A sugarcane recognition method

Abstract

The application discloses a sugarcane recognition method, comprising: acquiring a remote sensing image of a to-be-recognized area and inputting the remote sensing image into a sugarcane recognition model, wherein the sugarcane recognition model outputs a sugarcane spatial distribution monitoring and recognition result; the sugarcane recognition model is coupled by a sugarcane recognition YOLOv8 model and a sugarcane recognition U-Net model; the sugarcane recognition YOLOv8 model is used for detecting and recognizing a sugarcane planting rough boundary; and the sugarcane recognition U-Net model is used for performing second recognition on the basis of the sugarcane planting rough boundary to recognize and extract a sugarcane fine boundary. The sugarcane recognition YOLOv8 is used to extract a sugarcane result frame, rough boundary extraction of the sugarcane is realized, then the sugarcane recognition U-Net model is used to accurately recognize a fine boundary of a sugarcane planting area, rapid and accurate monitoring of a present situation and a change of the sugarcane is realized, and thus 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 present application relates to the field of remote sensing monitoring, in particular to a sugarcane recognition method. BACKGROUND

[0002] Sugarcane is one of the important sugar crops in the world, accounting for about 80% of the global sugar production. The Lingnan region is located in the tropical and subtropical regions, and has the natural advantage of planting sugarcane. Among them, Guangdong has a sowing area of 220.79 million mu in 2022, with a total output of 1292.05 million tons. With the increasing demand for sugarcane at home and abroad, the production of sugarcane has continuously improved the economic benefits, which has promoted the enthusiasm of the Lingnan region for developing economic crops such as sugarcane, so the monitoring of economic crops such as sugarcane plays an important role in improving 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 range of cultivated land in remote sensing satellite images, and then classifying the cultivated land using various classification algorithms. This method changes the traditional operation mode and can quickly extract information in cultivated land. However, due to the fragmented planting structure of cultivated land in the Lingnan region, the climate is cloudy, rainy and foggy, which makes it difficult to identify sugarcane in the Lingnan region using remote sensing technology. The YOLOv8 (You Only Look Once) model is known for its efficient single-step detection method, which can achieve real-time and fast detection of objects. The end-to-end training method simplifies the training and inference process, uses a single network structure for feature extraction and detection, improves the simplicity and efficiency of the model, and can achieve high detection accuracy. However, due to the rough grid division of the YOLOv8 model on the image for detection, the accuracy of small object positioning may be low, and the detection frame may be offset. In the case of dense objects, the detection performance of YOLOv8 may have overlapping frames or missed detection. The U-Net model is good at fine segmentation of ground objects and can accurately segment the boundaries of target objects. It can effectively train small data sets and has good generalization ability. However, the U-Net network requires high hardware conditions and large memory consumption due to the large number of convolution operations and feature map fusion. In addition, the U-Net model relies on accurately labeled data sets, which requires a large amount of work. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art and provide a sugarcane recognition method to achieve fast and accurate monitoring of the current situation and changes of sugarcane.

[0005] To achieve the above purpose, the technical solution of the present application is:

[0006] A sugarcane recognition method, comprising:

[0007] The remote sensing image of the to-be-identified area is acquired and input into a sugarcane identification model, and the sugarcane identification model outputs a sugarcane spatial distribution monitoring and identification result.

[0008] The sugarcane identification model is coupled by a sugarcane identification YOLOv8 model and a sugarcane identification U-Net model; the sugarcane identification YOLOv8 model is used for detecting and identifying a sugarcane planting rough boundary; and the sugarcane identification U-Net model is used for performing a second identification on the basis of the sugarcane planting rough boundary to identify and extract a sugarcane fine boundary.

[0009] Preferably, the sugarcane identification YOLOv8 model is trained by the following method:

[0010] The remote sensing image containing sugarcane is acquired, geometric correction and orthographic correction are performed on the remote sensing image containing sugarcane, sample patches are extracted and classified, and a vector sample library is formed;

[0011] The vector samples are segmented according to a set size to form training sample pairs, and part of the training samples are used as a training sample data set

[0012] The training sample data set is input into an original YOLOv8 model for training to obtain a sugarcane identification YOLOv8 model.

[0013] Preferably, the sugarcane identification YOLOv8 model is used to detect the sugarcane planting area image to obtain a 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 of inversion, translation, rotation, tilt and scaling are used to expand the training samples.

[0015] Preferably, the sugarcane identification U-Net model is trained by the following method:

[0016] The obtained rough boundary result box sample data of the sugarcane field is geocoded, and a corresponding area is cropped from the training sample pair to generate a sugarcane identification special independent training sample pair

[0017] The crop fine boundary sample is input into the improved U-Net model for training to obtain a trained crop classification U-Net model; the crop fine boundary sample includes a sugarcane crop fine boundary sample and a non-sugarcane crop fine boundary sample;

[0018] The sugarcane identification special independent training sample pair is subjected to data enhancement and input into the 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 comprises an encoder, a decoder and a connection jump connection; the jump connection is used to transmit the output feature map of each layer of the encoder to the corresponding layer of the decoder to gradually recover 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 2x2 max pooling.

[0021] Preferably, the decoder comprises an up-sampling layer and a convolution layer, the up-sampling layer increases the spatial dimension of the feature map to gradually recover the size of the input image, and the convolution layer contains two 3x3 convolutions for refining 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 the 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] A Canny edge detection algorithm and a binaryzation process are applied to the input training sample image to generate a binary real edge image, wherein 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 operation, an output segmentation result is obtained, and the output result is also subjected to Canny edge detection and binaryzation process; a predicted edge map similar to the real edge map is generated, denoted as P edge ;

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

[0027]

[0028] Where the numerator ∑ i G edge,i P edge,i represents the intersection of the real edge map and the predicted edge map at each pixel, and the denominator ∑ i G edge,i +∑ i P edge,i represents the total sum of pixels of the two edge maps.

[0029] Finally, L edge is combined with the main segmentation loss function L segThe final loss function of model training is combined with a coefficient lambda to control the weight of boundary loss:

[0030] L total = L seg + lambda L edge

[0031] Compared with the prior art, the present application has the beneficial effects that:

[0032] The present application uses sugarcane recognition YOLOv8 to extract sugarcane result frame, realizes sugarcane rough boundary extraction, and then uses sugarcane recognition U-Net model to accurately identify sugarcane planting land fine boundary, realizes rapid and accurate monitoring of sugarcane status and changes, thereby realizing the monitoring demand of sugar crops in high-standard farmland, and providing data support for agricultural protection management work. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The main flowchart of the sugarcane recognition method provided by the embodiment of the present application is shown in the figure;

[0034] Figure 2 The YOLOv8 network structure diagram is shown in the figure;

[0035] Figure 3 The bottleneck layer (Bottleneck) structure diagram is shown in the figure;

[0036] Figure 4 The Neck mixed feature network layer structure diagram is shown in the figure;

[0037] Figure 5 The Head prediction layer network structure diagram is shown in the figure;

[0038] Figure 6 The coupling flowchart of the sugarcane recognition YOLOv8 model and the sugarcane recognition U-Net model is shown in the figure;

[0039] Figure 7 The improved U-Net network structure diagram is shown in the figure;

[0040] Figure 8 The working principle diagram of multi-scale information fusion is shown in the figure;

[0041] Figure 9 The sample distribution is shown in the figure;

[0042] Figure 10 The flowchart of applying the sugarcane recognition method of the present application is shown in the figure;

[0043] Figure 11 The confusion matrix is shown in the figure;

[0044] Figure 12 The sugarcane fine boundary extraction result is shown in the figure DETAILED DESCRIPTION

[0045] Embodiments:

[0046] The technical solutions of the present application will be further described below in combination with the drawings and embodiments.

[0047] In this application, it should be noted that "coarse boundary" and "fine boundary" are relative adjectives and do not represent a specific numerical size limit.

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

[0049] Obtain the remote sensing image of the to-be-recognized area and input it into the sugarcane recognition model, and the sugarcane recognition model outputs the sugarcane spatial distribution monitoring and recognition result;

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

[0051] As can be seen, the present application uses the sugarcane recognition YOLOv8 to extract the sugarcane result frame, realizes the extraction of the coarse boundary of sugarcane, and then uses the sugarcane recognition U-Net model to accurately recognize the fine boundary of the sugarcane planting area, realizes the rapid and accurate monitoring of the current situation and changes of sugarcane, and thus realizes the monitoring demand of sugar crops in high-standard farmland, and provides data support for agricultural protection and management work.

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

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

[0054] Specifically, the field sample points are photographed by a UAV, the image resolution is 0.5 meters, the total cloud cover ratio is 11.11%, and the spatial reference adopts the 2000 national geodetic coordinate system. Then the collected remote sensing images are geometrically corrected and orthographically corrected, and 11147 sample patches are extracted using the artificial interpretation method. Finally, these samples are merged and divided into 12 categories, which are: vegetable land, sugarcane, corn, gordon euryale, abandoned land, forest land, other vegetable land, other garden land, rice (early stage), rice (middle stage), rice (late stage) and garden land. Form a vector sample library for image sample pair generation.

[0055] The ground object sample division criteria used in constructing the sample library are as follows: vegetable land refers to large-scale planting and uniform texture on the image; sugarcane, corn and gordon euryale are determined according to the actual investigation results; abandoned land is determined according to the image texture; forest land refers to trees; other vegetable land refers to small vegetable gardens in front of houses or land blocks with various types of melons, fruits and vegetables planted in disorder; other garden land refers to garden land without granular planting texture on the image, such as dragon fruit, grape and banana; rice (initial stage) refers to paddy fields without obvious planting characteristics on the image; rice (middle stage) refers to paddy fields with obvious planting characteristics but not mature on the image, with obvious gaps between the rice; rice (late stage) refers to paddy fields with obvious planting characteristics and basically mature rice on the image, without obvious gaps between the rice; garden land mainly refers to garden land with granular planting characteristics on the image, such as lychee, longan and citrus trees.

[0056] The vector samples are respectively segmented according to the set size to form training sample pairs, and part of the training samples are used as a training sample data set;

[0057] Specifically, based on the established vector sample library and 0.5-meter high-resolution optical images, the computer vision application function library OpenCV and the file operation module OS are called in Python to construct a reading, slicing and saving program for sample images and sample vectors. Before deep learning training, the sample images and sample vectors are respectively segmented according to the size of 640*640 according to the use requirements, 3320 independent training sample pairs are formed, and 2323 training data sets and 997 verification data sets are divided according to the proportion of 7:3, which are used for model training and verification. Due to the small amount of some classified samples, five data enhancement methods of flipping, translation, rotation, tilting and scaling are used to expand the training samples in the training process, so as 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 a sugarcane recognition YOLOv8 model. By using the sugarcane recognition YOLOv8 model to detect the sugarcane planting land image, the coarse boundary result frame sample data of the sugarcane land can be obtained. YOLO is a deep learning target detection model developed by Ultralytics Company, which can also be used for instance segmentation of images. In this embodiment, the latest YOLOv8 medium-scale model is used for detection of the coarse boundary of sugarcane planting, and the total parameters of the model are 2590 million, which accepts an input image of 640*640. The YOLOv8 network structure diagram is shown in 2-5: YOLOv8 mainly consists of three parts of Backbone main neural network, Neck mixed feature network layer and Head prediction layer network.

[0059] The backbone main neural network mainly extracts the image features of the rough boundary of sugarcane, and the structure thereof comprises an attention mechanism module, a cross-stage local network (CSP) and a spatial pyramid pooling structure (SPPF). The attention mechanism module mainly performs slicing on the picture before the picture enters the backbone main neural network, that is, one value is taken every other pixel in the picture to obtain four complementary pictures, and finally the newly generated pictures are subjected to convolution operation to obtain a 2-fold down-sampling feature map without information loss. The C2f network in the cross-stage local network (CSP) used by YOLOv8 not only ensures lightweight but also obtains rich gradient flow information, and the CBS in the C2f network is a convolutional 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 passes three Maxpools, and then concatenates each layer, so that the accuracy of objects at different scales is ensured, and the lightweight of objects is ensured.

[0060] The neck mixed feature network layer is composed of a convolutional layer and a C2f module, and the method for fusing multi-scale features is a path aggregation network (PAD) and a feature pyramid network (FPBN), which strengthens fusion and utilizes feature layer information at different scales.

[0061] The head prediction layer network adopts a decoupled head (Decoupled Head), and the structure thereof takes into account that the classification categories and the content information concerned by the positioning location are different, so two parallel branches are used to extract the classification category features (Cls.Loss) and the positioning location features (Bbox.Loss), respectively, and then 1 layer of 1*1 convolution is used to complete classification and positioning.

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

[0063] The obtained rough boundary result frame sample data of the sugarcane field are geocoded, and the corresponding regions are cropped from the training sample pairs to generate independent training sample pairs special for sugarcane recognition, so as to specially train and construct a sugarcane recognition U-Net model to extract more fine boundaries of sugarcane;

[0064] The crop fine boundary sample is input into the improved U-Net model for training to obtain a trained crop classification U-Net model; the crop fine boundary sample comprises a sugarcane crop fine boundary sample and a non-sugarcane crop fine boundary sample;

[0065] The independent training sample pairs special for sugarcane recognition are subjected to data enhancement, as shown in Figure 6 The trained crop classification U-Net model is input into the improved U-Net model for training for 300 rounds to obtain a sugarcane recognition U-Net model.

[0066] U-Net is a biomedical image segmentation design that appeared in 2015 in Olaf Ronneberger, Philipp Fischer and Thomas Brox, and after nearly a few years of development, it has also been widely used in remote sensing influenced semantic segmentation research. The model is used in this application to further fine identification of the cane rough boundary samples extracted by the YOLOv8 model, and to extract the accurate boundary of the cane. The total parameter quantity of the model is 26772728, and the input image is 640*640. The specific network structure diagram of the improved U-Net model used in this embodiment is shown in Figure 7. The U-Net structure 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] Among them, the encoder (Contracting Path) part includes convolution layer (Conv Layer) and pooling layer (MaxPooling Layer), each convolution layer includes two 3x3 convolution, followed by ReLU activation function, to extract spatial features, and the 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 layer and convolution layer, the up-sampling layer (Up-Convolution Layer / Transposed Convolution) is used to increase the spatial dimension of the feature map and gradually restore the size of the input image. While the convolution layer is the same structure as the encoder, containing two 3x3 convolution, used to refine the features.

[0069] The main role of the skip connection is to pass 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 original U-Net skip connection mechanism, the output of the encoder layer is not processed at all, but directly transmitted to the corresponding layer of the decoder. Such processing method may cause serious problems when segmenting targets of different scales. For the scene containing both large and small targets on the remote sensing image, the directly transmitted feature map may cause the small target to be ignored, while the large target may appear under-segmentation, and the complete contour cannot be accurately extracted. The characteristics of the cane plantation on the remote sensing image, which varies in size and shape, will exactly cause such a problem.

[0070] To solve this problem, the application improves the skip connection by introducing a multi-scale convolution strategy. Specifically, before each skip connection, different scale feature extraction is performed on the high-resolution features of the encoder part using convolution kernels of 2x2, 3x3 and 4x4, and then the different scale features are spliced together to form a feature map that integrates multi-scale information.

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

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

[0073] In addition, considering that the planting boundary of sugarcane is easy to be confused with other crops (such as corn, cassava, etc.), therefore, in the process of training the U-Net model for sugarcane recognition, in addition to calculating the regular cross-entropy loss, an edge loss calculation (Boundary Loss) is additionally added to improve the extraction accuracy of the boundary area, to ensure that the model can more accurately distinguish the boundary of sugarcane from other crops. The specific steps are as follows:

[0074] First, apply the Canny edge detection algorithm and binaryzation processing to the input training image to generate a binary real edge image, where the edge pixel value is 1 and the non-edge pixel value is 0, denoted as G edge .

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

[0076] The intersection and union of the real edge map and the edge prediction map are calculated to measure the degree of overlap between them. The edge loss function is specifically as follows:

[0077]

[0078] where ∑ i G edge,i P edge,i denotes the intersection (number of overlapping edge pixels) of the true and predicted edge maps at each pixel, and the denominator ∑ i G edge,i +∑ i P edge,i denotes the sum of pixels of both edge maps (union).

[0079] Finally, L edge is combined with the main segmentation loss function L seg as the final loss function for model training, with the weight of the boundary loss controlled by the coefficient λ:

[0080] L total = L seg + λL edge

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

[0082] Input 0.5-meter high-resolution images, set the initial learning rate and object weight, use the sugarcane recognition model for model prediction and image classification, and obtain the classified raster. Convert the obtained classified raster to vector, and finally obtain the classification result.

[0083] Use evaluation indicators such as accuracy (Accuracy), precision (Precision), recall (Recall), F1 value, intersection over union (Intersection over Union, IoU), and mean intersection over union (mIoU) to evaluate and analyze the results. The calculation formulas are as follows:

[0084] Accuracy:

[0085] Precision:

[0086] Recall:

[0087] F1 value:

[0088] Intersection over Union:

[0089] Mean Intersection over Union:

[0090] The calculation model training result evaluation index is analyzed, reasons and loss function changes are analyzed, parameters are optimized, model training effect is improved, and sugarcane recognition precision is increased.

[0091] The method of the application will be further verified and described below in combination with an application scenario example, as shown in Figure 10

[0092] The photos taken by the unmanned aerial vehicle are collected by manual interpretation in the indoor work, as shown in Figure 9

[0093] The 0.5-meter high-resolution sample image and sample vector are segmented by a slicing program according to the size of 640*640 to form 3320 independent training sample pairs, and 2323 training data sets and 997 verification data sets are divided according to the ratio of 7:3.

[0094] The training sample data set is put into the YOLOv8 model for training for 350 rounds, and five data enhancement methods of inversion, translation, rotation, tilt and scaling are used to expand the training samples during the training process to avoid the situation that the classification sample is too few. The trained YOLOv8 model is used to detect the sugarcane planting image to obtain the coarse boundary result box of the sugarcane field. These result boxes will be used as the input area of the U-Net model.

[0095] The obtained coarse boundary result box sample data of the sugarcane is geocoded, and the corresponding area is cropped from the independent training sample pair to generate multiple small independent training sample pairs for sugarcane recognition.

[0096] The fine boundary of the crop area of the cropped image sample is manually labeled or labeled using a semi-automatic tool to generate a corresponding fine boundary mask. The crop fine boundary sample pair is divided into a training set and a verification set to construct a crop recognition data set. Then, according to the requirements, a crop fine recognition U-Net model is constructed, and the model is trained and optimized using the crop fine boundary recognition data set.

[0097] Then, the small sample pair for sugarcane recognition is data-augmented, and the crop classification U-Net model that has been trained is input into the model for training for 300 rounds to obtain a sugarcane recognition U-Net model to extract the accurate boundary of the crop.

[0098] ​​Input the 0.5-meter high-resolution image of the study area, use the sugarcane recognition model coupled by the sugarcane recognition YOLOv8 model and the sugarcane recognition U-Net model to perform model prediction and image classification, and obtain the classified raster data. Calculate the classification evaluation index, analyze the loss function, and optimize the parameters. As shown in Figure 11 , the final classified raster is converted into a vector to obtain the classification result of the sugarcane, so as 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 garden land, pit pond water surface, sugarcane and abandoned land is the best, basically reaching or even exceeding 80% accuracy and recall rate, and the sugarcane recognition model has good prediction accuracy.

[0103] Finally, input the remote sensing image of the region to be recognized into the sugarcane recognition model, as shown in Figure 12 , the spatial distribution monitoring and recognition result of the sugarcane in the region to be recognized can be obtained, the boundary of the sugarcane can be accurately extracted, the rapid and accurate monitoring of the current situation and changes of the sugarcane can be realized, and data support can be provided for agricultural protection and management.

[0104] The above examples are only for illustrating the technical concept and characteristics of the present application, the purpose is to enable those skilled in the art to understand the content of the present application and to implement it, and cannot limit the protection scope of the present application. Any equivalent changes or modifications made according to the essence of the present application should be covered within the protection scope of the present application.

Claims

1. A method of sugarcane identification, characterized by, include: Remote sensing images of the area to be identified are acquired and input into the sugarcane identification model, which outputs the spatial distribution monitoring and identification results of sugarcane. The sugarcane recognition model is composed of a sugarcane recognition YOLOv8 model and a 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. The sugarcane recognition U-Net model was trained in the following manner: The coarse boundary result box sample data of the obtained sugarcane field is geocoded, and the corresponding area is cropped from the training sample pair to generate independent training sample pairs for sugarcane identification. The crop fine boundary samples are input 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 augmentation is performed on the independent training sample pairs specifically for sugarcane identification, and the data is then input into the pre-trained crop classification U-Net model for multiple training rounds to obtain the sugarcane identification U-Net model. The network structure of the sugarcane recognition U-Net model includes an encoder, a decoder, and skip connections; the skip connections are used to pass the output feature map of each layer of the encoder to the corresponding layer of the decoder, so as to gradually recover the boundary of the target during the decoding stage. Before each skip connection, 2x2, 3x3 and 4x4 convolutional kernels are used to extract features at different scales from the high-resolution features of the encoder part. These features at different scales are then stitched together to form a feature map that integrates multi-scale information.

2. The method of claim 1, wherein, The sugarcane recognition YOLOv8 model was trained in the following way: Acquire remote sensing images containing sugarcane, perform geometric and orthorectification on the remote sensing images containing sugarcane, extract sample patches and classify them to form a vector sample library; The vector samples are divided according to the set size to form training sample pairs, and a portion of the training samples are used as the training sample dataset. The training sample dataset is input into the original YOLOv8 model for training to obtain the sugarcane recognition YOLOv8 model.

3. The method of claim 1 or 2, wherein, The sugarcane recognition YOLOv8 model was used to detect images of sugarcane plantations, and sample data of coarse boundary result boxes of sugarcane fields were obtained.

4. The method of claim 2, wherein the step of identifying the sugarcane plant is performed by using the image recognition method. The training sample dataset is input into the original YOLOv8 model for training. Five data augmentation methods, namely inversion, translation, rotation, skew, and scaling, are used to expand the training samples.

5. The method of claim 1, wherein, The encoder includes convolutional layers and pooling layers. Each convolutional layer includes two 3x3 convolutions followed by a ReLU activation function to extract spatial features. The pooling uses 2x2 max pooling.

6. The method of claim 1, wherein, The decoder includes an upsampling layer and a convolutional layer. The upsampling layer increases the spatial dimension of the feature map, gradually restoring it to the size of the input image. The convolutional layer contains two 3x3 convolutions to refine the features.

7. The method of claim 1, wherein, During the training of the sugarcane recognition U-Net model: A Canny edge detection algorithm and binarization processing are applied to the input training sample image to generate a binary real edge image, where the value of an edge pixel is 1 and the value of a non-edge pixel is 0, denoted as ; The input training sample image is subjected to multi-layer convolution operation to obtain an output segmentation result. The output result is subjected to Canny edge detection and binary processing to generate a predicted edge map similar to the real edge map, denoted as ; The intersection and union of the real edge map and the edge prediction map are calculated to measure the degree of their overlap; the edge loss function is specifically as follows: wherein the molecule denotes the intersection of the real edge map and the predicted edge map at each pixel, the denominator denotes the sum of the pixels of both edge maps; Finally With the main segmentation loss function Combined with the final loss function used for model training, The coefficients control the weights of the boundary loss: 。

Images