Farmland remote sensing image analysis method, device, equipment, storage medium and product

Abstract

The application provides a farmland remote sensing image analysis method, device, equipment, storage medium and product, and relates to the technical field of image processing. The method comprises the following steps: acquiring a historical remote sensing image set and a current remote sensing image of a target farmland area; determining N target historical remote sensing images in the historical remote sensing image set based on an analysis result of data distribution differences between the current remote sensing image and the historical remote sensing image set, wherein each pixel in the target historical remote sensing image corresponds to a farmland type label, and N is a positive integer; migrating the farmland type label in the target historical remote sensing image to the current remote sensing image in a multi-source transfer learning manner to obtain a current remote sensing target image; and training a preset deep forest model based on the N target historical remote sensing images and the current remote sensing target image to obtain a trained farmland remote sensing image classification model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to a cultivated land remote sensing image analysis method, device, equipment, storage medium and product. BACKGROUND

[0002] Maintaining the quantity and quality of cultivated land is very important for food production, and in many places, there are different degrees of cultivated land abandonment, which will lead to problems such as waste of land resources and decline in cultivated land quality.

[0003] In related technologies, it is usually necessary to manually analyze the images of cultivated areas to determine abandoned cultivated land, which is usually slow and has a certain misjudgment rate.

[0004] Therefore, how to effectively detect abandoned cultivated land has become a problem to be solved in the industry. SUMMARY

[0005] The present application provides a cultivated land remote sensing image analysis method, device, equipment, storage medium and product to solve the problem of how to effectively detect abandoned cultivated land in the prior art.

[0006] The present application provides a cultivated land remote sensing image analysis method, comprising:

[0007] Obtaining a set of historical remote sensing images and a current remote sensing image of a target cultivated area;

[0008] Based on the data distribution difference analysis result of the current remote sensing image and the set of historical remote sensing images, determining N target historical remote sensing images in the set of historical remote sensing images, wherein each pixel in the target historical remote sensing image corresponds to a cultivated land type label, and N is a positive integer;

[0009] Transferring the cultivated land type label in the target historical remote sensing image to the current remote sensing image by multi-source transfer learning to obtain a current remote sensing target image;

[0010] Training a preset deep forest model based on N target historical remote sensing images and the current remote sensing target image to obtain a trained cultivated land remote sensing image classification model;

[0011] The cultivated land remote sensing image classification model is used to classify the cultivated land type of each pixel in the remote sensing image of the target cultivated area.

[0012] According to the cultivated land remote sensing image analysis method provided by the present application, after obtaining the trained cultivated land remote sensing image classification model, the method further comprises:

[0013] input the current remote sensing image into the trained cultivated land remote sensing image classification model, and output the cultivated land type of each pixel in the current remote sensing image;

[0014] obtain a target pixel in which the cultivated land type is a bare land type, and determine a deserted land block in a farmland corresponding to the target pixel based on the imaging time of the current remote sensing image and crop rotation information of the farmland corresponding to the target pixel.

[0015] According to the cultivated land remote sensing image analysis method provided by the application, based on the data distribution difference analysis result of the current remote sensing image and the historical remote sensing image set, N target historical remote sensing images in the historical remote sensing image set are determined, including:

[0016] Calculate the maximum mean difference value information of the current remote sensing image and each historical remote sensing image in the historical remote sensing image set;

[0017] Determine a historical remote sensing image subset in which the sum of the weights of the maximum mean difference value information is greater than a first preset threshold value, and the sum of the image quantities is less than a second preset threshold value, wherein the historical remote sensing image subset includes N target historical remote sensing images.

[0018] According to the cultivated land remote sensing image analysis method provided by the application, based on N target historical remote sensing images and the current remote sensing target image, a preset deep forest model is trained to obtain a trained cultivated land remote sensing image classification model, including:

[0019] Take one of the target historical remote sensing images or the current remote sensing target image as a training sample, and obtain a plurality of training samples;

[0020] The plurality of training samples are used to train the preset deep forest model, and the training is stopped under the condition that a preset training condition is met, and a trained cultivated land remote sensing image classification model is obtained.

[0021] According to the cultivated land remote sensing image analysis method provided by the application, the cultivated land type label includes: a water body type label, an artificial object type label, a forest land type label, a crop type label and a bare land type label.

[0022] The application also provides a cultivated land remote sensing image analysis device, comprising:

[0023] The acquisition module is configured to acquire a historical remote sensing image set and a current remote sensing image of a target cultivated land region;

[0024] determining module, configured to determine N target historical remote sensing images in the set of historical remote sensing images based on the data distribution difference analysis result of the current remote sensing image and the set of historical remote sensing images, wherein each pixel in the target historical remote sensing image corresponds to a cultivated land type label, and N is a positive integer;

[0025] migrating module, configured to migrate the cultivated land type label in the target historical remote sensing image to the current remote sensing image in a manner of multi-source migration learning to obtain a current remote sensing target image;

[0026] training module, configured to train a preset deep forest model based on the N target historical remote sensing images and the current remote sensing target image to obtain a trained cultivated land remote sensing image classification model;

[0027] The cultivated land remote sensing image classification model is used for classifying the cultivated land types of each pixel in the target cultivated land region.

[0028] According to the cultivated land remote sensing image analysis device provided by the application, the device is further used for:

[0029] inputting the current remote sensing image into the trained cultivated land remote sensing image classification model to output the cultivated land types of each pixel in the current remote sensing image;

[0030] obtaining a target pixel in which the cultivated land type is a bare land type, and determining a deserted land block in a farmland corresponding to the target pixel based on the imaging time of the current remote sensing image and crop rotation information of the farmland corresponding to the target pixel.

[0031] According to the cultivated land remote sensing image analysis device provided by the application, the device is further used for:

[0032] calculating maximum mean difference value information of the current remote sensing image and each historical remote sensing image in the set of historical remote sensing images;

[0033] determining a subset of historical remote sensing images in which the sum of weights of maximum mean difference value information is greater than a first preset threshold and the sum of image quantities is less than a second preset threshold, wherein the subset of historical remote sensing images includes N target historical remote sensing images.

[0034] According to the cultivated land remote sensing image analysis device provided by the application, the device is further used for:

[0035] taking one of the target historical remote sensing images or the current remote sensing target image as a training sample to obtain a plurality of training samples;

[0036] The preset deep forest model is trained by using a plurality of training samples, and the training is stopped when a preset training condition is met, so as to obtain a trained cultivated land remote sensing image classification model.

[0037] According to the cultivated land remote sensing image analysis device provided by the application, the cultivated land type label comprises a water body type label, a man-made object type label, a forest land type label, a crop type label and a bare land type label.

[0038] The application further provides an electronic device, including a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the cultivated land remote sensing image analysis method according to any one of the above when executing the program.

[0039] The application further provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program is executable on a processor to implement the cultivated land remote sensing image analysis method according to any one of the above.

[0040] The application further provides a computer program product, which includes a computer program, and the computer program is executable on a processor to implement the cultivated land remote sensing image analysis method according to any one of the above.

[0041] The cultivated land remote sensing image analysis method, device, equipment, storage medium and product provided by the application filter target historical remote sensing images with smaller feature distribution differences from historical remote sensing images for subsequent model training through current remote sensing images, further migrate cultivated land type labels in the target historical remote sensing images to the current remote sensing images through a multi-source transfer learning method, and assign annotation samples to new images, so that manual training sample collection on new images is not needed, model training and ground object classification are automated, and finally, a preset deep forest model is trained according to N target historical remote sensing images and a current remote sensing target image, so as to obtain a cultivated land remote sensing image classification model for classifying cultivated land types of each pixel in the target cultivated land remote sensing image, thereby automatically and efficiently identifying abandoned cultivated land. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0043] Figure 1 The cultivated land remote sensing image analysis method flowchart provided by the application is shown in the figure.

[0044] Figure 2A main flowchart of cultivated land remote sensing image analysis provided by the embodiment of the present application is shown in the figure.

[0045] Figure 3 A flowchart of abandoned land identification in the embodiment of the present application is shown in the figure.

[0046] Figure 4 A structure schematic diagram of the cultivated land remote sensing image analysis device described in the embodiment of the present application is shown in the figure.

[0047] Figure 5 A structure schematic diagram of the electronic device provided by the present application is shown in the figure. DETAILED DESCRIPTION

[0048] In order to make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below with reference to the drawings in the present application. Obviously, the described embodiments are some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0049] Figure 1 A flowchart of the cultivated land remote sensing image analysis method provided by the embodiment of the present application is shown in the figure, as shown in the figure, comprising: Figure 1

[0050] Step 110, obtaining a historical remote sensing image set and a current remote sensing image of a target cultivated land region.

[0051] Specifically, the target cultivated land region described in the embodiment of the present application can be specifically a cultivated land which needs to be analyzed for abandoned land condition, and the target cultivated land region can be set by the user as needed.

[0052] The historical remote sensing image set described in the embodiment of the present application can contain multiple historical remote sensing images of the target cultivated land region, and the historical remote sensing images can be obtained from the historical data already stored.

[0053] The current remote sensing image in the embodiment of the present application refers to the remote sensing image of the target cultivated land region obtained recently.

[0054] More specifically, since the resolution of the images obtained by different sensors will be different, after obtaining the initial historical remote sensing image set and the current remote sensing image, spatial consistency processing is performed thereon, the spatial resolution of the data is unified with that in the historical image library, the ground object discrimination error caused by inconsistent resolution is reduced, and finally the historical remote sensing image set and the current remote sensing image are obtained.

[0055] ​In step 120, based on the data distribution difference analysis result of the current remote sensing image and the historical remote sensing image set, N target historical remote sensing images in the historical remote sensing image set are determined, wherein each pixel in the target historical remote sensing image corresponds to a cultivated land type label, and N is a positive integer.

[0056] More specifically, since the same cultivated land area accumulates remote sensing images of different periods and different sensor types, a historical remote sensing image library of the area is formed, and the feature distribution of the same type of ground object on different images will have certain differences. Therefore, in the embodiment of the application, the target historical remote sensing image that matches the current remote sensing image can be further selected from the historical remote sensing image set as a training sample.

[0057] More specifically, the data distribution difference analysis result of the current remote sensing image and the historical remote sensing image set in the embodiment of the application can be the result of maximum mean difference analysis of the current remote sensing image and each historical remote sensing image.

[0058] Through the data distribution difference analysis result, the target historical remote sensing image with low distribution difference from the current remote sensing image in the historical remote sensing image set can be determined.

[0059] More specifically, the target historical remote sensing image in the embodiment of the application includes remote sensing images of multiple plots, and each remote sensing image of a plot is composed of at least one pixel. In the embodiment of the application, each remote sensing image of a plot can be labeled with a cultivated land type label, that is, each corresponding pixel also carries a cultivated land type label.

[0060] The cultivated land type label described in the embodiment of the application is used to mark the actual cultivated land type of each plot.

[0061] In step 130, the cultivated land type label in the target historical remote sensing image is migrated into the current remote sensing image by a multi-source transfer learning method, to obtain a current remote sensing target image.

[0062] Specifically, the transfer knowledge extracted by the multi-source transfer learning method is no longer limited to a single source domain data set, but comes from two or more source domain data sets, and can fully utilize the information in the existing multi-source data. The transfer learning technology can effectively transfer the cultivated land type label in the processed target historical remote sensing image to the current remote sensing image to a certain extent, and realize automatic labeling of the current remote sensing image to a certain extent.

[0063] In the embodiment of the present application, the information-rich part of the current remote sensing image can obtain the cultivated land type label through transfer learning, that is, after transfer learning, part of the pixels in the current remote sensing target image already carry the cultivated land type label. However, there are still many pixels in the current remote sensing target image that do not carry the cultivated land type label and cannot determine the corresponding cultivated land type, which needs to be further analyzed by the subsequent model.

[0064] The transfer component analysis (TCA) algorithm in the transfer learning method is used to perform transfer learning on the current remote sensing image according to each historical remote sensing image.

[0065] In step 140, the preset deep forest model is trained based on the N target historical remote sensing images and the current remote sensing target image to obtain a trained cultivated land remote sensing image classification model.

[0066] The cultivated land remote sensing image classification model is used to classify the cultivated land types of each pixel in the remote sensing image of the target cultivated land area.

[0067] Specifically, in the embodiment of the present application, the deep forest model has the characteristics of strong representation learning ability, fewer hyperparameters, strong model structure interpretability, and small computational overhead.

[0068] After multi-source transfer learning, the preset deep forest model is trained based on the N target historical remote sensing images and the current remote sensing target image. When the preset training condition is met, the training is stopped, and a trained cultivated land remote sensing image classification model is obtained. The model can classify the cultivated land types of each pixel in the remote sensing image of the target cultivated land area, and effectively identify the abandoned land in the target cultivated land area.

[0069] In the embodiment of the present application, the target historical remote sensing image with small feature distribution difference is selected from the historical remote sensing image set through the current remote sensing image for subsequent model training. Further, through the multi-source transfer learning method, the cultivated land type label in the target historical remote sensing image is transferred to the current remote sensing image to provide labeled samples for the new image, so that manual training samples do not need to be collected on the new image, realizing the automation of model training and feature classification. Finally, the preset deep forest model is trained based on the N target historical remote sensing images and the current remote sensing target image to obtain a cultivated land remote sensing image classification model for classifying the cultivated land types of each pixel in the remote sensing image of the target cultivated land area, thereby automatically and efficiently identifying the abandoned cultivated land.

[0070] Optionally, after obtaining the trained cultivated land remote sensing image classification model, the following steps are further included:

[0071] input the current remote sensing image into the trained cultivated land remote sensing image classification model, and output the cultivated land type of each pixel in the current remote sensing image;

[0072] obtain a target pixel in which the cultivated land type is a bare land type, and determine a deserted land block in a block corresponding to the target pixel based on an imaging time of the current remote sensing image and crop rotation information of the block corresponding to the target pixel.

[0073] Specifically, after obtaining the cultivated land type of each pixel in the current remote sensing image in the embodiment of the application, a weighted voting of multiple classification results is performed on each pixel, and the highest weight class of the pixel is given, so as to obtain the cultivated land class of each block.

[0074] However, since cultivation is greatly affected by climate and cultivation period, some crops planted in a block are only planted in the planting period, and other periods may belong to the bare land type in the feedback of the remote sensing image, but this does not mean that the block is really bare land.

[0075] Therefore, in the embodiment of the application, it is necessary to further combine the imaging time of the current remote sensing image and the crop rotation information of the block corresponding to the target pixel for judgment.

[0076] First, based on the block crop rotation information and the imaging time of the current remote sensing image, it is determined whether the crop of the block belongs to the planting period. The block crop rotation information can be determined according to the type of the crop that should be planted in the block in advance.

[0077] Further, if the cultivated land type of the pixel is a bare land type, it is necessary to further analyze whether the block corresponding to the pixel is in the planting period. If the block of the pixel is in the planting period, it is indicated that the block may be a deserted land block, and if the block of the pixel is not in the planting period, it is indicated that it is likely not a deserted land block.

[0078] More specifically, if a deserted land block is identified, the information is reported to a human for further confirmation of whether the block is a deserted land block.

[0079] In the embodiment of the application, the imaging time of the current remote sensing image and the crop rotation information of the block corresponding to the target pixel can be used to further determine the block of the bare land type output by the model, so as to effectively ensure the recognition accuracy of the deserted land block.

[0080] Optionally, based on the data distribution difference analysis result of the current remote sensing image and the historical remote sensing image set, N target historical remote sensing images in the historical remote sensing image set are determined, including:

[0081] The maximum mean difference value information of the current remote sensing image and each historical remote sensing image in the historical remote sensing image set is calculated.

[0082] determine a historical remote sensing image subset in which the sum of weights of maximum mean discrepancy value information of the historical remote sensing image set is greater than a first preset threshold value, and the sum of image quantities is less than a second preset threshold value, wherein the historical remote sensing image subset includes N target historical remote sensing images.

[0083] Specifically, in the present application, the maximum mean discrepancy (MMD) is used to measure the distribution difference between different images. MMD is one of the most commonly used difference measurement indicators in transfer learning research. It first maps the source domain data and the target domain data to the reproducing kernel hilbert space (RKHS), and then calculates the distance of the mean values of the two groups of data. The calculation formula of MMD is:

[0084]

[0085] In the formula, VMMD represents the value of MMD, XS and XT represent the source domain data and the target domain data respectively; nS and nT represent the number of source domain data and target domain data respectively; and ψ(·) represents the kernel function. The larger the value of MMD is, the greater the distribution difference between the two groups of data is. When the distributions of the two groups of data are completely consistent, the value of MMD is equal to 0.

[0086] First, the MMD of all images in the historical remote sensing image set and the current remote sensing image is calculated respectively, and then all MMDs are normalized. A group of historical images with a large enough sum of weight values and as few images as possible are selected as target historical remote sensing images through adaptive dynamic selection, so as to obtain a historical remote sensing image subset.

[0087] In the embodiment of the present application, the maximum mean discrepancy value information of the current remote sensing image and each historical remote sensing image in the historical remote sensing image set is calculated, so that the historical remote sensing image with the smallest possible difference from the current remote sensing image is effectively selected from the historical remote sensing image set as a training sample, which can effectively ensure that the subsequent trained model can better identify the current cultivated land type of the target cultivated land region.

[0088] Optionally, a preset deep forest model is trained based on the N target historical remote sensing images and the current remote sensing target image to obtain a trained cultivated land remote sensing image classification model, which includes:

[0089] One of the target historical remote sensing images or the current remote sensing target image is taken as a training sample, and a plurality of training samples are obtained.

[0090] The preset deep forest model is trained by using a plurality of training samples, and the training is stopped when a preset training condition is met, so as to obtain a trained cultivated land remote sensing image classification model.

[0091] Specifically, after multi-source transfer learning is performed, the model is trained based on the transferred target historical remote sensing image, and a trained cultivated land remote sensing image classification model is obtained.

[0092] The deep forest is used as the cultivated land remote sensing image classification model.

[0093] The deep forest has the characteristics of strong representation learning ability, fewer hyperparameters, strong model structure interpretability and small calculation overhead, although the deep neural network has strong performance, but it has the defects of relying on a large number of training samples, complex model and poor interpretability, more hyperparameters and high difficulty in parameter adjustment. If the powerful representation learning ability of the deep neural network is introduced into a suitable learning model, the performance of the model may be sufficient to compare the deep neural network, while the above-mentioned defects can be avoided, and therefore the deep forest (Deep Forest, DF) is proposed. The DF is a model of decision tree integration, mainly including two parts of multi-granularity scanning and cascade forest, which will be simply introduced below.

[0094] The multi-granularity scanning is used for enhancing the representation learning ability of the model. The multi-granularity scanning process of the sequence data is shown in the above figure, assuming that the original feature dimension of the input sample is equal to 400, and the category number is equal to 3. When the size of the sampling window is 100 dimensions and the sliding distance is 1, sliding sampling can be performed on the sample to obtain 301 sub-samples, and the feature dimension of each sub-sample is equal to 100. A sub-sample is input into a random forest to obtain the prediction probability of 3 categories, therefore 301 sub-samples are input into 2 random forests, and a total of 301*3*2=1806 category prediction probabilities can be obtained, which are spliced together as the multi-granularity features of the sample. In addition, a plurality of sampling windows with different sizes are used for multi-granularity scanning, which can further increase the dimension of the multi-granularity features. Since each sub-sample represents the local feature of the original sample, the multi-granularity scanning is equivalent to the structured up-sampling of the original feature, which enhances the representation learning ability of the model and helps to improve the performance of the model.

[0095] The representation learning ability of deep neural networks mainly comes from the layer-by-layer processing of original features. Inspired by this, the cascade forest is also composed of cascade layers that progress layer by layer, as shown in the above figure. In order to enhance the difference, two completely random forests and two random forests (hereinafter collectively referred to as random forests) are used in each cascade layer. The class prediction probability output by the random forest is taken as the enhanced feature of the sample. When the number of categories is 3, 4 random forests can output 12 class prediction probabilities as the enhanced features of the sample. The enhanced features and the multi-granularity features are spliced together as the input features of the next cascade layer.

[0096] In each cascade layer of the cascade forest, the training samples are divided into a growing set and a validation set. The growing set is used to train the model, and the validation set is used to verify the performance of the model. When the prediction accuracy of the model on the validation set no longer improves significantly, the generation of new cascade layers is stopped. This method not only avoids model overfitting, but also adaptively adjusts the complexity of the model, so that the model can adapt to training data of different scales. When the cascade forest stops growing, the class prediction probability of the sample output by the four random forests in the last cascade layer is averaged, and the class with the highest prediction probability is taken as the final prediction result of the sample.

[0097] The preset training condition described in the embodiments of the present application can be, for example, meeting a preset training number, such as training for 150 times. The preset training condition can also be, for example, meeting a preset training event, such as training for 30 minutes.

[0098] In the embodiments of the present application, the preset deep forest model can be trained by a plurality of training samples, and finally a model for classifying the types of each pixel in the remote sensing image of the target cultivated area is obtained, which can effectively help to identify abandoned land.

[0099] The cultivated land remote sensing image analysis device provided by the present application is described below. The cultivated land remote sensing image analysis device described below can be correspondingly referred to the cultivated land remote sensing image analysis method described above.

[0100] Optionally, Figure 2 The main flowchart of the cultivated land remote sensing image analysis provided by the embodiments of the present application is shown in Figure 2 As shown in the figure, it includes:

[0101] Since the resolution of images obtained by different sensors will differ, first, the input image is processed for spatial consistency, the spatial resolution of the data is unified with the data in the historical remote sensing image data set, and the error in ground object discrimination caused by inconsistent resolution is reduced.

[0102] The feature distribution of the same type of land cover can differ across different images. In transfer learning, these differences are referred to as inter-domain distribution differences. These inter-domain distribution differences affect the effectiveness of transfer learning; the smaller the differences between two domains, the easier it is to transfer knowledge between them.

[0103] Since remote sensing images from different periods and with different sensor types have been accumulated in the same area, a historical source domain remote sensing image library for the region has been formed. It is necessary to select a set of the most suitable source domain images from the historical source domain remote sensing image library based on the differences in data distribution to obtain a low-difference historical remote sensing image set.

[0104] Multi-source transfer learning is performed based on a low-difference historical remote sensing image set and the input remote sensing image. Then, the model is trained using the current remote sensing target image after transfer and the low-difference historical remote sensing image set. The deep forest model is used in the model training process. After the model training is completed, a trained farmland remote sensing image classification model is obtained. This classification model can be used to further analyze each pixel in the input remote sensing image. Combined with farmland patches, unplanted areas within the farmland range can be extracted. Then, combined with the crop phenology information of the region, abandoned farmland can be extracted.

[0105] Figure 3 This is a schematic diagram of the land abandonment identification process in an embodiment of this application, such as... Figure 3 As shown, the current plot is determined to be in the planting period based on the image imaging time and crop rotation information of the plot. If the plot is in the planting period, the plot category information is combined to determine whether there is crop planting. If there is no crop planting, it is judged to be suspected of being abandoned.

[0106] Figure 4 This is a schematic diagram of the structure of the farmland remote sensing image analysis device described in the embodiments of this application, such as... Figure 4 As shown, it includes: an acquisition module 410, a determination module 420, a transfer module 430, and a training module 440;

[0107] Among them, the acquisition module 410 is used to acquire the historical remote sensing image set and the current remote sensing image of the target cultivated land area;

[0108] The determining module 420 is used to determine N target historical remote sensing images in the historical remote sensing image set based on the data distribution difference analysis results between the current remote sensing image and the historical remote sensing image set. Each pixel in the target historical remote sensing image corresponds to a farmland type label, and N is a positive integer.

[0109] The migration module 430 is used to transfer the farmland type label in the target historical remote sensing image to the current remote sensing image through multi-source transfer learning to obtain the current remote sensing target image.

[0110] The training module 440 is configured to train a preset deep forest model based on the N target historical remote sensing images and the current remote sensing target image, and obtain a trained cultivated land remote sensing image classification model.

[0111] The cultivated land remote sensing image classification model is configured to classify the cultivated land types of each pixel in the remote sensing image of the target cultivated land region.

[0112] Optionally, the apparatus is further configured to:

[0113] input the current remote sensing image into the trained cultivated land remote sensing image classification model, and output the cultivated land types of each pixel in the current remote sensing image;

[0114] obtain a target pixel in which the cultivated land type is a bare land type, and determine a deserted land block in a farmland corresponding to the target pixel based on an imaging time of the current remote sensing image and crop rotation information of the farmland corresponding to the target pixel.

[0115] Optionally, the apparatus is further configured to:

[0116] calculate maximum mean difference value information of the current remote sensing image and each historical remote sensing image in the historical remote sensing image set;

[0117] determine a historical remote sensing image subset in which a sum of weights of maximum mean difference value information is greater than a first preset threshold and a sum of image quantities is less than a second preset threshold, and the historical remote sensing image subset includes N target historical remote sensing images.

[0118] Optionally, the apparatus is further configured to:

[0119] take one of the target historical remote sensing images or the current remote sensing target image as one training sample, and obtain a plurality of training samples;

[0120] train the preset deep forest model by using the plurality of training samples, and stop training under a preset training condition to obtain the trained cultivated land remote sensing image classification model.

[0121] Optionally, the cultivated land type label includes a water body type label, an artificial object type label, a forest land type label, a crop type label, and a bare land type label.

[0122] In the embodiment of the present application, the target historical remote sensing image with smaller feature distribution difference is selected from the historical remote sensing image set through the current remote sensing image for subsequent model training, and further through the multi-source transfer learning mode, the cultivated land type label in the target historical remote sensing image is transferred to the current remote sensing image, and the labeled sample is given to the new image, so that the model training and the automatic of ground object classification are realized without manual collection of training samples on the new image. Finally, the preset deep forest model is trained according to N target historical remote sensing images and the current remote sensing target image, and a cultivated land remote sensing image classification model for classifying the cultivated land type of each pixel in the remote sensing image of the target cultivated land region is obtained, so as to automatically and efficiently identify abandoned cultivated land.

[0123] Figure 5 is a structural schematic diagram of an electronic device provided by the present application, as Figure 5 shown, the electronic device can include: a processor 510, a communications interface 520, a memory 530 and a communications bus 540, wherein the processor 510, the communications interface 520, the memory 530 complete mutual communication through the communications bus 540. The processor 510 can call the logic instruction in the memory 530 to execute the cultivated land remote sensing image analysis method, the method comprising: acquiring a historical remote sensing image set and a current remote sensing image of a target cultivated land region;

[0124] Based on the data distribution difference analysis result of the current remote sensing image and the historical remote sensing image set, N target historical remote sensing images in the historical remote sensing image set are determined, wherein each pixel in the target historical remote sensing image corresponds to a cultivated land type label, and N is a positive integer;

[0125] The cultivated land type label in the target historical remote sensing image is transferred to the current remote sensing image through multi-source transfer learning, and a current remote sensing target image is obtained;

[0126] Based on N target historical remote sensing images and the current remote sensing target image, a preset deep forest model is trained to obtain a trained cultivated land remote sensing image classification model;

[0127] The cultivated land remote sensing image classification model is used for classifying the cultivated land type of each pixel in the remote sensing image of the target cultivated land region.

[0128] Further, the logic instructions in the memory 530 described above can be implemented in the form of software functional units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that make contributions to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0129] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the cultivated land remote sensing image analysis method provided by the above-mentioned methods, the method comprising: obtaining a historical remote sensing image set and a current remote sensing image of a target cultivated land region;

[0130] Based on the data distribution difference analysis result of the current remote sensing image and the historical remote sensing image set, determining N target historical remote sensing images in the historical remote sensing image set, wherein each pixel in the target historical remote sensing image corresponds to a cultivated land type label, and N is a positive integer;

[0131] Migrating the cultivated land type label in the target historical remote sensing image to the current remote sensing image by a multi-source transfer learning manner to obtain a current remote sensing target image;

[0132] Training a preset deep forest model based on the N target historical remote sensing images and the current remote sensing target image to obtain a trained cultivated land remote sensing image classification model;

[0133] The cultivated land remote sensing image classification model is used for classifying the cultivated land types of each pixel in the remote sensing image of the target cultivated land region.

[0134] In another aspect, the present application also provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the cultivated land remote sensing image analysis method provided by the above-mentioned methods, the method comprising: obtaining a historical remote sensing image set and a current remote sensing image of a target cultivated land region;

[0135] Determine N target historical remote sensing images in the set of historical remote sensing images based on the data distribution difference analysis result of the current remote sensing image and the set of historical remote sensing images, wherein each pixel in the target historical remote sensing image corresponds to a cultivated land type label, and N is a positive integer;

[0136] Migrate the cultivated land type label in the target historical remote sensing image to the current remote sensing image in a multi-source transfer learning manner to obtain a current remote sensing target image.

[0137] Train a preset deep forest model based on the N target historical remote sensing images and the current remote sensing target image to obtain a trained cultivated land remote sensing image classification model.

[0138] The cultivated land remote sensing image classification model is used to classify the cultivated land types of each pixel in the remote sensing image of the target cultivated land region.

[0139] The device embodiments described above are only schematic, wherein the units shown as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., they can be located in one place, or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0140] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary general hardware platforms, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiment.

[0141] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method of analyzing a remote sensing image of a cultivated field, characterized in that, The method comprises the following steps: obtain a set of historical remote sensing images and a current remote sensing image of a target cultivated area; determine N target historical remote sensing images in the set of historical remote sensing images based on a data distribution difference analysis result of the current remote sensing image and the set of historical remote sensing images, wherein each pixel in the target historical remote sensing images corresponds to a cultivated land type label, and N is a positive integer; migrate the cultivated land type label in the target historical remote sensing images to the current remote sensing image by multi-source transfer learning to obtain a current remote sensing target image; train a preset deep forest model based on the N target historical remote sensing images and the current remote sensing target image to obtain a trained cultivated land remote sensing image classification model; wherein the cultivated land remote sensing image classification model is used to classify the cultivated land type of each pixel in the remote sensing image of the target cultivated area; wherein after obtaining the trained cultivated land remote sensing image classification model, the method further comprises: inputting the current remote sensing image into the trained cultivated land remote sensing image classification model to output the cultivated land type of each pixel in the current remote sensing image; obtaining target pixels in which the cultivated land type is bare land type, and determining a abandoned land block in a land block corresponding to the target pixels based on the imaging time of the current remote sensing image and the crop rotation information of the land block corresponding to the target pixels; wherein based on the data distribution difference analysis result of the current remote sensing image and the set of historical remote sensing images, the N target historical remote sensing images in the set of historical remote sensing images are determined, comprising: calculating the maximum mean difference value information of the current remote sensing image and each historical remote sensing image in the set of historical remote sensing images; determining a subset of historical remote sensing images in which the sum of the weights of the maximum mean difference value information is greater than a first preset threshold and the sum of the image quantities is less than a second preset threshold, wherein the subset of historical remote sensing images includes N target historical remote sensing images.

2. The cultivated field remote sensing image analysis method according to claim 1, characterized in that, Training a preset deep forest model based on N target historical remote sensing images and a current remote sensing target image to obtain a trained cultivated land remote sensing image classification model comprises: taking one of the target historical remote sensing images or the current remote sensing target image as a training sample to obtain a plurality of training samples; training the preset deep forest model using the plurality of training samples, and stopping training when a preset training condition is met to obtain a trained cultivated land remote sensing image classification model.

3. The cultivated land remote sensing image analysis method according to claim 1, characterized by, The cultivated land type label includes a water body type label, a man-made object type label, a forest land type label, a crop type label, and a bare land type label.

4. An agricultural field remote sensing image analysis device, characterized by comprising: The method comprises the following steps: an obtaining module for obtaining a set of historical remote sensing images and a current remote sensing image of a target cultivated area; a determining module for determining N target historical remote sensing images in the set of historical remote sensing images based on a data distribution difference analysis result of the current remote sensing image and the set of historical remote sensing images, wherein each pixel in the target historical remote sensing images corresponds to a cultivated land type label, and N is a positive integer; a migration module, configured to migrate a cultivated land type label in the target historical remote sensing image to the current remote sensing image by multi-source transfer learning, to obtain a current remote sensing target image; a training module, configured to train a preset deep forest model based on N target historical remote sensing images and the current remote sensing target image, to obtain a trained cultivated land remote sensing image classification model; The cultivated land remote sensing image classification model is configured to classify the cultivated land type of each pixel in the remote sensing image of the target cultivated land region. The device is further configured to: input the current remote sensing image into the trained cultivated land remote sensing image classification model, and output the cultivated land type of each pixel in the current remote sensing image; obtain a target pixel in which the cultivated land type is a bare land type, and determine a fallow land block in a farmland corresponding to the target pixel based on an imaging time of the current remote sensing image and crop rotation information of the farmland corresponding to the target pixel. The device is further configured to: calculate maximum mean difference value information of the current remote sensing image and each historical remote sensing image in the historical remote sensing image set; determine a historical remote sensing image subset in which a sum of weights of maximum mean difference value information is greater than a first preset threshold and a sum of image quantities is less than a second preset threshold, and the historical remote sensing image subset includes N target historical remote sensing images.

5. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the cultivated land remote sensing image analysis method according to any one of claims 1 to 3 when executing the program.

6. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program implements the cultivated land remote sensing image analysis method according to any one of claims 1 to 3 when executed by the processor.

7. A computer program product comprising a computer program, characterized in that, The computer program implements the cultivated land remote sensing image analysis method according to any one of claims 1 to 3 when executed by the processor.

Images