Rape remote sensing identification method

Abstract

The invention belongs to the technical field of satellite remote sensing, and particularly relates to a rape remote sensing recognition method which comprises the following steps: S1, data acquisition and preprocessing; s2, sample size calculation and sample drawing; s3, model building and training are carried out, and a network model comprises a U2-Net network model and an HRNet network model; s4, model precision evaluation; s5, outputting a result; and S6, result verification. According to the method, a deep learning rape identification model which combines U2-Net and HRNet and faces a provincial scale is provided for regional differences, capturing of overall multi-scale information by the U2-Net is considered, the identification effect of the HRNet on small-region rape pattern spots is also considered, and an important basis is provided for macroscopically mastering rape planting and optimizing a planting structure in China. Meanwhile, input images do not need to be processed excessively during operation, specialized operation is simplified, and the method is high in reproducibility and easy to popularize and use.

Description

Technical Field

[0001] The present application belongs to the field of satellite remote sensing technology, and specifically relates to a rapeseed remote sensing identification method. Background Art

[0002] Rapeseed is a major cash crop and a major source of edible oil in my country. Timely and accurate acquisition of its planting distribution information is an important basis for strengthening production management and optimizing the spatial pattern of crop planting.

[0003] At present, satellite remote sensing technology is the main means to quickly and accurately obtain the spatial pattern and evolution of rapeseed planting. Remote sensing identification methods for rapeseed are mainly divided into traditional machine learning identification methods and deep learning identification methods that have emerged in recent years.

[0004] Among them, the characteristics of machine learning recognition methods are to design specific indexes or spectral, spatial and other features for rapeseed targets. However, machine learning recognition methods need to rely on researchers' prior knowledge to design spectral, spatial and other extraction features. Such features are often shallow features that cannot fully and accurately reflect the intrinsic semantic information of image data, and therefore lack the representation and utilization of high-level semantic features. Since such features are artificially designed, they must be repeatedly selected and cross-validated before they can be effectively applied, and their scalability and portability are not strong.

[0005] The deep learning rapeseed identification method can effectively avoid the shortcomings of the traditional rapeseed identification method, which relies on artificial screening features and has poor generalization, and has become a research hotspot in the field of remote sensing identification in recent years. Convolutional neural network is a new type of data-driven technology with the characteristics of multi-layer convolution. U-Net, DeepLab, HRNet, etc. are the main network structures popular in the field of deep learning in recent years. After multiple consecutive downsampling, they can extract richer semantic information from the input samples.

[0006] However, most of the existing deep learning research on rapeseed identification uses sub-meter high-resolution images. Due to the high cost of image acquisition and the large amount of data, its effective identification range is limited (suitable for village and town scale, at most county scale), making it difficult to promote and apply. Summary of the invention

[0007] In order to solve at least one technical problem existing in the prior art, the present application provides a rapeseed remote sensing identification method.

[0008] The present application discloses a method for remote sensing identification of rapeseed, comprising the following steps:

[0009] S1. Obtain an optical remote sensing image of rapeseed in a region to be measured, and preprocess the optical remote sensing image, wherein the region to be measured is a provincial-level region;

[0010] S2, according to the spatial distribution characteristics of rapeseed planting in the test area, determine the number, distribution and size of the sample frame, and select a predetermined number of original sample images from the optical remote sensing image preprocessed in step S1, and then disperse the original sample images into the sample frame, and then perform sliding window cropping and image segmentation processing, complete sample labeling in a human-computer interaction manner, and finally construct an image and label training data set by a cropping method to obtain a sample data set;

[0011] S3, augmenting the sample data set, and dividing the augmented sample data set into a training set, a validation set, and a test set, adjusting the model parameters of the U2-Net and HRNet network models, and importing the training set and validation set into the U2-Net and HRNet network models for training and optimization respectively;

[0012] S4, extract rapeseed from the U2-Net and HRNet network models trained and optimized in step S3, perform AND operation on the two model extraction results to obtain the final result, and then evaluate the accuracy of the result through the test set. If the accuracy requirement is met, proceed to step S5. If the accuracy requirement is not met, return to steps S2 and S3, check sample drawing, and adjust and optimize each network model until the requirement is met;

[0013] S5. Use the U2-Net and HRNet network models that meet the requirements in step S4, and apply them to the optical remote sensing images of rapeseed in the area to be tested, respectively, predict the images scene by scene, and splice the prediction results to obtain rapeseed planting distribution maps of the area to be tested by the two models. Finally, perform an AND operation on the two rapeseed planting distribution maps to obtain the final rapeseed spatial distribution map and rapeseed area of ​​the area to be tested.

[0014] In an optional embodiment, in step S1, the optical remote sensing image is L2 level surface reflectance data, and the preprocessing of the optical remote sensing image includes at least band combination, stitching and cropping, wherein the optical remote sensing image after band combination contains at least five bands: blue, green, red, near red and short-wave infrared.

[0015] In an optional implementation, when the sliding window clipping process is performed in step S2, the sliding window overlap rate is 40%.

[0016] In an optional implementation, when performing image segmentation processing in step S2, a segmentation object whose target is rapeseed is selected, and the rapeseed in the sample frame is quickly extracted based on at least the minimum distance and color matching features.

[0017] In an optional implementation, in step S2, after the image segmentation process, the following steps are further included:

[0018] The image is visually corrected to obtain a vector patch marked as rapeseed; the vector patch is verified for correctness and vector boundary fitting in the field, and the vector patch is corrected again based on the verification results to obtain a vector patch marked as rapeseed, completing the sample labeling.

[0019] In an optional implementation, in step S2, the vector image spots are cropped using a regular grid of 224*224 pixel size.

[0020] In an optional implementation, in step S3, the improved U2-Net and HRNet network models are trained based on the pytorch framework, wherein the model parameters and the settings of each model parameter are respectively:

[0021] The optimization function, which is set to the stochastic gradient descent optimizer;

[0022] Loss function, which is set to the cross entropy loss function;

[0023] The initial value of the learning rate is set to 0.001;

[0024] Learning rate, which is set to warmup strategy and cosine annealing strategy;

[0025] Batch size, which is set to 4.

[0026] In an optional implementation, in step S3, the optimization of the U2-Net and HRNet network models includes:

[0027] The sample set is remade by supplementing samples from missed areas and modifying samples from mistaken areas. The model to be optimized is retrained on the new sample set as the initialization model to generate a new model, and then prediction is made. This process is repeated many times until the model achieves the expected effect.

[0028] In an optional implementation, in step S4, the accuracy of the U2-Net and HRNet network models is evaluated by the following formula:

[0029]

[0030] Among them, F1 represents the F1 score that measures the accuracy of the rapeseed extraction model, R represents the recall rate, P represents the precision rate, TP represents the rapeseed area correctly extracted by the model, FN represents the rapeseed area missed by the model, and FP represents the rapeseed area incorrectly extracted by the model;

[0031] Among them, when F1≥0.9, it is determined that the accuracy requirement is met.

[0032] In an optional embodiment, the rapeseed remote sensing identification method further includes the following steps:

[0033] S6. Verify the rapeseed area of ​​the tested area obtained in step S5 by the following formula:

[0034]

[0035] Among them, P a Indicates the extraction accuracy based on sample area; P n Indicates the extraction accuracy based on the number of samples; A1 indicates the area of ​​rapeseed identified by remote sensing in the sample area; A2 indicates the actual area of ​​rapeseed in the sample area; N1 indicates the number of samples correctly extracted by the model; N2 indicates the total number of samples

[0036] This application has at least the following beneficial technical effects:

[0037] 1) A rapeseed recognition method combining object-oriented and deep learning was constructed;

[0038] Deep learning remote sensing recognition methods have high requirements for sample accuracy. Completely manual delineation methods are prone to omissions and inaccurate delineation boundaries. Pixel-based machine classification methods usually only use the spectral information of each pixel, ignoring the effective use of spatial context information, resulting in salt and pepper noise in the target sample. Object-oriented classification methods can avoid salt and pepper noise to a certain extent, but the level of automation needs to be improved. This application first adopts an object-oriented method to make sample labels, then accurately corrects the sample data through visual interpretation, and finally verifies the accuracy of the sample data through field investigation, ensuring the authenticity of rapeseed samples in many ways. The advantages of various methods such as object-oriented, internal correction, and field investigation are comprehensively used to construct provincial-scale rapeseed samples.

[0039] 2) Aiming at regional differences, a deep learning rapeseed recognition model combining U2-Net and HRNet was proposed;

[0040] The north-south differences in provincial rapeseed planting determine that it is difficult to accurately extract the province's rapeseed using only one model. The U2-Net network structure uses the fusion of residual U blocks at different scales and different receptive fields to capture the contextual information of multi-scale feature maps; while HRNet connects the convolution flows from high resolution to low resolution in parallel, maintaining high resolution, rather than restoring from low resolution to high resolution, and is more accurate in processing local details. In order to take into account the possibility of multiple planting conditions in the collection area, this application considers both U2-Net's capture of overall multi-scale information and HRNet's recognition effect on small-area rapeseed patches, and uses U2-Net and HRNet network structures to train deep learning models respectively to extract provincial rapeseed distribution. Finally, the results extracted by the two models are unioned to obtain a rapeseed planting distribution map for the province.

[0041] 3) Rapeseed extraction at provincial level;

[0042] In the past, most rapeseed extraction studies based on medium and high resolution images tended to focus on the spatial scales of villages, towns, counties and cities. Due to the large scope of the provincial scale and the difficulty of extracting rapeseed at different planting times, there are few reported rapeseed extraction methods at the provincial scale. This application uses free Sentinel-2 images and deep learning technology to build a rapeseed remote sensing identification model for a large spatial scale, providing an important basis for the country to master rapeseed planting and optimize planting structure at a macro level. At the same time, there is no need to process the input image too much during operation, which simplifies professional operations, and the method is highly reproducible and easy to promote and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a technical flow chart of the rapeseed remote sensing identification method of this application;

[0044] Figure 2 The rapeseed remote sensing identification method of this application adopts the 432 band combination display remote sensing image of Anhui Province from April to May by Sentinel-2;

[0045] Figure 3 It is the regional distribution map of rapeseed samples in Anhui Province used in the rapeseed remote sensing identification method of this application;

[0046] Figure 4 It is a schematic diagram of the structure of the U2-Net network model in the rapeseed remote sensing identification method of this application;

[0047] Figure 5 It is a schematic diagram of the HRNet network model structure in the rapeseed remote sensing identification method of this application;

[0048] Figure 6 It is the distribution map of rapeseed extraction results in Anhui Province used in the rapeseed remote sensing identification method of this application;

[0049] Figure 7 This is the distribution map of rapeseed field sampling survey in Anhui Province used in the rapeseed remote sensing identification method of this application. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the implementation of this application clearer, the technical solution in the embodiment of this application will be described in more detail below in conjunction with the drawings in the embodiment of this application. The described embodiment is a part of the embodiment of this application, not all of the embodiments.

[0051] The following is combined with Figure 1-7 The map of Anhui Province further illustrates the rapeseed remote sensing identification method of this application.

[0052] The present application discloses a method for remote sensing identification of rapeseed, which specifically comprises the following steps.

[0053] Step S1: data acquisition and preprocessing.

[0054] The image data used in this application is Sentinel-2 optical imagery, obtained from the official website of the European Space Agency. Based on the image feature comparison of rapeseed during its growth period, cloud-free images from April and May 2023, when rapeseed is in full bloom, were selected for download. The downloaded dataset is L2-level surface reflectance data, and band combination processing is performed in remote sensing software such as envi.

[0055] like Figure 2 As shown in the figure, the processed image contains at least five bands: blue, green, red, near-red, and short-wave infrared. The combined image is further pre-processed by stitching and cropping, and finally a cloud-free or low-cloud image of the entire Anhui province with a spatial resolution of 10 meters is obtained (432 band combination display).

[0056] Step S2: sample size calculation and sample drawing.

[0057] The drawing of rapeseed samples must ensure both representativeness and quantity. In Anhui Province, rapeseed is generally planted more in the south and less in the north. Compared with crops such as wheat and rice, the planting distribution is scattered, and due to the influence of terrain, altitude, and climate, the phenology of rapeseed varies from place to place. Therefore, the selection of sample areas should take into account the different conditions in different places, so as to ensure the diversity and representativeness of rapeseed samples. Considering the proportion of rapeseed planting area in each city in Anhui Province in 2022, the rapeseed planting area of ​​Anqing, Hefei, Lu'an, Wuhu, Chizhou, Ma'anshan, Xuancheng, and Tongling accounts for about 83% of the province, so the samples are mainly selected in these eight cities. Combined with the characteristics of remote sensing images, in accordance with the principle of presenting the diversity of rapeseed and the complexity of the background as much as possible in the sample frame area, this study set up sample frames in 16 different areas. For details, please see Table 1:

[0058] Table 1 Statistical data of rapeseed planting and distribution of rapeseed samples in this study

[0059]

[0060]

[0061] To train a high-accuracy deep learning extraction model, a certain number of sufficiently labeled training samples are required as a driver. Considering that the number of sample images required by deep convolutional neural networks is generally more than 5,000, as well as the means of data augmentation and expansion of the sample set such as horizontal flipping, vertical flipping, image rotation (optional rotation of 90°, 180°, 270°), about 1,600 original sample images are required in this application. In order to ensure sample diversity, we set up sample frames in 16 different regions according to the spatial distribution characteristics of rapeseed planting in Anhui Province, and 1,600 original sample images were scattered into 16 sample frames, so each sample frame needs to have about 100 sample images. According to the sample frame design as a square, each sample frame can have 10*10=100 sample images. The input sample image size of the convolutional neural network cannot be too large, generally 224*224 pixels in size.

[0062] In order to make the sample images more complete during training and avoid position errors during subsequent stitching, a sliding window cropping method is adopted in each sample frame, the sliding window overlap rate is set to 40%, and the horizontal and vertical sliding window steps are both 10 (that is, to ensure that there are 10*10 sample images). Taking the image with a spatial resolution of 10 meters in Sentinel-2 as an example, the side length of each sample frame is 1*224*10+9*(1-0.4)*224*10=14336m=14km.

[0063] After determining the number, distribution and size of the sample frames, the image in the sample area is segmented using an object-oriented method. In order to achieve better segmentation and subsequent accurate drawing of rapeseed samples, this application uses different segmentation scales based on different sample image features. Specifically, the target is the segmentation object of rapeseed, and the rapeseed in the sample frame is quickly extracted based on features such as minimum distance and color matching.

[0064] Furthermore, visual correction was performed in ArcGIS software to obtain vector patches marked as rapeseed. The correctness of the vector patches and the vector boundary fitting were checked in the field, and the patches were further corrected according to the verification results to obtain high-precision and representative vector patches marked as rapeseed (or sample patches) (see Figure 3). The sample frame and the samples in the frame can be added, checked, modified or deleted later according to the actual situation and extraction effect. After the sample is labeled, the regular grid (i.e., 224*224 pixel size) cropping method is used to combine the labeled image with the sample labeling result to construct an image and label training dataset, thus completing the production of the sample dataset.

[0065] Finally, 16 plots with a total area of ​​196 km were evenly distributed in Anhui Province. 2 In the sample area, a combination of object-oriented, visual interpretation and field investigation methods were used to accurately map the rapeseed samples within the sample frame, with a total of 26,500 surface patches.

[0066] Step S3: Model building and training.

[0067] Before model training, the prepared sample set is first augmented, including left-right flipping, up-down flipping, and counterclockwise rotation of 90°, 180°, and 270°. The augmented data set is 6 times the original data set. The augmented data set is then randomly divided into training set, validation set, and test set in a ratio of 6:2:2. Then, the training set and validation set are put into the model (U2-Net and HRNet network models) for training.

[0068] Among them, U2-Net (or U 2 -Net) network structure has shown excellent performance in the field of remote sensing target extraction. Therefore, this application uses the U2-Net network structure as the basic network.

[0069] like Figure 4 As shown in the figure, U2-Net is a two-level nested U-shaped structure. The outermost layer is a large U-shaped structure composed of an encoder (left part) and a decoder (right part). The encoder extracts the spatiotemporal features of the original image layer by layer through multi-layer convolution and pooling operations, and performs downsampling. The decoder is used to decode and reconstruct the features extracted by the encoder to obtain the classification result, which is an upsampling process. The feature signal after each downsampling is connected to the decoding part, so that the feature map restored by upsampling contains more low-level features. Each encoding block and decoding block is filled with a small residual U-block (RSU). RSU is to solve the problem of network degradation that occurs as the depth of the network increases. The residual U-block directly bypasses the input information to the output, so that the subsequent layers can directly learn the residual, which protects the integrity of the information while simplifying the target and difficulty, so that each encoding block / decoding block network only needs to learn the difference signal between the input and output, and adjust the connection weights of each layer according to the difference.

[0070] The above structure can extract multi-scale features and multi-level features more efficiently. In the encoding stage, the first 4 RSUs reduce the dimension of the feature map, increase the receptive field, and obtain multi-scale information. In order to avoid the loss of details caused by excessive downsampling, the last 2 RSUs use dilated convolution instead of pooling. In the decoding stage, a similar structure to the encoding stage is used, and the connection of the upsampled feature map from its previous stage and the downsampled feature map from its symmetric encoder stage is used as input to pass context information. Finally, the output of the multi-scale features at each level is spliced ​​and fused to generate the final result.

[0071] Different from the idea of ​​U2-Net structure that aggregates high resolution by upsampling low resolution, the biggest feature of HRNet network structure is that the convolution flow from high resolution to low resolution is connected in parallel instead of serially, such as Figure 5 As shown in Figure 1, the resolution of each stream in the neural network is the same. Since HRNet connects high-resolution and low-resolution networks in parallel instead of restoring from low resolution to high resolution, this method maintains high resolution and can more accurately represent the spatial features of the input information.

[0072] Furthermore, during the model building and training process, the present application first improves the performance of the model on the test set by adjusting the model's hyperparameters, where the hyperparameters (i.e., model parameters) include: optimization function, loss function, learning rate, batch size, and number of training times.

[0073] This application is based on the pytorch framework for model building and training, and the above hyperparameters are specifically set as follows:

[0074] The optimization function is the stochastic gradient descent optimizer; the loss function is the cross entropy loss function; the initial value of the learning rate is 0.001, and as the number of training iterations increases, the learning rate strategy is set to warmup and cosine annealing strategy to avoid the problem of a fixed learning rate falling into a singularity when searching for the optimal solution during training; the batch size is 4.

[0075] Furthermore, according to the actual prediction effect of the model on the image, the sample set can be remade by supplementing samples in the missed areas, modifying samples in the mistaken areas, etc. The model to be optimized is retrained on the new sample set as the initialization model to generate a new model, and then prediction is performed. The whole process may be iterated many times until the model achieves the expected effect, that is, the model optimization is completed.

[0076] Step S4: model accuracy evaluation.

[0077] Specifically, this step is to use the trained network model to perform accuracy test on the test set. After the accuracy evaluation meets the requirements (i.e., F1 is greater than or equal to 0.9), the rapeseed prediction is started. Otherwise, the sample drawing, parameter setting, etc. are rechecked, and the model is adjusted and optimized until the requirements are met.

[0078] Among them, the accuracy of U2-Net and HRNet network models is evaluated by the following formula:

[0079]

[0080] Among them, F1 represents the F1 score that measures the accuracy of the rapeseed extraction model, R represents the recall rate, P represents the precision rate, TP represents the rapeseed area correctly extracted by the model, FN represents the rapeseed area missed by the model, and FP represents the rapeseed area incorrectly extracted by the model.

[0081] In this embodiment, the F1 score based on the confusion matrix is ​​first used. According to formulas (1)-(3), the F1 values ​​of the U2-Net and HRNet network structure training models are 0.86 and 0.70 respectively. Then, the recognition results of the two models are fused to obtain a final rapeseed extraction F1 score of 0.94, indicating that the fusion of the two models can better extract the rapeseed planting area.

[0082] Step S5, apply the trained U2-Net model and HRNet model that meet the accuracy evaluation requirements to the images of Anhui Province, predict the images scene by scene, and then splice the prediction results of all scene images to obtain the rapeseed distribution area of ​​the province (i.e., rapeseed planting distribution map). Then, the two results are ANDed to obtain the final rapeseed spatial distribution map and rapeseed area of ​​Anhui Province. In this embodiment, the rapeseed area extracted is 4.3025 million mu.

[0083] Step S6: Result verification.

[0084] The rapeseed distribution area and area results collected in the field are used as the true value, and the rapeseed recognition accuracy is calculated using the following formula:

[0085]

[0086] Among them, P a Indicates the extraction accuracy based on sample area; P n It represents the extraction accuracy based on the number of samples; A1 represents the area of ​​rapeseed identified by remote sensing in the sample area; A2 represents the actual area of ​​rapeseed in the sample area; N1 represents the number of samples correctly extracted by the model; N2 represents the total number of samples.

[0087] In this embodiment, Figure 6As shown, the technical method disclosed in the present invention has been applied in Anhui Province for testing and used to extract rapeseed in Anhui Province in 2023.

[0088] Specifically, Anhui Province has 16 blocks with a total area of ​​196 km. 2 In the sample area, a combination of object-oriented, visual interpretation and field investigation methods were used to accurately map the rapeseed samples within the sample frame, with a total of 26,500 surface patches.

[0089] Among them, 60% is used for model training, 20% for model verification, and 20% for result testing. Based on the U2-Net and HRNet network structures, the deep learning models are trained separately, and then the scene-by-scene image prediction is performed. Then, the prediction results of all scene images are spliced ​​to obtain the rapeseed distribution area of ​​the province. Finally, the two results are ANDed to obtain the rapeseed extraction result of Anhui Province, and the extracted rapeseed area is 4.3025 million mu.

[0090] To verify the feasibility of the proposed method, the F1 score based on the confusion matrix was first used according to formula (1) to formula (3). The F1 values ​​of the U2-Net and HRNet network structure training models were 0.86 and 0.70 respectively. Then the recognition results of the two models were fused to obtain the final rapeseed extraction F1 score of 0.94.

[0091] At the same time, if Figure 7 As shown in the figure, a field sampling survey was conducted in Hefei, Ma'anshan, Anqing and other places in Anhui Province. According to the actual distribution of rapeseed, 462 plots with a total area of ​​5,005,677 m 2 The surface vector map was used as the basis to evaluate the extraction accuracy of rapeseed in Anhui Province.

[0092] According to formula (4) and formula (5), P is calculated a The value is 94.63%, P n The value is 90.91%. The above results show that the technical method of this application is feasible, and the deep learning model combining the U2-Net and HRNet network structures can be used for rapeseed remote sensing extraction in Anhui Province.

[0093] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A method for remote sensing identification of rapeseed, characterized in that: The following steps are involved: S1. Obtain an optical remote sensing image of rapeseed in a region to be measured, and preprocess the optical remote sensing image, wherein the region to be measured is a provincial-level region; S2, according to the spatial distribution characteristics of rapeseed planting in the test area, determine the number, distribution and size of the sample frame, and select a predetermined number of original sample images from the optical remote sensing image preprocessed in step S1, and then disperse the original sample images into the sample frame, and then perform sliding window cropping and image segmentation processing, complete sample labeling in a human-computer interaction manner, and finally construct an image and label training data set by a cropping method to obtain a sample data set; S3, augmenting the sample data set, and dividing the augmented sample data set into a training set, a validation set, and a test set, adjusting the model parameters of the U2-Net and HRNet network models, and importing the training set and validation set into the U2-Net and HRNet network models for training and optimization respectively; S4, extract rapeseed from the U2-Net and HRNet network models trained and optimized in step S3, perform AND operation on the two model extraction results to obtain the final result, and then evaluate the accuracy of the result through the test set. If the accuracy requirement is met, proceed to step S5. If the accuracy requirement is not met, return to steps S2 and S3, check sample drawing, and adjust and optimize each network model until the requirement is met; S5. Use the U2-Net and HRNet network models that meet the requirements in step S4, and apply them to the optical remote sensing images of rapeseed in the area to be tested, respectively, predict the images scene by scene, and splice the prediction results to obtain rapeseed planting distribution maps of the area to be tested by the two models. Finally, perform an AND operation on the two rapeseed planting distribution maps to obtain the final rapeseed spatial distribution map and rapeseed area of ​​the area to be tested.

2. The rapeseed remote sensing identification method according to claim 1, characterized in that: In step S1, the optical remote sensing image is L2 level surface reflectance data, and the preprocessing of the optical remote sensing image includes at least band combination, stitching and cropping, wherein the optical remote sensing image after band combination contains at least five bands: blue, green, red, near red and short-wave infrared.

3. The rapeseed remote sensing identification method according to claim 1, characterized in that: When the sliding window clipping process is performed in step S2, the sliding window overlap rate is 40%.

4. The rapeseed remote sensing identification method according to claim 1, characterized in that: When performing image segmentation processing in step S2, a segmentation object whose target is rapeseed is selected, and rapeseed in the sample frame is quickly extracted based on at least the minimum distance and color matching features.

5. The rapeseed remote sensing identification method according to claim 4, characterized in that: In the step S2, after the image segmentation process, the following steps are also included: The image is visually corrected to obtain a vector patch marked as rapeseed; the vector patch is verified for correctness and vector boundary fitting in the field, and the vector patch is corrected again based on the verification results to obtain a vector patch marked as rapeseed, completing the sample labeling.

6. The rapeseed remote sensing identification method according to claim 5, characterized in that: In step S2, the vector image spots are cut using a regular grid of 224*224 pixel size.

7. The rapeseed remote sensing identification method according to claim 1, characterized in that: In step S3, the improved U2-Net and HRNet network models are trained based on the pytorch framework, wherein the model parameters and the settings of each model parameter are respectively: The optimization function, which is set to the stochastic gradient descent optimizer; Loss function, which is set to the cross entropy loss function; The initial value of the learning rate is set to 0.001; Learning rate, which is set to warmup strategy and cosine annealing strategy; Batch size, which is set to 4.

8. The rapeseed remote sensing identification method according to claim 1, characterized in that: In step S3, the optimization of the U2-Net and HRNet network models includes: The sample set is remade by supplementing samples from missed areas and modifying samples from mistaken areas. The model to be optimized is retrained on the new sample set as the initialization model to generate a new model, and then prediction is made. This process is repeated many times until the model achieves the expected effect.

9. The rapeseed remote sensing identification method according to claim 1, characterized in that: In step S4, the accuracy of the U2-Net and HRNet network models is evaluated by the following formula: Among them, F1 represents the F1 score that measures the accuracy of the rapeseed extraction model, R represents the recall rate, P represents the precision rate, TP represents the rapeseed area correctly extracted by the model, FN represents the rapeseed area missed by the model, and FP represents the rapeseed area incorrectly extracted by the model; Among them, when F1≥0.9, it is determined that the accuracy requirement is met.

10. The rapeseed remote sensing identification method according to claim 1, characterized in that: The following steps are also included: S6. Verify the rapeseed area of ​​the tested area obtained in step S5 by the following formula: Among them, P a Indicates the extraction accuracy based on sample area; P n It represents the extraction accuracy based on the number of samples; A1 represents the area of ​​rapeseed identified by remote sensing in the sample area; A2 represents the actual area of ​​rapeseed in the sample area; N1 represents the number of samples correctly extracted by the model; N2 represents the total number of samples.

Images