extract the optimal (most conducive to classification) polarization feature from the multi-dimensional polarization profile texture features through the morphological attribute profile features.

[0077] S206, inputting the optimal polarization profile texture features and polarization texture spatial channel features corresponding to each block polarization texture feature space into the convolutional attention fusion layer to obtain the polarization texture fusion spatial features corresponding to each block;

[0078] Furthermore, in this embodiment, the convolutional attention fusion layer is configured with a convolutional attention network.

[0079] S207, inputting the polarization texture fusion spatial features and the crop category features contained in the crop type recognition library simultaneously into the discriminant sublayer and the classification sublayer in the discriminant output layer, and obtaining the proportion of blocks containing crops and the probability of correct crop classification respectively;

[0080] Furthermore, in this embodiment, the discriminant sublayer is configured with a softmax activation function, and the classification sublayer is configured with a fully connected function.

[0081] S208. Obtain the percentage of blocks that actually contain crops based on the labels marked in S202, and use the percentage of blocks that actually contain crops as a training threshold for the reverse recognition network sub-model, while also setting a probability threshold for correct crop classification.

[0082] S209. The reverse recognition network sub-model is trained according to the two thresholds set in S208. When the proportion of blocks containing crops is equal to the proportion of blocks actually containing crops and the probability of correct crop classification is greater than the probability threshold of correct crop classification, a trained reverse recognition network sub-model is obtained.

[0083] S210, inputting the sequence of land images to be identified after the block division obtained in real time in S105 and the number of divided blocks into the trained reverse recognition network sub-model to obtain the proportion of blocks containing crops and the sequence of potential abandoned land to be identified that does not contain crops;

[0084] S211. Obtain the corresponding proportion of blocks in the potential abandoned land that does not contain crops through the sequence of blocks to be identified and the number of divided blocks.

[0085] In this process, first, the land image sequence to be identified after block division is accurately divided into a block sequence set, and the crop type is marked for each block to ensure the accuracy and consistency of data annotation; the polarization feature layer of the polarization attention network is introduced to effectively extract the polarization texture feature space of each block, enhancing the model's ability to understand complex environments; the extended morphological multi-attribute profile layer solves the problem of high-resolution image texture features that are difficult to describe with a single attribute profile by integrating different attribute profile filters, further enriching the feature representation; the optimal classification feature extraction layer uses the random forest algorithm to select the polarization features that are most conducive to classification from the multi-dimensional polarization profile texture features, thereby improving the classification accuracy. Accuracy; the convolutional attention fusion layer combines the optimal polarization profile texture features and polarization texture spatial channel features to generate richer polarization texture fusion spatial features, enhancing the model's discrimination of different block types; the discriminant output layer configures the softmax activation function and the fully connected function to calculate the proportion of blocks containing crops and the probability of correct classification, respectively, ensuring the reliability and interpretability of the output results; by setting the training threshold and the classification correct probability threshold, the model can stop training when the optimal performance is reached, avoiding overfitting or underfitting, and ensuring the stability and generalization ability of the model; this model not only reduces manual intervention and improves work efficiency, but also ensures the accuracy and consistency of the recognition results.

[0086] S3. Input the sequence of blocks of potential abandoned land that do not contain crops and the vector data into the forward recognition network sub-model of the bidirectional recognition network model to obtain the proportion of abandoned land blocks and the corresponding marked blocks;

[0087] Furthermore, the steps of constructing the forward recognition network sub-model in this embodiment include:

[0088] S301. Mark the potential abandoned land blocks that do not contain crops as abandoned land, uncultivated land, and forest, grassland and wetland;

[0089] S302: Inputting the polarization texture fusion spatial features, vector data, and vegetation index features of the marked potential abandoned land blocks that do not contain crops into the MobileNet layer of the forward recognition network sub-model to extract non-crop vegetation features and obtain non-crop vegetation features;

[0090] S303: Input the non-agricultural crop vegetation characteristics, the corresponding block rotation information, and the number characteristics of the divided blocks into the abandonment discrimination layer in the forward recognition network sub-model to obtain the proportion of abandoned area blocks and the corresponding marked blocks.

[0091] By constructing a forward recognition network sub-model, this process can efficiently and accurately identify the proportion of abandoned area blocks and corresponding marked blocks in the potential abandoned land blocks that do not contain crops; it uses polarization texture to fuse spatial features, vector data, vegetation index features and other multi-source information, combined with block rotation information and block quantity characteristics, to improve the accuracy and reliability of abandoned land identification.

[0092] S4. Obtaining historical land rotation information and forest, grassland and wetland block information data corresponding to different seasons in the area to be identified, and inputting the obtained data into the configured seasonal non-abandoned land proportion prediction model for training to obtain non-abandoned land block proportion information corresponding to different seasonal points;

[0093] S5. Input the information of the proportion of potentially abandoned blocks, the proportion of abandoned blocks, and the proportion of non-abandoned blocks at the corresponding time point into the logical discrimination layer of the bidirectional recognition network model to obtain a true value approximation error value;

[0094] S6. Feedback the true value approximation error value to the bidirectional recognition network model for training until the true value approximation error value is less than or equal to a preset value, and output the abandoned area block ratio and the corresponding partitioned marked blocks obtained by the forward recognition network sub-model.

[0095] Furthermore, the steps of training the bidirectional recognition network model in this embodiment include:

[0096] S501, subtracting the proportion of abandoned land blocks from the proportion of potentially abandoned land blocks to obtain the error in the proportion of non-abandoned land identification;

[0097] S502: Subtract the percentage of non-abandoned land blocks at the abandoned land identification time point obtained in step S4 from the non-abandoned land identification percentage error to obtain a true value approximation error value;

[0098] S503. Set a true value approximation error threshold, feed the true value approximation error value back to the bidirectional recognition network model for training, and when the true value approximation error value is less than the true value approximation error threshold, obtain a trained bidirectional recognition network model.

[0099] This process significantly improves the accuracy and reliability of abandoned land identification by introducing historical land rotation information and forest, grassland and wetland block information data, and combining it with a seasonal non-abandoned land proportion prediction model. First, by obtaining land rotation information and forest, grassland and wetland block information data corresponding to different seasons and inputting them into the seasonal non-abandoned land proportion prediction model for training, the proportion of non-abandoned area blocks corresponding to each seasonal point is obtained, providing an important time series reference for subsequent discrimination. Second, in the logical discrimination layer, the proportion of abandoned area blocks is subtracted from the proportion of abandoned area blocks by the proportion of potential abandoned area blocks to calculate the non-abandoned land identification proportion error. Combined with the non-abandoned area block proportion at the time point obtained in the S4 process, the true value approximation error is further calculated. This process ensures that the model can dynamically adjust its recognition ability of abandoned and non-abandoned land, adapt to changes in different seasons, and enhance its robustness and generalization ability in complex environments.

[0100] Example 2

[0101] See also Figure 4 Another embodiment provided by the present invention is a wasteland identification system based on a bidirectional network of a truth value approximation principle, comprising: an image marking module, a reverse identification module, a forward identification module, and a logic judgment output module;

[0102] An image marking module is used to acquire a sequence of land images to be identified and to mark the blocks to be identified; the image marking module includes an image acquisition unit and a block automatic marking unit;

[0103] An image acquisition unit is used to acquire a sequence of land images to be identified; a block automatic labeling unit is used to automatically label land blocks based on the sequence of land images to be identified using a vector labeling model, and obtain the sequence of land images to be identified after block division, the number of divided blocks, and the vector data corresponding to the divided blocks;

[0104] The reverse recognition module is used to obtain a sequence of potential abandoned land blocks that do not contain crops and the corresponding proportion of potential abandoned land blocks based on the configured crop type recognition library, the sequence of land images to be identified after block division, the vector data, and the number of divided blocks.

[0105] The forward recognition module is used to input the sequence of blocks of potential abandoned land that do not contain crops and vector data into the forward recognition network sub-model of the bidirectional recognition network model to obtain the proportion of abandoned land blocks and the corresponding marked blocks;

[0106] A logic discrimination output module is used for true value logic discrimination approximation of abandoned area blocks; the logic discrimination output module includes an experience information unit, a logic discrimination unit, and a true value approximation training unit;

[0107] The empirical information unit is used to obtain land rotation information and forest, grassland and wetland block information data corresponding to different seasons in the historical identified area, and input the obtained data into the configured seasonal non-abandoned land proportion prediction model for training to obtain the non-abandoned area block proportion information corresponding to different seasonal points;

[0108] A logical discrimination unit is used to input the proportion of potentially abandoned blocks, the proportion of abandoned blocks, and the proportion of non-abandoned blocks at the corresponding time point into the logical discrimination layer of the bidirectional recognition network model to obtain a true value approximation error value;

[0109] The true value approximation training unit is used to feed back the true value approximation error value to the bidirectional recognition network model for training until the true value approximation error value is less than or equal to the preset value, and output the proportion of abandoned area blocks and the corresponding divided marked blocks obtained by the forward recognition network sub-model.

[0110] Example 3

[0111] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, a method for identifying abandoned land using a bidirectional network based on the principle of true value approximation is implemented.

[0112] A computer-readable storage medium stores computer instructions, which, when executed, execute an abandoned land identification method based on a bidirectional network and the principle of true value approximation.

[0113] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also change, modify, replace and modify the above-mentioned embodiments without departing from the scope of protection of the purpose of the present invention and the claims, and all of these are protected by the present invention.

Claims

1. A method for identifying abandoned land using a bidirectional network based on the principle of true value approximation, characterized by: include: S1. Obtain a sequence of land images to be identified, input the sequence of land images to be identified into a vector marking model for automatic land block marking, and obtain the sequence of land images to be identified after block division, the number of divided blocks, and vector data of the corresponding blocks; S2. Configure a crop type recognition library and input the crop type recognition library, the image sequence of land to be identified after block division, the vector data, and the number of divided blocks into the reverse recognition network sub-model of the configured bidirectional recognition network model to obtain a sequence of potential abandoned land to be identified that does not contain crops and the corresponding proportion of potential abandoned land blocks; S3. Input the sequence of blocks of potential abandoned land that do not contain crops and the vector data into the forward recognition network sub-model of the bidirectional recognition network model to obtain the proportion of abandoned land blocks and the corresponding marked blocks; S4. Obtaining historical land rotation information and forest, grassland and wetland block information data corresponding to different seasons in the area to be identified, and inputting the obtained data into the configured seasonal non-abandoned land proportion prediction model for training to obtain non-abandoned land block proportion information corresponding to different seasonal points; S5. Input the information of the proportion of potentially abandoned blocks, the proportion of abandoned blocks, and the proportion of non-abandoned blocks at the corresponding time point into the logical discrimination layer of the bidirectional recognition network model to obtain a true value approximation error value; S6. Feeding back the true value approximation error value to the bidirectional recognition network model for training until the true value approximation error value is less than or equal to a preset value, outputting the abandoned area block ratio and the corresponding partitioned marked blocks obtained by the forward recognition network sub-model; The steps of constructing the reverse recognition network sub-model include: S201: Input the block-divided land image sequence to be identified and the number of divided blocks into the segmentation sublayer of the reverse recognition network submodel, and perform block segmentation according to the block boundary of each mark to obtain a block sequence set. ,in, Indicates the i The first j blocks, I Indicates the length of the land image sequence to be identified, J Indicates the number of blocks in each land image to be identified; S202: Mark each block with crops with the corresponding crop type, and mark the blocks without crops with 0, to obtain a sequence set of blocks with completed crop marking; S203: Input the crop-labeled block sequence set into the polarization feature layer in the reverse recognition network sub-model to obtain the polarization texture feature space corresponding to each block; S204, inputting the polarization texture feature space into the extended morphology multi-attribute profile layer to obtain the multi-dimensional polarization profile texture features corresponding to the polarization texture feature space of each block; S205. Input the multidimensional polarization profile texture features into the optimal classification feature extraction layer to obtain the optimal polarization profile texture features corresponding to the polarization texture feature space of each block. At the same time, input the multidimensional polarization profile texture features into the Gabor filter layer to obtain the polarization texture space channel features corresponding to the polarization texture feature space of each block. S206, inputting the optimal polarization profile texture features and polarization texture spatial channel features corresponding to each block polarization texture feature space into the convolutional attention fusion layer to obtain the polarization texture fusion spatial features corresponding to each block; S207, inputting the polarization texture fusion spatial features and the crop category features contained in the crop type recognition library simultaneously into the discriminant sublayer and the classification sublayer in the discriminant output layer, and obtaining the proportion of blocks containing crops and the probability of correct crop classification respectively; S208. Obtain the percentage of blocks that actually contain crops based on the labels marked in S202, and use the percentage of blocks that actually contain crops as a training threshold for the reverse recognition network sub-model, while also setting a probability threshold for correct crop classification. S209. The reverse recognition network sub-model is trained according to the two thresholds set in S208. When the proportion of blocks containing crops is equal to the proportion of blocks actually containing crops and the probability of correct crop classification is greater than the probability threshold of correct crop classification, a trained reverse recognition network sub-model is obtained.

2. The abandoned land identification method based on a bidirectional network and the truth value approximation principle according to claim 1 is characterized in that: The step of automatically marking land blocks includes: S101, preprocessing the acquired sequence of land images to be identified through filtering and brightness enhancement algorithms to obtain a preprocessed sequence of land images to be identified; S102: Using the block area, block boundary and shape, and plot slope data in the vector data, the corresponding blocks are vector-marked. The brightness and pixel color of the ridges inside the blocks and at the block boundaries are secondary-marked using the YUV color space. The marked blocks are numbered to obtain a marked sequence of land images to be identified. S103, constructing a farmland block division model, inputting the marked land image sequence to be identified into the farmland block division model to extract image content information and extract brightness and chromaticity difference features of the field ridges inside the block and at the block boundary, and constructing a comprehensive training loss function using the extracted brightness and chromaticity difference features and image content information extraction loss; S104, setting a comprehensive loss threshold and a training cycle, embedding the constructed comprehensive training loss function into the farmland block division model for training, and obtaining a trained farmland block division model when the comprehensive training loss function meets the comprehensive loss threshold or meets the training cycle; S105: The trained farmland block division model is configured in an image acquisition device, and the captured image is automatically divided into blocks and vector data marked to obtain a sequence of land images to be identified after the block division is completed and the number of divided blocks.

3. The abandoned land identification method based on a bidirectional network and the truth value approximation principle according to claim 2 is characterized in that: The steps of constructing the forward recognition network sub-model include: S301. Mark the potential abandoned land blocks that do not contain crops as abandoned land, uncultivated land, and forest, grassland and wetland; S302: Inputting the polarization texture fusion spatial features, vector data, and vegetation index features of the marked potential abandoned land blocks that do not contain crops into the MobileNet layer of the forward recognition network sub-model to extract non-crop vegetation features and obtain non-crop vegetation features; S303: Input the non-agricultural crop vegetation characteristics, the corresponding block rotation information, and the number characteristics of the divided blocks into the abandonment discrimination layer in the forward recognition network sub-model to obtain the proportion of abandoned area blocks and the corresponding marked blocks.

4. The abandoned land identification method based on a bidirectional network and the truth value approximation principle according to claim 3 is characterized in that: The steps of training the bidirectional recognition network model include: S501, subtracting the proportion of abandoned land blocks from the proportion of potentially abandoned land blocks to obtain the error in the proportion of non-abandoned land identification; S502: Subtract the percentage of non-abandoned land blocks at the abandoned land identification time point obtained in step S4 from the non-abandoned land identification percentage error to obtain a true value approximation error value; S503. Set a true value approximation error threshold, feed the true value approximation error value back to the bidirectional recognition network model for training, and when the true value approximation error value is less than the true value approximation error threshold, obtain a trained bidirectional recognition network model.

5. A system for identifying abandoned land using a bidirectional network based on the principle of truth approximation, which is used to implement the method for identifying abandoned land using a bidirectional network based on the principle of truth approximation according to any one of claims 1 to 4, characterized in that: include: Image marking module, reverse recognition module, forward recognition module and logic discrimination output module; The image marking module includes an image acquisition unit and a block automatic marking unit; the image acquisition unit is used to acquire a sequence of land images to be identified; The automatic block marking unit is used to automatically mark land blocks based on the sequence of land images to be identified using a vector marking model, and obtain the sequence of land images to be identified after block division, the number of divided blocks, and vector data corresponding to the divided blocks; The reverse recognition module is used to obtain a sequence of potential abandoned land to be identified that does not contain crops and the corresponding proportion of potential abandoned land blocks based on the configured crop type recognition library, the sequence of land images to be identified after block division, the vector data, and the number of divided blocks, using the reverse recognition network sub-model in the bidirectional recognition network model; The steps of constructing the reverse recognition network sub-model include: S201: Input the block-divided land image sequence to be identified and the number of divided blocks into the segmentation sublayer of the reverse recognition network submodel, and perform block segmentation according to the block boundary of each mark to obtain a block sequence set. ,in, Indicates the i The first j blocks, I Indicates the length of the land image sequence to be identified, J Indicates the number of blocks in each land image to be identified; S202: Mark each block with crops with the corresponding crop type, and mark the blocks without crops with 0, to obtain a sequence set of blocks with completed crop marking; S203: Input the crop-labeled block sequence set into the polarization feature layer in the reverse recognition network sub-model to obtain the polarization texture feature space corresponding to each block; S204, inputting the polarization texture feature space into the extended morphology multi-attribute profile layer to obtain the multi-dimensional polarization profile texture features corresponding to the polarization texture feature space of each block; S205. Input the multidimensional polarization profile texture features into the optimal classification feature extraction layer to obtain the optimal polarization profile texture features corresponding to the polarization texture feature space of each block. At the same time, input the multidimensional polarization profile texture features into the Gabor filter layer to obtain the polarization texture space channel features corresponding to the polarization texture feature space of each block. S206, inputting the optimal polarization profile texture features and polarization texture spatial channel features corresponding to each block polarization texture feature space into the convolutional attention fusion layer to obtain the polarization texture fusion spatial features corresponding to each block; S207, inputting the polarization texture fusion spatial features and the crop category features contained in the crop type recognition library simultaneously into the discriminant sublayer and the classification sublayer in the discriminant output layer, and obtaining the proportion of blocks containing crops and the probability of correct crop classification respectively; S208. Obtain the percentage of blocks that actually contain crops based on the labels marked in S202, and use the percentage of blocks that actually contain crops as a training threshold for the reverse recognition network sub-model, while also setting a probability threshold for correct crop classification. S209. The reverse recognition network sub-model is trained according to the two thresholds set in S208. When the proportion of blocks containing crops is equal to the proportion of blocks actually containing crops and the probability of correct crop classification is greater than the probability threshold of correct crop classification, a trained reverse recognition network sub-model is obtained.

6. The abandoned land identification system based on a two-way network and the truth value approximation principle according to claim 5, characterized in that: The forward recognition module is used to input the sequence of blocks of potential abandoned land that do not contain crops and vector data to be identified into the forward recognition network sub-model in the bidirectional recognition network model to obtain the proportion of abandoned land blocks and the corresponding marked blocks; The logic discrimination output module includes an experience information unit, a logic discrimination unit and a true value approximation training unit; The experience information unit is used to obtain land rotation information and forest, grassland and wetland block information data corresponding to different seasons in the historical area to be identified, and input the obtained data into the configured seasonal non-abandoned land ratio prediction model for training to obtain non-abandoned area block ratio information corresponding to different seasonal points; The logic discrimination unit is used to input information on the proportion of potentially abandoned blocks, the proportion of abandoned blocks, and the proportion of non-abandoned blocks at a corresponding time point into the logic discrimination layer of the bidirectional recognition network model to obtain a true value approximation error value; The true value approximation training unit is used to feed back the true value approximation error value to the bidirectional recognition network model for training until the true value approximation error value is less than or equal to a preset value, and output the abandoned area block ratio and the corresponding partitioned marked blocks obtained by the forward recognition network sub-model.

7. A computer-readable storage medium, characterized in that Computer instructions are stored thereon, and when the computer instructions are executed, the abandoned land identification method based on the bidirectional network of the truth value approximation principle as described in any one of claims 1 to 4 is executed.

Images