Sugarcane identification method

Abstract

The invention discloses a sugarcane identification method, which comprises the steps that a remote sensing image of a to-be-identified area is acquired and input into a sugarcane identification model, and the sugarcane identification model outputs a sugarcane spatial distribution monitoring identification result; the sugarcane identification model is formed by coupling a sugarcane identification YOLOv8 model and a sugarcane identification U-Net model; the sugarcane identification YOLOv8 model is used for detecting and identifying a rough boundary of sugarcane planting; the sugarcane identification U-Net model is used for carrying out second identification on the basis of a sugarcane planting coarse boundary so as to identify and extract a sugarcane fine boundary. According to the method, a sugarcane result frame is extracted by using sugarcane identification YOLOv8, so that sugarcane coarse boundary extraction is realized; and then the sugarcane identification U-Net model is used to accurately identify the fine boundary of the sugarcane planting field, and rapid and accurate monitoring of the current situation and change of the sugarcane is realized, so that the monitoring demand of sugar crops in a high-standard farmland range is realized, and data support is provided for agricultural protection and management work.

Description

Technical Field

[0001] The invention relates to the field of remote sensing monitoring, and in particular to a sugarcane identification method. Background Art

[0002] Sugarcane is one of the important sugar crops in the world, accounting for about 80% of the global sugar production. Lingnan region is located in tropical and subtropical areas and has natural advantages in growing sugarcane. Among them, Guangdong has a sown area of ​​2.2079 million mu in 2022, with a total output of 12.9205 million tons. With the gradual increase in demand for sugarcane at home and abroad, the economic benefits of sugarcane production have been continuously improved, which has prompted the enthusiasm of Lingnan region to develop cash crop sugarcane. Therefore, the monitoring of cash crop sugarcane plays an important role in increasing farmers' income and stabilizing agricultural production.

[0003] With the development of remote sensing technology, the current method of quickly obtaining planting information in cultivated land has changed to identifying the scope of cultivated land in remote sensing satellite images and then using various classification algorithms to classify cultivated land. This method has changed the traditional operation mode and can quickly extract information in cultivated land. However, due to the fragmented planting structure of cultivated land in Lingnan region and the cloudy, rainy and foggy climate, remote sensing technology has difficulties in identifying sugarcane in Lingnan region. The YOLOv8 (You Only Look Once) model is known for its efficient single-step detection method, which can realize real-time and fast detection of objects. The end-to-end training method simplifies the training and reasoning process, and uses a single network structure to extract features and detect at the same time, which improves the simplicity and efficiency of the model and can achieve higher detection accuracy. However, since the YOLOv8 model performs coarse grid division on the image for detection, the accuracy of positioning small objects may be low, and the detection frame is prone to offset. In scenes with dense objects, the detection performance of YOLOv8 may have overlapping frames or missed detections. The U-Net model is good at fine-grained segmentation of objects and can accurately segment the boundaries of target objects. It can also be effectively trained for small data sets and has good generalization capabilities. However, since the U-Net network contains a large number of convolution operations and feature map fusion, it has high requirements on hardware conditions and high memory consumption. In addition, the U-Net model relies on accurately labeled data sets, which requires a lot of work. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies of the above-mentioned prior art and provide a sugarcane identification method to achieve rapid and accurate monitoring of the current status and changes of sugarcane.

[0005] To achieve the above object, the technical solution of the present invention is:

[0006] A sugarcane identification method, comprising:

[0007] Acquire a remote sensing image of the area to be identified and input it into a sugarcane identification model, wherein the sugarcane identification model outputs a sugarcane spatial distribution monitoring and identification result;

[0008] The sugarcane recognition model is formed by coupling the sugarcane recognition YOLOv8 model and the sugarcane recognition U-Net model; the sugarcane recognition YOLOv8 model is used to detect and recognize the coarse boundary of sugarcane planting; the sugarcane recognition U-Net model is used to perform a second recognition based on the coarse boundary of sugarcane planting to identify and extract the fine boundary of sugarcane.

[0009] Preferably, the sugarcane recognition YOLOv8 model is trained in the following manner:

[0010] Acquire remote sensing images containing sugarcane, perform geometric correction and orthorectification on the remote sensing images containing sugarcane, extract sample spots and classify them to form a vector sample library;

[0011] The vector samples are divided into pairs according to the set size to form training sample pairs, and some training samples are used as training sample data sets.

[0012] The training sample data set is input into the original YOLOv8 model for training to obtain the sugarcane recognition YOLOv8 model.

[0013] Preferably, the sugarcane recognition YOLOv8 model is used to detect the sugarcane planting area image to obtain the rough boundary result box sample data of the sugarcane field.

[0014] Preferably, in the process of inputting the training sample data set into the original YOLOv8 model for training, five data enhancement methods, namely, inversion, translation, rotation, tilt and scaling, are used to expand the training samples.

[0015] Preferably, the sugarcane recognition U-Net model is trained in the following manner:

[0016] The obtained rough boundary result frame sample data of the sugarcane field is geocoded, and the corresponding area is cut out from the training sample pair to generate an independent training sample pair dedicated to sugarcane identification.

[0017] Inputting crop fine boundary samples into the improved U-Net model for training to obtain a trained crop classification U-Net model; the crop fine boundary samples include sugarcane crop fine boundary samples and non-sugarcane crop fine boundary samples;

[0018] Data enhancement is performed on the sugarcane identification-specific independent training sample pairs, and the samples are input into a trained crop classification U-Net model for training for multiple rounds to obtain a sugarcane identification U-Net model.

[0019] Preferably, the network structure of the sugarcane recognition U-Net model includes an encoder, a decoder, and a jump connection; the jump connection is used to transfer the output feature map of each layer of the encoder to the corresponding layer of the decoder, so as to gradually restore the boundary of the target in the decoding stage.

[0020] Preferably, the encoder comprises a convolution layer and a pooling layer, each convolution layer comprises two 3x3 convolutions followed by a ReLU activation function to extract spatial features, and the pooling uses a 2x2 maximum pooling.

[0021] Preferably, the decoder comprises an upsampling layer and a convolution layer, the upsampling layer increases the spatial dimension of the feature map and gradually restores it to the size of the input image, and the convolution layer comprises two 3x3 convolutions for refining the features.

[0022] Preferably, before each jump connection, 2x2, 3x3 and 4x4 convolution kernels are used to extract features of different scales from the high-resolution features of the encoder part, and then these features of different scales are spliced ​​together to form a feature map that integrates multi-scale information.

[0023] Preferably, in the process of training the sugarcane recognition U-Net model:

[0024] Apply the Canny edge detection algorithm and binarization processing to the input training sample image to generate a binary true edge image, in which the value of the edge pixel is 1 and the value of the non-edge pixel is 0, denoted as G edge ;

[0025] After the input training sample image is subjected to multi-layer convolution operations, the output segmentation result is obtained. The output result is also subjected to Canny edge detection and binarization processing; a predicted edge map similar to the real edge map is generated, denoted as P edge ;

[0026] The intersection and union of the true edge map and the edge prediction map are calculated to measure their overlap degree; the specific formula of the edge loss function is as follows:

[0027]

[0028] The numerator ∑ i G edge,i P edge,i represents the intersection of the true edge map and the predicted edge map at each pixel, and the denominator ∑ i G edge,i +∑ i P edge,i represents the sum of pixels of two edge maps;

[0029] Finally, L edge With the main segmentation loss function L segCombined with the final loss function for model training, the λ coefficient controls the weight of the boundary loss:

[0030] L total =L seg +λL edge

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The present invention uses sugarcane recognition YOLOv8 to extract the sugarcane result frame and realize the extraction of rough boundaries of sugarcane; and then uses the sugarcane recognition U-Net model to accurately identify the fine boundaries of sugarcane planting areas, thereby realizing rapid and accurate monitoring of the current status and changes of sugarcane, thereby meeting the monitoring needs of sugar crops within the scope of high-standard farmland and providing data support for agricultural protection and management. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A main flow chart of the sugarcane identification method provided by an embodiment of the present invention;

[0034] Figure 2 This is the YOLOv8 network structure diagram;

[0035] Figure 3 This is the bottleneck layer (Bottleneck) structure diagram;

[0036] Figure 4 This is the structure diagram of the Neck hybrid feature network layer;

[0037] Figure 5 This is the network structure diagram of the Head prediction layer;

[0038] Figure 6 The coupling flow chart of the sugarcane identification YOLOv8 model and the sugarcane identification U-Net model;

[0039] Figure 7 This is the improved U-Net network structure diagram;

[0040] Figure 8 This is a working principle diagram of multi-scale information fusion;

[0041] Fig. 9 is the sample distribution;

[0042] Fig.10 A flowchart of the sugarcane identification method of the present invention is provided;

[0043] Fig.11 is the confusion matrix;

[0044] Fig.12 Result of sugarcane fine boundary extraction DETAILED DESCRIPTION

[0045] Example:

[0046] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0047] In the present application, it should be noted that “coarse boundary” and “fine boundary” are relative adjectives and do not represent limitations on specific numerical values.

[0048] The sugarcane identification method provided in this embodiment mainly includes the following steps:

[0049] Acquire a remote sensing image of the area to be identified and input it into a sugarcane identification model, wherein the sugarcane identification model outputs a sugarcane spatial distribution monitoring and identification result;

[0050] The sugarcane recognition model is formed by coupling the sugarcane recognition YOLOv8 model and the sugarcane recognition U-Net model; the sugarcane recognition YOLOv8 model is used to detect and recognize the coarse boundary of sugarcane planting; the sugarcane recognition U-Net model is used to perform a second recognition based on the coarse boundary of sugarcane planting to identify and extract the fine boundary of sugarcane.

[0051] It can be seen that the present invention uses sugarcane recognition YOLOv8 to extract the sugarcane result frame and realize the extraction of rough boundaries of sugarcane; and then uses the sugarcane recognition U-Net model to accurately identify the fine boundaries of sugarcane planting areas, thereby realizing rapid and accurate monitoring of the status quo and changes of sugarcane, thereby meeting the monitoring needs of sugar crops within the scope of high-standard farmland and providing data support for agricultural protection and management.

[0052] In a specific embodiment, the sugarcane recognition YOLOv8 model is trained in the following manner:

[0053] Acquire remote sensing images containing sugarcane, perform geometric correction and orthorectification on the remote sensing images containing sugarcane, extract sample spots and classify them to form a vector sample library.

[0054] Specifically, the sample points were photographed by drones in the field, with an image resolution of 0.5 meters, a total cloud cover of 11.11%, and a spatial reference of the 2000 National Geodetic Coordinate System. The collected remote sensing images were then geometrically corrected and orthorectified in the office, and 11,147 sample patches were extracted using manual interpretation methods. Finally, these samples were merged and divided into 12 categories, namely: vegetable fields, sugarcane, corn, gorgon fruit, abandoned land, forest land, other vegetable fields, other gardens, rice (early stage), rice (mid-stage), rice (late stage), and gardens. A vector sample library was formed for the generation of image sample pairs.

[0055] The classification standards of ground feature samples used in constructing the sample library are as follows: vegetable fields refer specifically to vegetable gardens with large-scale planting and relatively uniform texture on the image; sugarcane, corn, and water chestnut are determined based on actual field survey results; abandoned land is judged based on image texture; woodland refers specifically to woods; other vegetable fields refer to small vegetable gardens in front of and behind houses, or plots where various fruits and vegetables are planted in a disorderly manner; other gardens refer to gardens that do not have granular planting textures on the image, such as pitaya, grapes, and bananas; rice (early stage) refers to paddy fields that have not been planted or have unclear planting features on the image; rice (middle stage) refers to paddy fields that have obvious planting features on the image but are not mature, with obvious gaps between rice; rice (late stage) refers to rice fields that have obvious planting features on the image and are basically mature, with no obvious gaps between rice; gardens are mainly gardens with granular planting features on the image, such as lychee, longan, citrus and other fruit trees.

[0056] The vector samples are divided into respective segments according to the set size to form training sample pairs, and some of the training samples are used as training sample data sets;

[0057] Specifically, using the established vector sample library, based on 0.5-meter high-resolution optical images, the computer vision application function library OpenCV and the file operation module OS were called in Python to build a program for reading, slicing and saving sample images and sample vectors. Before deep learning training, the sample images and sample vectors were segmented in 640*640 sizes according to the usage requirements to form 3320 independent training sample pairs, and 2323 training data sets and 997 validation data sets were divided in a ratio of 7:3 for model training and validation. Since there were too few samples in some categories, five data enhancement methods, namely flipping, translation, rotation, tilting and scaling, were used to expand the training samples during the training process to improve the training effect of the model and improve the generalization ability of the model.

[0058] Finally, the training sample data set is input into the original YOLOv8 model for training to obtain the sugarcane recognition YOLOv8 model; by using the sugarcane recognition YOLOv8 model to detect the sugarcane planting area image, the rough boundary result box sample data of the sugarcane field can be obtained. YOLO is a deep learning target detection model developed by Ultralytics, and can also be used for instance segmentation of images. This embodiment uses the latest YOLOv8 mesoscale model to detect the rough boundaries of sugarcane planting. The total model parameters are 25.9 million, and it accepts 640*640 input images. The YOLOv8 network structure diagram is shown in Figure 2-5: YOLOv8 can be mainly divided into three parts: Backbone neural network, Neck mixed feature network layer and Head prediction layer network.

[0059] The backbone neural network is mainly used to extract the image features of the rough boundary of sugarcane. Its structure includes an attention mechanism module, a cross-stage local network (CSP) and a spatial pyramid pooling structure (SPPF). The attention mechanism module mainly slices the image before entering the backbone neural network, that is, taking a value for every other pixel in an image to obtain 4 complementary images. Finally, the newly generated image is convolved to obtain a 2x downsampled feature map without information loss. YOLOv8 uses the C2f network in the cross-stage local network (CSP) to not only ensure lightweight but also obtain rich gradient flow information. The CBS in the C2f network is the convolution layer, and the bottleneck layer (Bottleneck) uses a 1*1 convolutional neural network. The spatial pyramid pooling structure (SPPF) at the end of the backbone sequentially transmits 3 Maxpools and then connects each layer in series, which not only ensures the accuracy of objects at different scales, but also ensures the lightweight of objects.

[0060] The Neck hybrid feature network layer is composed of convolutional layers and C2f modules. The method for multi-scale feature fusion is path aggregation network (PAD) and feature pyramid network (FPBN), which strengthens the fusion and utilizes the feature layer information at different scales.

[0061] The head prediction layer network adopts a decoupled detection head (Decoupled Head). Its structure takes into account the different content information of the classification category and the positioning position. Therefore, two parallel branches extract the features of the classification category (Cls.Loss) and the features of the positioning position (Bbox.Loss), and then use 1 layer of 1*1 convolution to complete the classification and positioning respectively.

[0062] In a specific embodiment, the sugarcane recognition U-Net model is trained in the following manner:

[0063] The obtained rough boundary result box sample data of the sugarcane field is geocoded, and the corresponding area is cropped from the training sample pair to generate independent training sample pairs dedicated to sugarcane identification. The sugarcane identification U-Net model is specially trained to extract finer boundaries of sugarcane.

[0064] Inputting crop fine boundary samples into the improved U-Net model for training to obtain a trained crop classification U-Net model; the crop fine boundary samples include sugarcane crop fine boundary samples and non-sugarcane crop fine boundary samples;

[0065] Data enhancement is performed on the sugarcane recognition dedicated independent training sample pairs, such as Figure 6 As shown in the figure, the trained crop classification U-Net model is input and trained for 300 more rounds to obtain the sugarcane recognition U-Net model.

[0066] U-Net appeared in the biomedical image segmentation design of Olaf Ronneberger, Philipp Fischer and Thomas Brox in 2015. After the development in recent years, it has also been widely used in the semantic segmentation research of remote sensing. The present invention uses this model to further finely identify the rough boundary samples of sugarcane extracted by the YOLOv8 model and extract the precise boundary of sugarcane. The total number of model parameters is 26772728, and it accepts 640*640 input images. The specific network structure diagram of the improved U-Net model used in this embodiment is shown in Figure 7. The U-Net structural design is derived from the encoder-decoder structure, but compared with the typical encoder-decoder network, U-Net adds skip connections between the two, which can better capture and utilize the detailed information in the image.

[0067] The encoder (Contracting Path) part includes convolutional layers (Conv Layer) and pooling layers (MaxPooling Layer). Each convolutional layer includes two 3x3 convolutions followed by ReLU activation function to extract spatial features. Pooling uses 2x2 maximum pooling to reduce the size of the feature map and increase the receptive field.

[0068] The decoder part (Expansive Path) contains up-sampling layers and convolution layers. The up-sampling layer (Up-Convolution Layer / Transposed Convolution) is used to increase the spatial dimension of the feature map and gradually restore it to the size of the input image. The convolution layer has the same structure as the encoder and contains two 3x3 convolutions for feature refinement.

[0069] The main function of the jump connection is to transfer the output feature map of each layer of the encoder to the corresponding layer of the decoder, so as to gradually restore the boundary of the target in the decoding stage. However, in the jump connection mechanism of the original U-Net, the output of the encoder layer is directly passed to the corresponding layer of the decoder without any processing. This processing method may cause serious problems when segmenting targets of different scales. For scenes that contain both large and small targets in remote sensing images, the directly transmitted feature map may cause small targets to be ignored, while large targets will be under-segmented and their complete contours cannot be accurately extracted. The different sizes and shapes of sugarcane plantations in remote sensing images will cause such problems.

[0070] To solve this problem, the present invention improves the skip connection by introducing a multi-scale convolution strategy. Specifically, before each skip connection, 2x2, 3x3 and 4x4 convolution kernels are used to extract features of different scales from the high-resolution features of the encoder part, and then these features of different scales are spliced ​​together to form a feature map that integrates multi-scale information.

[0071] like Figure 8 As shown in the figure, this multi-scale information fusion design enables the decoder to utilize both local (captured by small convolution kernels) and global (captured by large convolution kernels) information, so that it can better maintain the integrity of the target when processing targets of different resolutions. For large targets, it can ensure the accuracy of integrity and boundaries, while for small targets, it can better restore their details. By introducing local and global contextual information, this method effectively improves the model's ability to capture local details and global information, greatly enhances the ability to restore fine boundaries and small targets in segmentation tasks, and significantly improves the overall accuracy of segmentation.

[0072] The obtained rough boundary samples of sugarcane planting are geocoded, the corresponding areas are cropped in the original independent training sample pairs, and then a U-Net model for crop segmentation is specially constructed to extract finer boundaries of sugarcane.

[0073] In addition, considering that the planting boundary of sugarcane is easily confused with other crops (such as corn, cassava, etc.), when calculating the loss function in the process of training the sugarcane recognition U-Net model, in addition to calculating the conventional cross entropy loss, an additional boundary loss calculation is added to improve the extraction accuracy of the boundary area to ensure that the model can more accurately distinguish the boundary between sugarcane and other crops. The specific steps are as follows:

[0074] First, the Canny edge detection algorithm and binarization processing are applied to the input training image to generate a binary real edge image, in which the value of edge pixels is 1 and the value of non-edge pixels is 0, denoted as G edge .

[0075] During the model training process, the input image is subjected to the multi-layer convolution operation of the model to obtain the output segmentation result, and the output result is also subjected to Canny edge detection and binarization. A predicted edge map similar to the real edge map is generated, denoted as P edge .

[0076] Calculate the intersection and union of the true edge map and the edge prediction map to measure their overlap. The specific formula of the edge loss function is as follows:

[0077]

[0078] The numerator ∑ i G edge,i P edge,i represents the intersection of the true edge map and the predicted edge map at each pixel (the number of overlapping edge pixels), and the denominator ∑ i G edge,i +∑ i P edge,i Represents the pixel sum (union) of two edge maps.

[0079] Finally, L edge With the main segmentation loss function L seg Combined with the final loss function for model training, the λ coefficient controls the weight of the boundary loss:

[0080] L total =L seg +λL edge

[0081] In addition, in order to ensure the accuracy of the sugarcane recognition model formed by coupling the sugarcane recognition YOLOv8 model and the sugarcane recognition U-Net model, the accuracy of the sugarcane recognition model needs to be verified:

[0082] Input 0.5-meter high-resolution images, set the initial learning rate and object weights, use the sugarcane recognition model to perform model prediction and image classification, and obtain classified grids. Convert the obtained classified grids to vectors to obtain the classification results.

[0083] The results are evaluated and analyzed using evaluation indicators such as accuracy, precision, recall, F1 value, intersection over union (IoU) and mean intersection over union (mIoU). The calculation formula is as follows:

[0084] Accuracy:

[0085] Accuracy:

[0086] Recall rate:

[0087] F1 value:

[0088] Intersection ratio:

[0089] Average intersection-union ratio:

[0090] Calculate the evaluation indicators of model training results, analyze the causes and changes in loss functions, optimize parameters, improve model training effects, and increase sugarcane recognition accuracy.

[0091] The method of the present invention is further verified and explained below in conjunction with an application scenario example. Fig.10 As shown:

[0092] The sample points in Lianjiang City and Gaoyao City in the study area were photographed by drones, and the photos taken were then manually interpreted to collect samples. Fig. 9 As shown in the figure, a total of 11,147 samples in 12 categories were collected: vegetable fields, sugarcane, corn, water chestnuts, abandoned land, forest land, other vegetable fields, other gardens, rice (early stage), rice (middle stage), rice (late stage) and gardens.

[0093] The 0.5-meter high-resolution sample images and sample vectors were segmented into 640*640 sizes using a slicing program to form 3320 independent training sample pairs, which were then divided into 2323 training data sets and 997 validation data sets in a ratio of 7:3.

[0094] The training sample data set was put into the YOLOv8 model for training for 350 rounds. During the training process, five data enhancement methods, namely inversion, translation, rotation, tilt, and scaling, were used to expand the training samples to avoid the situation where there were too few classified samples. The trained YOLOv8 model was used to detect the sugarcane planting images and obtain the rough bounding box of the sugarcane field. These result boxes will be used as the input area of ​​the U-Net model.

[0095] The obtained sugarcane rough boundary result frame sample data is geocoded, and the corresponding area is cut out from the independent training sample pair to generate multiple small independent training sample pairs dedicated to sugarcane identification.

[0096] Manually or semi-automatically annotate the fine boundaries of crop areas such as sugarcane in the cropped image samples and generate the corresponding fine boundary masks. Divide the crop fine boundary samples into training sets and validation sets to construct a crop recognition dataset. Then, build a crop fine recognition U-Net model according to the needs, and use the crop fine boundary recognition dataset to train and optimize the model.

[0097] Then, data enhancement was performed on small sample pairs dedicated to sugarcane identification and input into the trained crop classification U-Net model for 300 rounds of training to obtain the sugarcane identification U-Net model to extract the precise boundaries of the crop.

[0098] Input 0.5-meter high-resolution images of the study area, use the sugarcane recognition model formed by coupling the sugarcane recognition YOLOv8 model and the sugarcane recognition U-Net model to perform model prediction and image classification, and obtain classified raster data. Calculate the classification evaluation index, analyze the loss function, and optimize the parameters. Fig.11 As shown in the figure, the final classified raster is converted into a vector, and finally the classification result of sugarcane is obtained to verify the prediction accuracy of the sugarcane recognition model.

[0099] The final performance on the validation data is as follows:

[0100] Table 1 Classification evaluation

[0101]

[0102] From the classification evaluation, it can be seen that the accuracy of gardens, ponds, sugarcane and abandoned land is the best, basically reaching or even exceeding 80% accuracy and recall rate. The sugarcane recognition model has good prediction accuracy.

[0103] Finally, the remote sensing image of the area to be identified is input into the sugarcane identification model. Fig.12 As shown, the spatial distribution monitoring and identification results of sugarcane in the area to be identified can be obtained, the sugarcane boundary can be accurately extracted, and the current status and changes of sugarcane can be quickly and accurately monitored to provide data support for agricultural protection and management.

[0104] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable ordinary technicians in the field to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made based on the essence of the content of the present invention should be included in the protection scope of the present invention.

Claims

1. A sugarcane identification method, characterized in that: include: Acquire a remote sensing image of the area to be identified and input it into a sugarcane identification model, wherein the sugarcane identification model outputs a sugarcane spatial distribution monitoring and identification result; The sugarcane recognition model is formed by coupling the sugarcane recognition YOLOv8 model and the sugarcane recognition U-Net model; the sugarcane recognition YOLOv8 model is used to detect and recognize the coarse boundary of sugarcane planting; the sugarcane recognition U-Net model is used to perform a second recognition based on the coarse boundary of sugarcane planting to identify and extract the fine boundary of sugarcane.

2. The sugarcane identification method according to claim 1, characterized in that: The sugarcane recognition YOLOv8 model is trained in the following way: Acquire remote sensing images containing sugarcane, perform geometric correction and orthorectification on the remote sensing images containing sugarcane, extract sample spots and classify them to form a vector sample library; The vector samples are divided into pairs according to the set size to form training sample pairs, and some training samples are used as training sample data sets. The training sample data set is input into the original YOLOv8 model for training to obtain the sugarcane recognition YOLOv8 model.

3. The sugarcane identification method according to claim 1 or 2, characterized in that: The sugarcane recognition YOLOv8 model is used to detect the sugarcane planting area image to obtain the rough boundary result box sample data of the sugarcane field.

4. The sugarcane identification method according to claim 2, characterized in that: The training sample data set is input into the original YOLOv8 model for training, and five data enhancement methods, namely, inversion, translation, rotation, tilt, and scaling, are used to expand the training samples.

5. The sugarcane identification method according to claim 3, characterized in that: The sugarcane recognition U-Net model is trained in the following way: The obtained rough boundary result frame sample data of the sugarcane field is geocoded, and the corresponding area is cut out from the training sample pair to generate an independent training sample pair dedicated to sugarcane identification; Inputting crop fine boundary samples into the improved U-Net model for training to obtain a trained crop classification U-Net model; the crop fine boundary samples include sugarcane crop fine boundary samples and non-sugarcane crop fine boundary samples; Data enhancement is performed on the sugarcane identification-specific independent training sample pairs, and the samples are input into a trained crop classification U-Net model for training for multiple rounds to obtain a sugarcane identification U-Net model.

6. The sugarcane identification method according to claim 1 or 5, characterized in that: The network structure of the sugarcane recognition U-Net model includes an encoder, a decoder, and a jump connection; the jump connection is used to transfer the output feature map of each layer of the encoder to the corresponding layer of the decoder, so as to gradually restore the boundary of the target in the decoding stage.

7. The sugarcane identification method according to claim 6, characterized in that: The encoder includes a convolution layer and a pooling layer, each convolution layer includes two 3x3 convolutions followed by a ReLU activation function to extract spatial features, and the pooling uses a 2x2 maximum pooling.

8. The sugarcane identification method according to claim 6, characterized in that: The decoder includes an upsampling layer and a convolution layer. The upsampling layer increases the spatial dimension of the feature map and gradually restores it to the size of the input image. The convolution layer contains two 3x3 convolutions for refining the features.

9. The sugarcane identification method according to claim 6, characterized in that: Before each jump connection, 2x2, 3x3, and 4x4 convolution kernels are used to extract features of different scales from the high-resolution features of the encoder part, and then these features of different scales are spliced ​​together to form a feature map that integrates multi-scale information.

10. The sugarcane identification method according to claim 5, characterized in that: In the process of training the sugarcane recognition U-Net model: Apply the Canny edge detection algorithm and binarization processing to the input training sample image to generate a binary true edge image, in which the value of the edge pixel is 1 and the value of the non-edge pixel is 0, denoted as G edge ; After the input training sample image is subjected to multi-layer convolution operations, the output segmentation result is obtained. The output result is also subjected to Canny edge detection and binarization processing; a predicted edge map similar to the real edge map is generated, denoted as P edge ; The intersection and union of the true edge map and the edge prediction map are calculated to measure their overlap degree; the specific formula of the edge loss function is as follows: The numerator ∑ i G edge,i P edge,i Represents the intersection of the true edge map and the predicted edge map at each pixel, and the denominator Σ i G edge,i +Σ i P edge,i represents the sum of pixels of two edge maps; Finally, L edge With the main segmentation loss function L seg Combined with the final loss function for model training, the λ coefficient controls the weight of the boundary loss: THE total =L seg +λL edge 。

Images