Idle farmland extraction method in complex farming areas, model training method and device, storage medium and terminal

Abstract

The present invention discloses a method for extracting idle farmland in complex cultivated areas, a model training method and device, a storage medium, and a terminal. The model training method includes obtaining a preliminary image feature dataset of a target area within a preset time period, obtaining a time series image dataset based on the preliminary image feature dataset, extracting texture features from the preliminary image feature data, and obtaining a texture spectral feature dataset; training a model based on the texture spectral feature dataset and the time series image dataset to obtain a teacher model, training a preset to-be-trained model based on the teacher model to obtain a student model, and using the student model and the teacher model as a model for extracting idle farmland in complex cultivated areas. The method summarizes different situations of idle farmland, clarifies the scope of idle farmland monitoring, and combines the ground feature characteristics and idle process of idle farmland to improve the recognition accuracy and generalization ability of idle farmland under complex planting patterns.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and in particular to a method for extracting idle farmland in a complex farming area, a model training method and device, a storage medium and a terminal. Background Art

[0002] Effectively obtaining information on idle cultivated land is the basic prerequisite for exploring the driving factors, distribution characteristics and influencing factors of idle cultivated land. It is also an important basis for formulating policies and measures according to local conditions and evaluating the ecological environment and socioeconomic effects.

[0003] Traditionally, idle farmland generally refers to long periods of inactivity. However, under current policies promoting farmland protection and a broader food security perspective, idle farmland has become increasingly diverse and complex. Existing research on idle farmland suffers from two main issues: First, unclear definitions of idle farmland targets lead to incomplete monitoring; second, complex cropping patterns exist, with fragmented and diverse land patterns and inconsistent planting schedules. These patterns include multiple cropping patterns for grain crops, and even multiple cropping patterns that rotate grain and non-grain crops. Currently, most research focuses on idle farmland itself, a concept that is not suitable for centralized cropping systems.

[0004] Therefore, it can be seen that the current idle cultivated land monitoring methods have the following problems: incomplete idle cultivated land monitoring, poor generalization, low idle cultivated land extraction accuracy, and failure to consider the complex planting patterns of cultivated land when judging idle cultivated land, especially for the extraction of idle cultivated land in complex farming areas. Summary of the Invention

[0005] In view of the shortcomings of the existing technology mentioned above, the purpose of this application is to provide a method for extracting idle farmland in complex farming areas, a model training method and device, a storage medium and a terminal, so as to solve the problems of incomplete idle farmland monitoring, poor generalization, low idle farmland extraction accuracy and failure to consider complex planting patterns of farmland when judging idle farmland in the current idle farmland monitoring method in complex farming areas.

[0006] In a first aspect, the present application provides a training method for extracting idle farmland from complex cultivated areas, comprising:

[0007] Acquire a preliminary image feature data set of a target area within a preset time period, wherein the preliminary image feature data set includes a preliminary image feature data subset of each piece of cultivated land within the target area, and the preliminary image feature data subset includes a preliminary image feature data group and a corresponding cultivated land label;

[0008] Based on the preliminary image feature dataset, a normalized vegetation index of each preset time interval of each piece of cultivated land in the target area is obtained, and the normalized vegetation index of all the preset time intervals of each piece of cultivated land is formed into time series image data of the corresponding cultivated land, and all the time series image data form a time series image dataset;

[0009] Extracting texture features of each preliminary image feature data in the preliminary image feature data set, and adding the texture features in the form of new bands to the corresponding preliminary image feature data in the preliminary image feature data set to obtain a texture spectral feature data set;

[0010] The Transformer model is trained based on the texture spectral feature dataset and the time series image dataset to obtain a teacher model, and a preset to-be-trained model is trained based on the teacher model to obtain a student model, and the teacher model is used as a decision maker and all the student models are used as input ends to form an idle farmland extraction model in a complex cultivated area;

[0011] Among them, the preset model to be trained includes a preset neural network model, a decision tree model and a long short-term memory network model. The data corresponding to the cultivated land label of non-idle cultivated land and the data corresponding to the cultivated land label of non-grain idle cultivated land in the texture spectral feature data set are used as the training data set of the preset neural network model, the data corresponding to the cultivated land label of year-round idle cultivated land in the time series image data set are used as the training data set of the long short-term memory network model, and the data corresponding to the cultivated land label of seasonal idle cultivated land in the texture spectral feature data set and the time series image data set are used as the training data set of the decision tree model; the preset time period is greater than two years.

[0012] In one embodiment of the present application, obtaining a preliminary image feature dataset of a target area within a preset time period includes:

[0013] An original image dataset of a target area within a preset time period is obtained, original image data in the original image dataset is preprocessed to obtain a preliminary image dataset, preliminary image data in the preliminary image dataset is cropped based on cultivated land range data to obtain a cultivated land range image dataset, and cultivated land corresponding to the cultivated land range image data in the cultivated land range image dataset is labeled based on the idle labeling criterion to obtain a preliminary image feature dataset.

[0014] In one embodiment of the present application, the original image data is satellite remote sensing image data, and atmospheric correction is performed on the satellite remote sensing image data to achieve pre-processing of the image data.

[0015] In one embodiment of the present application, the idle land marking criteria include idle farmland types and non-idle farmland types, and the idle farmland types include seasonal idle farmland, annual idle farmland, perennial idle farmland and non-grain idle farmland.

[0016] In one embodiment of the present application, training a preset to-be-trained model based on the teacher model to obtain a student model includes:

[0017] Using feature-based knowledge distillation, the features acquired by the teacher model are aligned with the features of the preset neural network model in terms of feature dimensions and are transferred to the middle layer of the preset neural network model to obtain a neural network student model;

[0018] Feature-based knowledge distillation is used to align the features acquired by the teacher model with the features of the long short-term memory network model and pass them to the middle layer of the long short-term memory network model to obtain a long short-term memory network student model.

[0019] In one embodiment of the present application, training a preset to-be-trained model based on the teacher model to obtain a student model further includes:

[0020] The soft labels obtained by the Softmax function of the teacher model are used to assist in training the decision tree model using target-based Logits distillation to obtain a decision tree student model.

[0021] In one embodiment of the present application, texture features in the preliminary image feature data are extracted using a gray level co-occurrence matrix.

[0022] In a second aspect, the present application provides a model training device for extracting idle farmland in complex cultivated areas, comprising a preliminary image feature dataset acquisition module, a time series image dataset acquisition module, a texture spectrum feature dataset acquisition module, and a model training module;

[0023] The preliminary image feature data set acquisition module is used to acquire a preliminary image feature data set of a target area within a preset time period, wherein the preliminary image feature data set includes a preliminary image feature data subset of each piece of cultivated land in the target area, and the preliminary image feature data subset includes a preliminary image feature data group and a corresponding cultivated land label;

[0024] The time series image dataset acquisition module is used to obtain the normalized vegetation index of each preset time interval of each piece of cultivated land in the target area based on the preliminary image feature dataset, and form the normalized vegetation index of all the preset time intervals of each piece of cultivated land into the time series image data of the corresponding cultivated land, and all the time series image data form a time series image dataset;

[0025] The texture spectral feature data set acquisition module is used to extract the texture features of each preliminary image feature data in the preliminary image feature data set, and add the texture features in the form of new bands to the corresponding preliminary image feature data in the preliminary image feature data set to obtain a texture spectral feature data set;

[0026] The model training module is used to train the Transformer model based on the texture spectral feature dataset and the time series image dataset to obtain a teacher model, and to train a preset to-be-trained model based on the teacher model to obtain a student model, and to form an idle farmland extraction model in a complex cultivated area with the teacher model as the decision maker and all student models as input ends;

[0027] Among them, the preset model to be trained includes a preset neural network model, a decision tree model and a long short-term memory network model. The data corresponding to the cultivated land label of non-idle cultivated land and the data corresponding to the cultivated land label of non-grain idle cultivated land in the texture spectral feature data set are used as the training data set of the preset neural network model, the data corresponding to the cultivated land label of year-round idle cultivated land in the time series image data set are used as the training data set of the long short-term memory network model, and the data corresponding to the cultivated land label of seasonal idle cultivated land in the texture spectral feature data set and the time series image data set are used as the training data set of the decision tree model; the preset time period is greater than two years.

[0028] In a third aspect, the present application provides a method for extracting idle farmland in a complex farming area, comprising:

[0029] Acquire a target image feature data set of a to-be-extracted area within a target time period, wherein the target image feature data set includes a target image feature data subset of each piece of cultivated land within the to-be-extracted area, and the target image feature data subset includes a target image feature data group;

[0030] Based on the target image feature dataset, a normalized vegetation index of each preset time interval of each piece of cultivated land in the area to be extracted is obtained, and the normalized vegetation index of all the target time zones of each piece of cultivated land is used to form target time series image data of the corresponding cultivated land, and all the target time series image data form a target time series image dataset;

[0031] Extracting texture features of each target image feature data in the target image feature data set, and adding the texture features in the form of new bands to the corresponding target image feature data in the target image feature data set to obtain a target texture spectral feature data set;

[0032] The target texture spectral feature data set is transmitted to the neural network student model in the idle farmland extraction model in the complex farming area to obtain the initial distribution data of non-idle farmland and non-grain idle farmland, the target time series image data set is transmitted to the long short-term memory network student model in the idle farmland extraction model in the complex farming area to obtain the initial distribution data of annual idle farmland and perennial farmland, the target texture spectral feature data and the target time series image data are transmitted to the decision tree student model in the idle farmland extraction model in the complex farming area to obtain the initial distribution data of seasonal idle farmland, the target texture spectral feature data set, the target time series image data set, the initial distribution data of non-idle farmland and non-grain idle farmland, the initial distribution data of annual idle farmland and perennial farmland, and the initial distribution data of seasonal idle farmland are all input into the teacher model in the idle farmland extraction model in the complex farming area to obtain the distribution of farmland in the area to be extracted;

[0033] The idle farmland extraction model for complex farming areas is obtained based on the idle farmland extraction model training method for complex farming areas.

[0034] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the idle farmland extraction model training method for complex farmland areas or implements the idle farmland extraction method for complex farmland areas.

[0035] In a fifth aspect, the present application provides a terminal comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the idle farmland extraction model training method for complex farmland areas or the idle farmland extraction method for complex farmland areas.

[0036] Compared with the prior art, one or more embodiments of the above solutions may have the following advantages or beneficial effects:

[0037] The training method for extracting idle farmland in complex cultivated areas, provided by an embodiment of the present invention, generates farmland samples by setting up multiple farmland types. These samples with different characteristics are then fed into different student models for training. A transformer is used as a teacher model to improve the student model, allowing it to quickly learn the knowledge of the teacher model. The trained student models are then integrated into a model for extracting idle farmland in complex cultivated areas, improving the model's robustness and achieving higher recognition accuracy. Furthermore, the accuracy and generalization of idle farmland recognition under complex planting patterns are enhanced. This method summarizes different idle farmland scenarios and clarifies the scope of idle farmland monitoring. Combining the features and idleness processes of idle farmland, it uses feature-based rules and change detection-based recognition methods to generate samples, achieving accurate identification of idle farmland. Furthermore, through time series tracking and texture validation, interference from similar features such as tillage and fallow is avoided. Continuous monthly data is used to reflect the detailed changes in vegetation growth during different growth periods, effectively addressing the identification of different idle farmland types.

[0038] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purposes and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] 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:

[0040] Figure 1 Shown is a flow chart of the idle farmland extraction model training method for complex farming areas described in an embodiment of the present application.

[0041] Figure 2 Shown is a process diagram of the idle farmland extraction model training method for complex farming areas described in an embodiment of the present application.

[0042] Figure 3 Shown is a structural schematic diagram of the idle farmland extraction model training device in complex farming areas described in an embodiment of the present application.

[0043] Figure 4 Shown is a flow chart of the method for extracting idle farmland in complex farming areas described in an embodiment of the present application.

[0044] Figure 5 Shown is a schematic diagram of the structure of the terminal described in an embodiment of the present application. DETAILED DESCRIPTION

[0045] The following describes the embodiments of the present invention in detail with reference to the accompanying drawings and examples, so that the present invention can fully understand how to apply technical means to solve technical problems and achieve technical effects, and thus implement the invention accordingly. It should be noted that, as long as no conflict exists, the various embodiments of the present invention and the various features of the embodiments can be combined with each other, and the resulting technical solutions are all within the scope of protection of the present invention.

[0046] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present application. Therefore, the illustrations only show components related to the present application and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0047] The large-scale, high-resolution, and precise idle farmland monitoring work currently being carried out in my country has failed to integrate the two different perspectives of idle farmland characteristics and the idle farmland process. It only considers the idea of ​​idle farmland itself, and fails to maximize the advantages of high-frequency and high-resolution images based on the overall structure of farmland planting.

[0048] The following embodiments of the present application provide a method for extracting idle farmland in complex farming areas, a model training method and device, a storage medium and a terminal, which solve the problems of the current idle farmland monitoring methods, such as incomplete idle farmland monitoring, poor generalization, low idle farmland extraction accuracy, and failure to consider complex planting patterns of farmland when judging idle farmland.

[0049] Complex farming areas refer to cultivated land areas with frequent adjustments in planting structure, diverse planting structure, and large differences in planting structure within the region. Frequent adjustments in planting structure mean that the planting pattern is different every year. For example, this year it is rice-rice rotation, and next year it is wheat-rice rotation. Diverse planting structure means that general farming areas may plant a certain crop at the same time, but complex farming areas will have large time differences. Large differences in planting structure within the region mean that similar areas in general farming areas plant roughly the same things, but similar areas in complex farming areas plant very different things.

[0050] The following will describe in detail the principles and implementation methods of a method for extracting idle farmland in a complex farming area, a model training method and device, a storage medium, and a terminal of this embodiment in conjunction with the accompanying drawings, so that those skilled in the art can understand the method for extracting idle farmland in a complex farming area, a model training method and device, a storage medium, and a terminal of this embodiment without the need for creative work.

[0051] like Figure 1 As shown, this embodiment provides a training method for extracting idle farmland in complex farming areas, including the following steps.

[0052] Step S101: obtaining a preliminary image feature dataset of a target area within a preset time period.

[0053] The preliminary image feature data set includes a preliminary image feature data subset of each piece of cultivated land in the area to be extracted, and the preliminary image feature data subset includes a preliminary image feature data group and a cultivated land label of the corresponding cultivated land.

[0054] Step S102: Based on the preliminary image feature dataset, the normalized vegetation index of each preset time interval of each piece of cultivated land in the target area is obtained, and the normalized vegetation index of all preset time zones of each piece of cultivated land is formed into time series image data of the corresponding cultivated land, and all time series image data form a time series image dataset.

[0055] It should be noted that since the time series image data is formed based on the normalized vegetation index of all preset time zones corresponding to a single piece of cultivated land, each time series image data corresponds to a piece of cultivated land. Since each piece of cultivated land has its corresponding cultivated land label in the preliminary image feature dataset, each time series image data has its corresponding cultivated land label. Further, the time series image dataset includes multiple time series image data and the cultivated land label corresponding to each time series image data.

[0056] Step S103 , extracting texture features of each preliminary image feature data in the preliminary image feature data set, and adding the texture features in the form of new bands to the corresponding preliminary image feature data in the preliminary image feature data set to obtain a texture spectrum feature data set.

[0057] Step S104: Train the Transformer model based on the texture spectral feature dataset and the time series image dataset to obtain a teacher model, and train the preset model to be trained based on the teacher model to obtain a student model. Use the teacher model as the decision maker and all student models as input ends to form an idle farmland extraction model in complex cultivated areas.

[0058] Among them, the preset models to be trained include a preset neural network model (which can be a residual neural network (ResNet)), a decision tree model and a long short-term memory network model (LSTM). The data corresponding to the cultivated land label of non-idle cultivated land and the data corresponding to the cultivated land label of non-grain idle cultivated land in the texture spectral feature dataset are used as training data sets for the preset neural network model, the data corresponding to the cultivated land label of annual idle cultivated land in the time series image dataset are used as training data sets for the long short-term memory network model, and the data corresponding to the cultivated land label of seasonal idle cultivated land in the texture spectral feature dataset and the time series image dataset are used as training data sets for the decision tree model; the preset time period is greater than two years.

[0059] The training method for extracting idle farmland in complex cultivated areas, provided by an embodiment of the present invention, creates farmland samples by setting up multiple farmland types. Samples with different characteristics are then fed into different student models for training. A transformer is used as a teacher model to improve the student model, allowing the student model to quickly learn the knowledge of the teacher model. Finally, the trained multiple student models are integrated into a model for extracting idle farmland in complex cultivated areas, thereby improving the model's robustness and achieving higher recognition accuracy. This improves the recognition accuracy and generalization capability of idle farmland under complex planting patterns. This method addresses the current problems of incomplete idle farmland monitoring, poor generalization, low idle farmland extraction accuracy, and the failure to consider complex planting patterns when determining idle farmland.

[0060] The following combination Figure 2 The content shown in FIG. 1 is a detailed description of steps S101 to S104 for obtaining the idle farmland extraction model in the complex farming area in this embodiment.

[0061] Specifically, in step S101 , a preliminary image feature dataset of a target area within a preset time period is obtained.

[0062] Set the target region. This can be a sufficiently large area of ​​multiple cultivated lands, or a sufficiently large area formed by some non-adjacent cultivated land areas. This setting ensures that the cultivated land types are sufficiently rich and the final training dataset has a sufficient amount of data. The target region can be set as a complex cultivated area.

[0063] The raw image data of the target area captured from a high altitude within a preset time period is obtained using existing data acquisition devices. The raw image data obtained is typically satellite remote sensing image data. Furthermore, the raw image data can be set to 2-meter resolution cloud-free satellite remote sensing image data. Alternatively, other image data reflecting the cultivated land conditions in the target area can be obtained as raw image data. It should be noted that the above data acquisition method can be set according to actual circumstances. For example, the raw image data of the target area can be collected once at noon each day, or twice a day in the morning and afternoon. In this embodiment, all raw image data collected for the target area within the preset time period are combined into a raw image data set.

[0064] It should be noted that to ensure the practicality of the image data, in the process of forming the original image dataset, it is necessary to select images that can guarantee continuous monthly coverage in the original image data, and use the intersection range of the collected images to reasonably crop the images to obtain complete original image data. The continuous monthly coverage area is the area where the corresponding image exists almost continuously during the preset time period. For example, if some areas are blocked by dark clouds or large objects for a long time, the image data of this area will not be able to achieve temporal continuity and must be discarded. However, the probability of such events occurring is low, so the discarding action will not have a significant impact on the final result.

[0065] After obtaining a raw image dataset of the target area within a preset time period, each raw image data in the raw image dataset can be preprocessed to obtain corresponding preliminary image data. All preliminary image data are then aggregated to form a preliminary image dataset. When the raw image data is satellite remote sensing image data, atmospheric correction can be used to preprocess the raw image data to remove atmospheric influences on the image. Each preliminary image data in the preliminary image dataset is then cropped based on the cultivated land extent data to obtain the cultivated land extent image data corresponding to each preliminary image data. Since each preliminary image data is specific to the target area, the aforementioned cropping process involves cropping each preliminary image data in the same manner, thereby cropping each preliminary image data into cultivated land extent image data corresponding to multiple plots of cultivated land according to the cultivated land extent. Finally, all cultivated land extent image data are aggregated into a cultivated land extent image dataset. It should be noted that the cultivated land extent corresponding to each cultivated land extent image data obtained after cropping is clear, i.e., the extent of each plot of cultivated land corresponding to the cultivated land extent image data has been determined. All cultivated land image data for the same cultivated land in the cultivated land image dataset is then used as the corresponding preliminary image feature data group. The cultivated land corresponding to each preliminary image feature data group is labeled using the idle labeling criterion, thereby obtaining a preliminary image feature dataset. This cultivated land labeling process can be implemented in conjunction with camera or field verification. The cultivated land area data is obtained by pre-automated extraction of cultivated land areas and verification using camera and field sampling.

[0066] Furthermore, since each piece of cultivated land in the target area has its corresponding preliminary image feature data group and corresponding cultivated land label, if the preliminary image feature data group and the corresponding cultivated land label corresponding to each piece of cultivated land are grouped into a preliminary image feature data subset of the corresponding cultivated land, then each piece of cultivated land has its corresponding preliminary image feature data subset, and all preliminary image feature data subsets form a preliminary image feature data set.

[0067] Idle land marking criteria include both idle and non-idle farmland types. Specific idle land types include seasonal idle land, year-round idle land, perennial idle land, and non-grain idle land. Non-idle farmland refers to farmland that is not idle year-round. Idle land refers to idle land. Seasonal idle land refers to farmland that should be planted with crops for two or more seasons in a year but remains idle for one or more seasons. Year-round idle land refers to farmland that has not been planted with any crops for a year. Perennial idle land refers to farmland that has not been planted with any crops for more than two consecutive years. Non-grain idle land refers to farmland planted with non-grain crops within permanent basic farmland, including fruit trees, seedlings, greenhouses, and grassland. When marking non-grain idle land, the permanent basic farmland boundary must be used as an auxiliary marker. Table 1 presents the idle land marking criteria in tabular form.

[0068] Table 1

[0069]

[0070] It should be noted that when labeling farmland types, the idle land labeling criteria can be used to digitize each farmland type. For example, seasonal idle land, year-round idle land, perennial idle land, and non-grain idle land can be labeled as 100, 150, 200, and 250, respectively; non-idle farmland is labeled as a negative sample and is marked as 0. The idle land labeling criteria can also be set to other reasonable methods, and each idle and non-idle farmland type can also be labeled using other labeling methods. This embodiment does not impose any additional restrictions on this.

[0071] From the above content, it can be seen that perennially idle cultivated land needs to be determined based on at least two years of cultivated land data. Therefore, in order to ensure that the idle cultivated land extraction model for complex cultivated areas finally obtained can judge perennially idle cultivated land, the preset time period set in this embodiment needs to be greater than two years.

[0072] Step S102: Based on the preliminary image feature dataset, the normalized vegetation index of each preset time interval of each piece of cultivated land in the target area is obtained, and the normalized vegetation index of all preset time zones of each piece of cultivated land is formed into time series image data of the corresponding cultivated land, and all time series image data form a time series image dataset.

[0073] Each piece of cultivated land in the target area has a corresponding subset of preliminary image feature data. The preliminary image feature data group in the preliminary image feature data subset includes all preliminary image feature data within a preset time period for the corresponding cultivated land. If the preset time period is divided into multiple preset time intervals, each preset time interval has multiple corresponding preliminary image feature data. At this time, based on the preliminary image feature data corresponding to each preset time interval of each piece of cultivated land, the normalized difference vegetation index (NDVI) of the corresponding cultivated land corresponding to the preset time interval can be calculated. For example, if the preset time period is one year and the preset time interval is monthly, the normalized difference vegetation index of a piece of cultivated land in January can be calculated based on the preliminary image feature data corresponding to the month of January. Similarly, the normalized difference vegetation index of the cultivated land from February to December can also be calculated separately. It should be noted that since the preliminary image feature data group of each piece of cultivated land in the target area has its corresponding cultivated land label, and the normalized vegetation index corresponding to each piece of cultivated land is obtained based on the preliminary image feature data in the corresponding preliminary image feature data group, each piece of cultivated land has its corresponding multiple normalized vegetation indices and cultivated land labels. At this time, all the normalized vegetation indices of each piece of cultivated land are arranged in chronological order, and the corresponding cultivated land labels are added to the corresponding sequence to obtain the time series image data corresponding to each piece of cultivated land. Finally, all the time series image data are combined to obtain the time series image dataset.

[0074] Step S103 , extracting texture features of each preliminary image feature data in the preliminary image feature data set, and adding the texture features in the form of new bands to the corresponding preliminary image feature data in the preliminary image feature data set to obtain a texture spectrum feature data set.

[0075] Specifically, the texture features of each preliminary image feature data in the preliminary image feature dataset are extracted using a gray level co-occurrence matrix, and then the extracted texture features are added to the corresponding preliminary image feature data in the preliminary image feature dataset in the form of new bands, thereby converting the preliminary image feature data in the preliminary image feature dataset into texture spectral feature data. Further, a subset of the preliminary image feature data in the preliminary image feature dataset is converted into a texture spectral feature data subset. The texture spectral feature data subset should include a texture spectral feature data group and a corresponding cultivated land label. All texture spectral feature data subsets can be aggregated into a texture spectral feature dataset.

[0076] It should be noted that the texture features of the preliminary image feature data can also be extracted by other reasonable methods, which will not be described in detail here.

[0077] Step S104: Train the Transformer model based on the texture spectral feature dataset and the time series image dataset to obtain a teacher model, and train the preset model to be trained based on the teacher model to obtain a student model. Use the teacher model as the decision maker and all student models as input ends to form an idle farmland extraction model in complex cultivated areas.

[0078] This embodiment adopts a teacher-student model training method to implement model training, specifically setting the Transformer model as the teacher model to be trained, setting the preset model to be trained as the student model to be trained, and further setting the preset model to be trained to include a preset neural network model, a decision tree model and a long short-term memory network model.

[0079] The teacher-student model training process is as follows: first, the texture spectral feature dataset and time series image dataset obtained in the above steps serve as the total training dataset for the teacher model. The total training dataset is then divided into a training dataset and a test dataset according to a certain ratio for training the Transformer model. Training is stopped when the overall model training accuracy reaches a preset level (the preset level can be set to above 95%) to obtain the teacher model. The preset neural network model, decision tree model, and long short-term memory network model in the preset to-be-trained model are then trained based on the teacher model to obtain the neural network student model corresponding to the preset neural network model, the decision tree student model corresponding to the decision tree model, and the long short-term memory network student model corresponding to the long short-term memory network model. Finally, the idle farmland extraction model for complex cultivated areas is formed using the teacher model as the decision maker and the neural network student model, decision tree student model, and long short-term memory network student model as input. During the training of the preset to-be-trained model based on the teacher model, data corresponding to farmland labeled as non-idle farmland and data corresponding to farmland labeled as non-grain-producing idle farmland in the texture spectral feature dataset can be used as the training dataset for the preset neural network model. Furthermore, the data corresponding to the cultivated land labeled as non-idle cultivated land in the texture spectral feature dataset and the data corresponding to the cultivated land labeled as non-grain idle cultivated land in the texture spectral feature dataset are combined as the neural network model training dataset, and in the teacher-student model training process, feature-based knowledge distillation is used to align the feature dimensions of the features obtained by the teacher model with the features of the preset neural network model and pass them to the middle layer of the preset neural network model to obtain the neural network student model.

[0080] During the training of the pre-set training model based on the teacher model, data corresponding to cultivated land labeled as year-round idle in the time series image dataset can be used as the training dataset for the LSTM network model. Furthermore, data corresponding to cultivated land labeled as year-round idle in the time series image dataset can be used as the training dataset for the LSTM network. The LSTM network model is trained using the time series image dataset to identify year-round idle cultivated land. If the same cultivated land is identified as year-round idle for two or more consecutive years, it is considered perennially idle. During the training of the teacher-student model, feature-based knowledge distillation is used to align the features acquired by the teacher model with those of the LSTM network model, and the features are transferred to the middle layer of the LSTM network model to obtain the LSTM network student model.

[0081] During the training of the pre-set model based on the teacher model, data corresponding to the seasonally idle farmland label in the texture spectral feature dataset and the time series image dataset can be used as the training dataset for the decision tree model. Furthermore, data corresponding to the seasonally idle farmland label in the texture spectral feature dataset and the time series image dataset can be used as the training dataset for the decision tree model. On the one hand, the texture spectral feature data can be used to determine whether crops are planted, and on the other hand, the time series variation characteristics can be used to determine whether crops have not been planted for multiple consecutive months. During the training of the teacher and student models, target-based logits distillation is used to use the soft labels obtained by the teacher model's softmax function to assist in training the decision tree model, thereby obtaining a decision tree student model.

[0082] In this training process, the student model is trained based on the teacher model, allowing it to quickly learn the teacher model's knowledge. The teacher model's predictions and validation set data are then used to train the student model. Ultimately, the teacher model serves as the decision maker, and all student models serve as input to form a model for extracting idle farmland from complex cultivated areas. Compared to pre-training the student model as a teacher model, the teacher model in the resulting model uses a multi-head attention mechanism to fuse the results of the three student models. This ensures both accuracy and synchronization between the teacher and student models, effectively ensuring overall precision.

[0083] This method summarizes the different situations of idle cultivated land, clarifies the scope of idle cultivated land monitoring, combines the ground features of idle cultivated land with the idle process, and uses both feature-based rule-based and change detection-based recognition methods to generate samples for accurate identification of idle cultivated land. Although samples of different idle situations were generated, the actual situation is more complex. Planting patterns are affected by factors such as climate, soil, and economy in different regions. Even the planting time of the same crop can vary by more than ten days, and there are also situations where different crops are mixed. A single model alone is difficult to handle and its accuracy is still lacking. By training multiple models and integrating them, the robustness of the model is improved and higher accuracy is achieved. Through time series tracking and texture, the interference of similar ground features such as tillage and fallow is effectively avoided. Continuous monthly data is used to reflect the subtle changes in different vegetation growth periods, which solves the problem of identifying different types of idle cultivated land in a targeted manner.

[0084] like Figure 3 As shown, this embodiment also provides a model training device for extracting idle farmland in complex farming areas, including a preliminary image feature dataset acquisition module, a time series image dataset acquisition module, a texture spectrum feature dataset acquisition module and a model training module;

[0085] The preliminary image feature data set acquisition module is used to obtain a preliminary image feature data set of the target area within a preset time period. The preliminary image feature data set includes a preliminary image feature data subset of each piece of cultivated land in the target area. The preliminary image feature data subset includes a preliminary image feature data group and a corresponding cultivated land label.

[0086] The time series image dataset acquisition module is used to obtain the normalized vegetation index of each preset time interval of each piece of cultivated land in the target area based on the preliminary image feature dataset, and to form the time series image data of the corresponding cultivated land with the normalized vegetation index of all preset time zones of each piece of cultivated land. All time series image data form a time series image dataset.

[0087] The texture spectral feature dataset acquisition module is used to extract the texture features of each preliminary image feature data in the preliminary image feature dataset, and add the texture features to the corresponding preliminary image feature data in the preliminary image feature dataset in the form of new bands to obtain the texture spectral feature dataset.

[0088] The model training module is used to train the Transformer model based on the texture spectral feature dataset and the time series image dataset to obtain a teacher model, and to train the preset model to be trained based on the teacher model to obtain a student model. The teacher model is used as the decision maker, and all student models are used as input to form an idle farmland extraction model in complex cultivated areas.

[0089] Among them, the preset models to be trained include a preset neural network model, a decision tree model and a long short-term memory network model. The data corresponding to the cultivated land label of non-idle cultivated land and the data corresponding to the cultivated land label of non-grain idle cultivated land in the texture spectral feature data set are used as training data sets for the preset neural network model, the data corresponding to the cultivated land label of annual idle cultivated land in the time series image data set are used as training data sets for the long short-term memory network model, and the data corresponding to the cultivated land label of seasonal idle cultivated land in the texture spectral feature data set and the time series image data set are used as training data sets for the decision tree model; the preset time period is greater than two years.

[0090] This device summarizes the different situations of idle farmland, clarifies the scope of idle farmland monitoring, combines the ground features of idle farmland with the idle process, and uses both feature-based rule-based and change detection-based recognition methods to create samples for accurate identification of idle farmland. Although samples of different idle situations have been created, the actual situation is more complicated. Planting patterns are affected by factors such as climate, soil, and economy in different regions. Even the planting time of the same crop may deviate by more than ten days, and there may be mixed planting of different crops. A single model alone is difficult to cope with and its accuracy is still lacking. By training multiple models and integrating them, the robustness of the model is improved and higher accuracy is achieved. Through time series tracking and texture, the interference of similar ground features such as tillage and fallow is effectively avoided. Continuous monthly data is used to reflect the subtle changes in the growth period of different vegetation, which solves the problem of identifying different types of idle farmland in a targeted manner.

[0091] like Figure 4 As shown, this embodiment also provides a method for extracting idle farmland in complex farming areas, including the following steps.

[0092] Step S301: Obtain a target image feature dataset of the area to be extracted within a target time period. The target image feature dataset includes a target image feature data subset of each piece of farmland in the area to be extracted, and the target image feature data subset includes a target image feature data group.

[0093] The target original image data of the area to be extracted, taken from a high altitude during a target time period, is obtained through existing data acquisition devices. The target original image data obtained is usually satellite remote sensing image data. It can further be set to 2-meter resolution cloud-free satellite remote sensing image data. Other image data that can reflect the cultivated land conditions in the area to be extracted can also be obtained as the target original image data. It should be noted that the above data acquisition method can be set according to actual conditions. For example, it can be set to collect target original image data of the area to be extracted once at 12 noon every day, or it can be set to collect image data of the area to be extracted twice a day in the morning and at noon. In this embodiment, all target original image data collected for the area to be extracted during the target time period are combined into a target original image data set.

[0094] It should be noted that in order to ensure the practicality of the image data, in the process of forming the target original image data set, it is necessary to select images that can guarantee continuous monthly coverage areas from the target original image data, and use the intersection range of the collected images to reasonably crop the images to obtain the target original image data with complete data.

[0095] After obtaining the target original image dataset of the area to be extracted within the target time period, each target original image data in the target original image dataset can be preprocessed to obtain the corresponding target preliminary image data. All target preliminary image data are then aggregated to form the target preliminary image dataset. When the target original image data is satellite remote sensing image data, atmospheric correction can be used to preprocess the target original image data to remove the influence of the atmosphere on the image. The target preliminary image data in the target preliminary image dataset are then clipped based on the cultivated land extent data to obtain the cultivated land extent image data corresponding to each target preliminary image data. That is, each target preliminary image data is clipped in the same manner, and then each target preliminary image data is clipped according to the cultivated land extent into target cultivated land extent image data for multiple plots of cultivated land. Finally, all target cultivated land extent image data are aggregated into the cultivated land extent image dataset. Then, the target cultivated land range image data corresponding to each piece of cultivated land is used as the target image feature data of the corresponding cultivated land, and all the target image feature data corresponding to the piece of cultivated land are grouped into a target image feature data group, and the target image feature data subset is set to include only the target image feature data group. Finally, the target image feature data subsets of all cultivated land in the area to be extracted are grouped into a target image feature data set.

[0096] It should be noted that the target image feature data subset in this embodiment does not include the farmland label.

[0097] Step S302: Based on the target image feature dataset, the normalized vegetation index of each preset time interval of each piece of cultivated land in the area to be extracted is obtained, and the normalized vegetation index of all target time zones of each piece of cultivated land is formed into the target time series image data of the corresponding cultivated land, and all target time series image data form the target time series image dataset.

[0098] Each plot of cultivated land in the area to be extracted has a corresponding subset of target image feature data. This subset includes all target image feature data for the corresponding plot within a target time period. If the target time period is divided into multiple preset time intervals, each preset time interval will have multiple corresponding target image feature data. Based on the target image feature data corresponding to each preset time interval for each plot of cultivated land, the Normalized Difference Vegetation Index (NDVI) for that preset time interval can be calculated. For example, if the target time period is one year and the preset time intervals are monthly, the NDVI for a particular plot of cultivated land can be calculated based on the target image feature data corresponding to January. Similarly, the NDVI for the same plot of cultivated land can be calculated from February to December. Since each plot of cultivated land has multiple corresponding NDVIs, all NDVIs for each plot of cultivated land can be arranged in chronological order to obtain the time series image data corresponding to each plot of cultivated land. Finally, all the time series image data are aggregated to obtain a time series image dataset. It should be noted that the time series image data in this embodiment does not include farmland labels.

[0099] Step S303 , extracting texture features of each target image feature data in the target image feature dataset, and adding the texture features to the corresponding target image feature data in the target image feature dataset in the form of new bands to obtain a target texture spectrum feature dataset.

[0100] Specifically, the texture features of each target image feature data in the target image feature dataset are extracted through the gray level co-occurrence matrix, and then the extracted texture features are added to the corresponding target image feature data in the form of new bands to realize the conversion of the target image feature data in the target image feature dataset into target texture spectral feature data. Finally, all the target texture spectral feature data corresponding to each piece of cultivated land are grouped into the target texture spectral feature data group of the corresponding cultivated land, and the target texture spectral feature data subset is set to include only the target texture spectral feature data group. Finally, the target texture spectral feature data subsets of all the cultivated land in the area to be extracted are grouped into the target texture spectral feature dataset.

[0101] It should be noted that the texture features of the target image feature data can also be extracted by other reasonable methods, which will not be described in detail here.

[0102] Step S304, transmitting the target texture spectral feature data set to the neural network student model in the idle farmland extraction model in the complex farming area to obtain the initial distribution data of non-idle farmland and non-grain idle farmland, transmitting the target time series image data set to the long short-term memory network student model in the idle farmland extraction model in the complex farming area to obtain the initial distribution data of annual idle farmland and perennial farmland, transmitting the target texture spectral feature data and the target time series image data to the decision tree student model in the idle farmland extraction model in the complex farming area to obtain the initial distribution data of seasonal idle farmland, and inputting the target texture spectral feature data set, the target time series image data set, the initial distribution data of non-idle farmland and non-grain idle farmland, the initial distribution data of annual idle farmland and perennial farmland, and the initial distribution data of seasonal idle farmland into the teacher model in the idle farmland extraction model in the complex farming area to obtain the distribution of farmland in the area to be extracted.

[0103] The idle farmland extraction model for complex cultivated areas in the above embodiment is used as the idle farmland extraction model for complex cultivated areas obtained by the idle farmland extraction model training method for complex cultivated areas, wherein the idle farmland extraction model for complex cultivated areas includes a neural network student model, a long short-term memory network student model, a decision tree student model, and a teacher model. The target texture spectral feature data set is transferred to the neural network student model in the idle farmland extraction model for complex cultivated areas to obtain the initial distribution of non-idle cultivated land and the initial distribution of non-grain idle cultivated land in the area to be extracted, that is, to obtain the initial distribution data of non-idle cultivated land and non-grain idle cultivated land in the area to be extracted. The target time series image data set is transferred to the long short-term memory network student model in the idle farmland extraction model for complex cultivated areas to obtain the initial distribution of year-round idle cultivated land and the initial distribution of perennial cultivated land in the area to be extracted, that is, to obtain the initial distribution data of year-round idle cultivated land and perennial cultivated land in the area to be extracted. The target texture spectral feature data and target time series image data are fed into the decision tree student model in the idle farmland extraction model for complex cultivated areas to obtain the seasonal idle farmland distribution in the extraction area, i.e., to obtain the initial distribution data of seasonal idle farmland in the extraction area. Finally, the target texture spectral feature dataset, target time series image dataset, initial distribution data of non-idle farmland and non-grain idle farmland, initial distribution data of annual idle farmland and perennial farmland, and initial distribution data of seasonal idle farmland are all fed into the teacher model in the idle farmland extraction model for complex cultivated areas to obtain the distribution of farmland in the extraction area. This process integrates the three targeted student models into a single model through the teacher model Transformer. During extraction, the relatively specialized student model first determines the probability of the input farmland belonging to a certain idle type, and the relatively versatile transformer then obtains the final judgment result.

[0104] This method summarizes the different scenarios of idle farmland and clarifies the scope of idle farmland monitoring. Combining the features of idle farmland with its idleness process, it uses both feature-based rules and change detection-based recognition methods to generate samples for accurate identification of idle farmland. By using time-series tracking and texture analysis, it effectively avoids interference from similar features such as tillage and fallow land. Using continuous monthly data to capture the subtle changes in vegetation across different growth periods, it effectively addresses the identification of different idle farmland types.

[0105] The present application also provides a computer-readable storage medium. Those skilled in the art will appreciate that all or part of the steps in the above-described method can be performed by instructing a processor through a program. The program can be stored in a computer-readable storage medium, which is a non-transitory medium, such as random access memory, read-only memory, flash memory, a hard disk, a solid-state drive, magnetic tape, a floppy disk, an optical disc, or any combination thereof. The storage medium can be any available medium accessible by a computer, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a magnetic tape), an optical medium (e.g., a digital video disc (DVD)), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0106] like Figure 5 As shown, an embodiment of the present application provides a terminal.

[0107] The terminal of this embodiment includes a processor and a memory connected to each other; the memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory, so that all or part of the steps in the above embodiment method can be implemented when the terminal executes them.

[0108] The beneficial effects of all or part of the steps of the method in the above embodiment are the same as the beneficial effects obtained by applying the terminal provided by the embodiment of the present invention, and will not be described in detail here.

[0109] It should be noted that the memory may include random access memory (RAM) and non-volatile memory (non-volatile memory), such as at least one disk storage device. Similarly, the processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0110] Although the embodiments disclosed herein are as described above, the contents described herein are merely embodiments for facilitating understanding of the present invention and are not intended to limit the present invention. Any person skilled in the art may make any modifications and variations in the form and details of the embodiments without departing from the spirit and scope of the present invention. However, the scope of protection of the present invention shall remain subject to the scope defined by the appended claims.

Claims

1. A training method for extracting idle farmland in complex cultivated areas, characterized by: include: Acquire a preliminary image feature data set of a target area within a preset time period, wherein the preliminary image feature data set includes a preliminary image feature data subset of each piece of cultivated land within the target area, and the preliminary image feature data subset includes a preliminary image feature data group and a corresponding cultivated land label; Based on the preliminary image feature dataset, a normalized vegetation index of each preset time interval of each piece of cultivated land in the target area is obtained, and the normalized vegetation index of all the preset time intervals of each piece of cultivated land is formed into time series image data of the corresponding cultivated land, and all the time series image data form a time series image dataset; Extracting texture features of each preliminary image feature data in the preliminary image feature data set, and adding the texture features in the form of new bands to the corresponding preliminary image feature data in the preliminary image feature data set to obtain a texture spectral feature data set; A teacher model is obtained by training a model based on the texture spectral feature dataset and the time series image dataset, a student model is obtained by training a preset to-be-trained model based on the teacher model, and an idle farmland extraction model in a complex cultivated area is formed with the teacher model as a decision maker and all the student models as input ends; Wherein, the preset model to be trained includes a preset neural network model and a long short-term memory network model, and the data corresponding to the cultivated land label of non-idle cultivated land and the data corresponding to the cultivated land label of non-grain idle cultivated land in the texture spectral feature data set are used as the training data set of the preset neural network model, and the data corresponding to the cultivated land label of year-round idle cultivated land in the time series image data set are used as the training data set of the long short-term memory network model; the preset model to be trained also includes a decision tree model, and the data corresponding to the cultivated land label of seasonal idle cultivated land in the texture spectral feature data set and the time series image data set are used as the training data set of the decision tree model; Training a preset model to be trained based on the teacher model to obtain a student model includes: Using feature-based knowledge distillation, the features acquired by the teacher model are aligned with the features of the preset neural network model in terms of feature dimensions and are transferred to the middle layer of the preset neural network model to obtain a neural network student model; Using feature-based knowledge distillation, the features acquired by the teacher model are aligned with the features of the long short-term memory network model in terms of feature dimensions and are transferred to the middle layer of the long short-term memory network model to obtain a long short-term memory network student model; The soft labels obtained by the Softmax function of the teacher model are used to assist in training the decision tree model using target-based Logits distillation to obtain a decision tree student model.

2. The model training method according to claim 1, characterized in that Obtaining a preliminary image feature dataset of the target area within a preset time period includes: Acquire an original image dataset of a target area within a preset time period, pre-process the original image data in the original image dataset to obtain a preliminary image dataset, crop the preliminary image data in the preliminary image dataset based on the cultivated land range data to obtain a cultivated land range image dataset, and perform cultivated land labeling on the cultivated land corresponding to the cultivated land range image data in the cultivated land range image dataset based on an idle labeling criterion to obtain a preliminary image feature dataset; Wherein, the preset time period is greater than two years.

3. The model training method according to claim 2, characterized in that The original image data is satellite remote sensing image data, and atmospheric correction is performed on the satellite remote sensing image data to achieve preprocessing of the image data.

4. The model training method according to claim 2, characterized in that The idle land marking criteria include idle farmland types and non-idle farmland types, and the idle farmland types include seasonal idle farmland, annual idle farmland, perennial idle farmland and non-grain idle farmland.

5. A model training device for extracting idle farmland in complex farming areas, characterized in that: It includes a preliminary image feature dataset acquisition module, a time series image dataset acquisition module, a texture spectrum feature dataset acquisition module and a model training module; The preliminary image feature data set acquisition module is used to acquire a preliminary image feature data set of a target area within a preset time period, wherein the preliminary image feature data set includes a preliminary image feature data subset of each piece of cultivated land in the target area, and the preliminary image feature data subset includes a preliminary image feature data group and a corresponding cultivated land label; The time series image dataset acquisition module is used to obtain the normalized vegetation index of each preset time interval of each piece of cultivated land in the target area based on the preliminary image feature dataset, and form the normalized vegetation index of all the preset time intervals of each piece of cultivated land into the time series image data of the corresponding cultivated land, and all the time series image data form a time series image dataset; The texture spectral feature data set acquisition module is used to extract the texture features of each preliminary image feature data in the preliminary image feature data set, and add the texture features in the form of new bands to the corresponding preliminary image feature data in the preliminary image feature data set to obtain a texture spectral feature data set; The model training module is used to train the model based on the texture spectral feature dataset and the time series image dataset to obtain a teacher model, train the preset to-be-trained model based on the teacher model to obtain a student model, and use the teacher model as a decision maker and all the student models as input ends to form an idle farmland extraction model in complex cultivated areas; Wherein, the preset model to be trained includes a preset neural network model and a long short-term memory network model, and the data corresponding to the cultivated land label of non-idle cultivated land and the data corresponding to the cultivated land label of non-grain idle cultivated land in the texture spectral feature data set are used as the training data set of the preset neural network model, and the data corresponding to the cultivated land label of year-round idle cultivated land in the time series image data set are used as the training data set of the long short-term memory network model; the preset model to be trained also includes a decision tree model, and the data corresponding to the cultivated land label of seasonal idle cultivated land in the texture spectral feature data set and the time series image data set are used as the training data set of the decision tree model; Training a preset model to be trained based on the teacher model to obtain a student model includes: Using feature-based knowledge distillation, the features acquired by the teacher model are aligned with the features of the preset neural network model in terms of feature dimensions and are transferred to the middle layer of the preset neural network model to obtain a neural network student model; Using feature-based knowledge distillation, the features acquired by the teacher model are aligned with the features of the long short-term memory network model in terms of feature dimensions and are transferred to the middle layer of the long short-term memory network model to obtain a long short-term memory network student model; The soft labels obtained by the Softmax function of the teacher model are used to assist in training the decision tree model using target-based Logits distillation to obtain a decision tree student model.

6. A method for extracting idle farmland in complex farming areas, characterized in that: include: Acquire a target image feature data set of a to-be-extracted area within a target time period, wherein the target image feature data set includes a target image feature data subset of each piece of cultivated land within the to-be-extracted area, and the target image feature data subset includes a target image feature data group; Based on the target image feature dataset, a normalized vegetation index of each preset time interval of each piece of cultivated land in the area to be extracted is obtained, and the normalized vegetation index of all the target time zones of each piece of cultivated land is used to form target time series image data of the corresponding cultivated land, and all the target time series image data form a target time series image dataset; Extracting texture features of each target image feature data in the target image feature data set, and adding the texture features in the form of new bands to the corresponding target image feature data in the target image feature data set to obtain a target texture spectral feature data set; The target texture spectral feature dataset, the target time series image dataset, the target texture spectral feature data and the target time series image data are respectively transmitted to the corresponding student models in the idle farmland extraction model in the complex farming area to obtain the initial distribution data of non-idle farmland and non-grain idle farmland, the initial distribution data of annual idle farmland and perennial farmland, and the initial distribution data of seasonal idle farmland, and the target texture spectral feature dataset, the target time series image dataset, the initial distribution data of non-idle farmland and non-grain idle farmland, the initial distribution data of annual idle farmland and perennial farmland, and the initial distribution data of seasonal idle farmland are all input into the teacher model in the idle farmland extraction model in the complex farming area to obtain the distribution of farmland in the area to be extracted; Wherein, the idle farmland extraction model in complex farming areas is obtained based on the idle farmland extraction model training method in complex farming areas according to any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the idle farmland extraction model training method for complex farmland areas according to any one of claims 1 to 4 or the idle farmland extraction method for complex farmland areas according to claim 6.

8. A terminal comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the idle farmland extraction model training method for complex farmland areas according to any one of claims 1 to 4 or the idle farmland extraction method for complex farmland areas according to claim 6.

Images