A jasmine remote sensing image planting area identification method based on texture features and global attention

Abstract

This invention discloses a method for identifying jasmine planting areas in remote sensing images based on texture features and global attention, comprising the following steps: Step S1. Dataset creation; Step S2. Image segmentation network construction: using a hierarchical encoder, a small convolutional kernel module, and a lightweight decoder as an image classification network; Step S3. Model training; Step S4. Model inference: loading the trained image segmentation network model parameters from step (3) into the image segmentation network from step (2), and sequentially inputting the remote sensing image data of jasmine planting areas from the test set in step (1) into the network for inference to obtain segmentation prediction image data. Based on the correlation characteristics of remote sensing images of jasmine planting areas, this invention improves the network structure of the algorithm model, enhances the speed and accuracy of segmentation and recognition, reduces model size, and improves model generalization.

Description

Technical Field

[0001] This invention relates to the field of jasmine remote sensing image segmentation technology, and more specifically, to a method for identifying planting areas in jasmine remote sensing images based on texture features and global attention. Background Technology

[0002] Due to the unique cultivation methods and growing environment of jasmine, artificially planted jasmine flowers are neatly arranged and concentrated. In satellite remote sensing images, jasmine planting areas exhibit distinctive texture features that are easily discernible. Therefore, deep learning network models targeting these texture features can be designed to improve the accuracy of identifying jasmine planting areas in remote sensing images. However, in some regions, jasmine planting areas are scattered and overlapping, making it difficult to estimate annual jasmine yields. Furthermore, the high rate of overlapping measurements leads to significant government duplication of subsidies. Therefore, developing a sophisticated and efficient model for identifying overlapping planting areas, and achieving accurate, efficient, and rapid identification of subsidized planting areas for improved varieties, will help solve the problem of planting area identification across the entire jasmine industry.

[0003] In recent years, with the development of computer vision technology, artificial intelligence algorithms based on image classification and object detection can achieve segmentation and recognition of remote sensing images. However, for high-resolution remote sensing images, traditional image segmentation methods usually involve cropping high-resolution images into multiple low-resolution images. Simultaneously, the image's location encoding information needs to be recorded in the network model, resulting in the network model only learning low-resolution image features. Furthermore, the network model has many parameters, large size, and insufficient recognition accuracy and speed, leading to less than ideal performance and making it difficult to apply in production and daily life. This patent proposes a hierarchical network structure that can generate high-resolution shallow features and low-resolution fine features. It eliminates the location encoding module, uses small convolutional kernels to represent location information, making it easier to capture texture and multi-scale feature information. Simultaneously, it uses a lightweight decoding module to construct the network model and introduces a global attention mechanism in the decoder to solve the problems of traditional remote sensing image segmentation methods. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying jasmine planting areas in remote sensing images based on texture features and global attention. Taking advantage of the correlation between remote sensing images of jasmine planting areas, the method improves the network structure of the algorithm model, thereby increasing the speed and accuracy of segmentation and identification, reducing model size, and enhancing model generalization.

[0005] To achieve the above objectives, a method for identifying planting areas in remote sensing images of jasmine flowers based on texture features and global attention is provided, comprising the following steps:

[0006] Step S1. Dataset creation: Collect a large amount of remote sensing image data of jasmine planting areas, and divide the image data into training set, validation set and test set according to a certain ratio;

[0007] Step S2. Constructing an image segmentation network: A hierarchical encoder, a small convolutional kernel module, and a lightweight decoder are used as an image classification network. The hierarchical encoder is used to divide the input image data into several image blocks and transform the image blocks into feature map data with multi-scale features and prominent texture features. The small convolutional kernel module is used to label the position information of each image block. The lightweight decoder is used to decode the feature map data to obtain the segmentation prediction image data of the input image data and output it. The segmentation prediction image data is image data that can predict the planting area of ​​jasmine flowers.

[0008] Step S3. Model training: Input the remote sensing image data of the jasmine planting area from the training set of step (1) into the image segmentation network in step (2), and use the network constructed in step (3) to train the parameters of the image segmentation network. During the training process, input the validation set of step (1) into the image segmentation network for validation; finally, obtain the image segmentation network model parameters.

[0009] Step S4. Model Inference: Load the image segmentation network model parameters trained in step (3) into the image segmentation network in step (2), and input the remote sensing image data of the jasmine planting area in the test set in step (1) into the network for inference to obtain the segmentation prediction image data.

[0010] Specifically, in step S1, the image data is divided into training set, validation set and test set in a ratio of 0.8:0.1:0.1.

[0011] Specifically, the design of the layered encoder includes the following steps:

[0012] (1) Divide the input image data into several image blocks;

[0013] (2) Convert the image patch into a feature vector using the following formula:

[0014] (1);

[0015] in, Indicates the height of the input image. This indicates the width of the input image. This represents the dimension of each image patch. Indicates the dimension of the i-th image patch;

[0016] (3) Add scaling factor ;

[0017] (4) Using formula (2) Feature maps transformed The first feature map data of dimension is then transformed using formula (3) into Second feature map data of dimension:

[0018] ;

[0019] ;

[0020] Where N is the number of all image patches;

[0021] The second feature map data is used as the feature map data input to the lightweight decoder.

[0022] Specifically, in step S3, the optimizer for training the image segmentation network is AdamW, and the learning rate is set to 0.01.

[0023] Specifically, the small convolutional kernel module employs a feedforward neural network and utilizes the feedforward neural network's... The convolutional kernel is used to represent the positional information between image blocks.

[0024] Specifically, the lightweight decoder consists of several decoders, each of which is composed of three interconnected sub-layers, including a multi-head attention mechanism, a normalization layer, and a feedforward fully connected network.

[0025] Specifically, after step S2, the remote sensing image of the jasmine planting area is normalized to [-1, 1].

[0026] Specifically, the segmentation prediction image data is image data in which the planting area of ​​jasmine flowers is selected and colored.

[0027] The technical principles and beneficial effects of this invention are as follows:

[0028] 1. The image segmentation network of this invention is configured with a hierarchical encoder and a lightweight decoder:

[0029] The hierarchical encoder divides the input image data into several image patches, transforms each patch into a feature vector using the following formula, and adds a scaling factor. The data is then converted into first feature map data and second feature map data, and finally the second feature map data is output to reduce the computational complexity of the image segmentation network and highlight the texture feature information.

[0030] The lightweight decoder consists of several decoders, each composed of three interconnected sub-layers. These sub-layers include a multi-head attention mechanism, a normalization layer, and a feedforward fully connected network, with residual connections directly used. Compared to traditional decoders, the lightweight decoder has a simpler structure and better performance. While avoiding redundant computation, the lightweight decoder uses larger convolutional kernels to expand the model's receptive area to the image. It utilizes the feature information output from the convolutional layers to generate global attention features, introducing a global attention mechanism. During training, it considers all inputs from the encoder sequence to decode the feature map data, obtaining segmentation prediction image data from the input image data, and outputting this segmentation prediction image data, which is capable of predicting the planting area of ​​jasmine flowers.

[0031] This invention obtains high-precision image segmentation network model parameters and an image segmentation network through long-term and repeated training. The segmentation prediction image data obtained using this image segmentation network can generate high-resolution shallow features and low-resolution fine features, making it easier to capture texture and multi-scale feature information. At the same time, the network model is constructed using a lightweight decoding module, which can solve the problems of traditional remote sensing image segmentation methods faster and more accurately, reduce model size, and improve model generalization.

[0032] 2. This invention uses small convolutional kernel modules to label the position information of each image block, eliminating the need to record the image position encoding information in the network model, which prevents existing network models from learning only low-resolution image features. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0035] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0036] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0037] It should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. In addition, the terms "first," "second," and "third," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0038] Furthermore, terms such as "horizontal," "vertical," and "sag" do not imply that components must be absolutely horizontal or suspended, but rather that they can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal relative to "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0039] like Figure 1 An embodiment of the present invention provides a method for identifying planting areas of jasmine in remote sensing images based on texture features and global attention, comprising the following steps:

[0040] Step S1. Dataset creation: Collect a large amount of remote sensing image data of jasmine planting areas, and divide the image data into training set, validation set and test set according to a certain ratio;

[0041] Step S2. Constructing an image segmentation network: A hierarchical encoder, a small convolutional kernel module, and a lightweight decoder are used as an image classification network. The hierarchical encoder is used to divide the input image data into several image blocks and transform the image blocks into feature map data with multi-scale features and prominent texture features. The small convolutional kernel module is used to label the position information of each image block. The lightweight decoder is used to decode the feature map data to obtain the segmentation prediction image data of the input image data and output it. The segmentation prediction image data is image data that can predict the planting area of ​​jasmine flowers.

[0042] Step S3. Model training: Input the remote sensing image data of the jasmine planting area from the training set of step (1) into the image segmentation network in step (2), and use the network constructed in step (3) to train the parameters of the image segmentation network. During the training process, input the validation set of step (1) into the image segmentation network for validation; finally, obtain the image segmentation network model parameters.

[0043] Step S4. Model Inference: Load the image segmentation network model parameters trained in step (3) into the image segmentation network in step (2), and input the remote sensing image data of the jasmine planting area in the test set in step (1) into the network for inference to obtain the segmentation prediction image data.

[0044] Specifically, in step S1, the image data is divided into training set, validation set and test set in a ratio of 0.8:0.1:0.1.

[0045] Specifically, the design of the layered encoder includes the following steps:

[0046] (1) Divide the input image data into several image blocks;

[0047] (2) Convert the image patch into a feature vector using the following formula:

[0048] (1);

[0049] in, Indicates the height of the input image. This indicates the width of the input image. This represents the dimension of each image patch. Indicates the dimension of the i-th image patch;

[0050] (3) Add scaling factor ;

[0051] (4) Using formula (2) Feature maps transformed The first feature map data of dimension is then transformed using formula (3) into Second feature map data of dimension:

[0052] ;

[0053] ;

[0054] Where N is the number of all image patches;

[0055] The second feature map data is used as the feature map data input to the lightweight decoder.

[0056] Specifically, in step S3, the optimizer for training the image segmentation network is AdamW, and the learning rate is set to 0.01.

[0057] Specifically, the small convolutional kernel module employs a feedforward neural network and utilizes the feedforward neural network's... The convolutional kernel is used to represent the positional information between image blocks.

[0058] Specifically, the lightweight decoder consists of several decoders, each of which is composed of three interconnected sub-layers, including a multi-head attention mechanism, a normalization layer, and a feedforward fully connected network.

[0059] Specifically, after step S2, the remote sensing image of the jasmine planting area is normalized to [-1, 1].

[0060] Specifically, the segmentation prediction image data is image data in which the planting area of ​​jasmine flowers is selected and colored.

[0061] Although embodiments of the present invention have been described in conjunction with the accompanying drawings, the patent owner may make various modifications or alterations within the scope of the appended claims, as long as they do not exceed the protection scope described in the claims of the present invention, they shall be within the protection scope of the present invention.

Claims

1. A method for identifying planting areas of jasmine in remote sensing images based on texture features and global attention, characterized in that, Includes the following steps: Step S1. Dataset creation: Collect a large amount of remote sensing image data of jasmine planting areas, and divide the image data into training set, validation set and test set according to a certain ratio; Step S2. Constructing an image segmentation network: A hierarchical encoder, a small convolutional kernel module, and a lightweight decoder are used as an image classification network. The hierarchical encoder is used to divide the input image data into several image blocks and transform the image blocks into feature map data with multi-scale features and prominent texture features. The small convolutional kernel module is used to label the position information of each image block. The lightweight decoder is used to decode the feature map data to obtain the segmentation prediction image data of the input image data and output it. The segmentation prediction image data is image data that can predict the planting area of ​​jasmine flowers. The design of the layered encoder includes the following steps: (1) Divide the input image data into several image blocks; (2) Convert the image patch into a feature vector using the following formula: （1）； in, Indicates the height of the input image. This indicates the width of the input image. This represents the dimension of each image patch. Indicates the dimension of the i-th image patch; (3) Add scaling factor ; (4) Using formula (2) Feature maps transformed The first feature map data of dimension is then transformed using formula (3) into Second feature map data of dimension: ； ； Where N is the number of all image patches; the small convolutional kernel module employs a feedforward neural network, and uses the feedforward neural network's... The convolutional kernels are used to represent the positional information between image blocks; the lightweight decoder consists of several decoders, each of which consists of three sub-layers connected together, and the sub-layers contain a multi-head attention mechanism, a normalization layer and a feedforward fully connected network; The second feature map data is used as the feature map data input to the lightweight decoder; Step S3. Model Training: Input the remote sensing image data of the jasmine planting area from the training set of Step S1 into the image segmentation network in Step S2. Use the network constructed in Step S2 to train the parameters of the image segmentation network. During the training process, input the validation set from Step S1 into the image segmentation network for validation. Finally, obtain the image segmentation network model parameters. Step S4. Model Inference: Load the image segmentation network model parameters trained in Step S3 into the image segmentation network of Step S2, and input the remote sensing image data of the jasmine planting area in the test set of Step S1 into the network for inference to obtain segmentation prediction image data.

2. The method for identifying planting areas of jasmine in remote sensing images based on texture features and global attention as described in claim 1, characterized in that, In step S1, the image data is divided into training set, validation set and test set in a ratio of 0.8:0.1:0.

1.

3. The method for identifying planting areas of jasmine in remote sensing images based on texture features and global attention as described in claim 1, characterized in that, In step S3, the optimizer for training the image segmentation network is AdamW, and the learning rate is set to 0.

01.

4. The method for identifying planting areas of jasmine in remote sensing images based on texture features and global attention as described in claim 1, characterized in that, After step S2, the remote sensing image of the jasmine planting area is normalized to [-1, 1].

5. The method for identifying planting areas of jasmine in remote sensing images based on texture features and global attention as described in claim 1, characterized in that, The segmentation prediction image data is image data in which the planting area of ​​jasmine flowers is selected and colored.

Images