Intelligent Identification Method of Rhododendron lapponicum Based on UAV Sensors

Abstract

The present invention discloses an intelligent recognition method for Rhododendron lapponicum based on drone sensors, which relates to the technical field of image recognition. The method includes: obtaining an image sample set and an identification sample set of the Rhododendron lapponicum planting area, performing identification modality classification to obtain Q Rhododendron identification modality areas; inputting the drone remote sensing image sample set into the Rhododendron lapponicum detection and segmentation module to obtain the Rhododendron lapponicum image sample set; introducing a Rhododendron identification learning channel to generate a Rhododendron Q-modality identification unit; performing generalization enhancement learning on the Rhododendron Q-modality identification unit; obtaining a remote sensing acquisition image of the Rhododendron lapponicum planting area and generating a Rhododendron Q-modality identification result. The present invention solves the technical problems in the prior art that the accuracy of identifying the Rhododendron lapponicum planting area is poor, the generalization ability is insufficient, and it is difficult to adapt to multiple identification modalities, and achieves the technical effect of realizing the accurate identification of the Rhododendron lapponicum planting area and improving the accuracy and generalization ability of Rhododendron lapponicum identification.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and particularly to an intelligent recognition method for Rhododendron lapponicum based on an unmanned aerial vehicle (UAV) sensor. Background Art

[0002] In the current fields of ecological protection and agricultural research, the accurate recognition and monitoring of Rhododendron lapponicum planting areas are of crucial significance. As a rare plant resource, Rhododendron lapponicum not only plays an indispensable role in the ecosystem, but its planting situation also has a profound impact on the maintenance of ecological balance and the development of the agricultural industry. However, there are many problems to be solved urgently in the existing technologies for identifying Rhododendron lapponicum planting areas. On the one hand, traditional recognition methods often rely on manual observation. This method is not only inefficient but also easily affected by the subjective factors of observers, resulting in a significant reduction in the accuracy of recognition results and making it difficult to meet the requirements of large-scale and high-precision recognition. On the other hand, some existing automated recognition technologies have serious deficiencies in generalization ability when facing complex and changeable natural environments and different recognition modalities, and cannot effectively adapt to the Rhododendron lapponicum recognition tasks in different scenarios.

[0003] The existing technologies have technical problems such as poor accuracy, insufficient generalization ability, and difficulty in adapting to multiple recognition modalities in the recognition of Rhododendron lapponicum planting areas. Summary of the Invention

[0004] The present application provides an intelligent recognition method for Rhododendron lapponicum based on a UAV sensor, which is used to solve the technical problems of poor accuracy, insufficient generalization ability, and difficulty in adapting to multiple recognition modalities in the recognition of Rhododendron lapponicum planting areas in the existing technologies.

[0005] In view of the above problems, the present application provides an intelligent recognition method for Rhododendron lapponicum based on a UAV sensor, and the method includes:

[0006] Obtain a drone remote sensing image sample set of the Rhododendron lapponicum planting area and a Rhododendron lapponicum recognition sample set, and perform recognition modality classification on the Rhododendron lapponicum recognition sample set to obtain Q Rhododendron recognition modality regions, where Q is a positive integer greater than 1; input the drone remote sensing image sample set into the Rhododendron lapponicum detection and segmentation module to obtain a Rhododendron lapponicum image sample set; introduce a Rhododendron recognition learning channel, perform loss optimization learning on the Rhododendron lapponicum image sample set and the Q Rhododendron recognition modality regions to generate a Rhododendron Q-modality recognition unit; introduce a GAN network and a predetermined remote sensing scene distribution to perform generalization enhancement learning on the Rhododendron Q-modality recognition unit to obtain a Rhododendron modality recognition integration module, and combine the Rhododendron lapponicum detection and segmentation module to build a Rhododendron lapponicum recognition module; obtain a remote sensing acquisition image of the Rhododendron lapponicum planting area according to the drone sensor, input the remote sensing acquisition image into the Rhododendron lapponicum recognition module, and generate a Rhododendron Q-modality recognition result.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] Obtain a drone remote sensing image sample set of the Rhododendron lapponicum planting area and a Rhododendron lapponicum recognition sample set, and perform recognition modality classification on the Rhododendron lapponicum recognition sample set to obtain Q Rhododendron recognition modality regions; input the drone remote sensing image sample set into the Rhododendron lapponicum detection and segmentation module to obtain a Rhododendron lapponicum image sample set; perform loss optimization learning on the Rhododendron lapponicum image sample set and the Q Rhododendron recognition modality regions to generate a Rhododendron Q-modality recognition unit; perform generalization enhancement learning to obtain a Rhododendron modality recognition integration module, and build a Rhododendron lapponicum recognition module; obtain a remote sensing acquisition image of the Rhododendron lapponicum planting area according to the drone sensor, input it into the Rhododendron lapponicum recognition module, and generate a Rhododendron Q-modality recognition result. It achieves the technical effect of realizing the accurate recognition of the Rhododendron lapponicum planting area and improving the accuracy and generalization ability of Rhododendron lapponicum recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0010] Figure 1 It is a schematic flowchart of the intelligent Rhododendron lapponicum recognition method based on a drone sensor provided in the embodiment of this application;

[0011] Figure 2 It is a schematic flowchart of generating a Rhododendron Q-modality recognition unit in the intelligent Rhododendron lapponicum recognition method based on a drone sensor provided in the embodiment of this application. Specific implementation manner

[0012] This application provides an intelligent recognition method for Rhododendron lapponicum based on drone sensors, aiming to solve the technical problems of poor accuracy, insufficient generalization ability, and difficulty in adapting to multiple recognition modalities in the identification of Rhododendron lapponicum planting areas in the prior art.

[0013] Next, the technical solutions in the embodiments of this application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without creative efforts shall fall within the protection scope of this application.

[0014] Embodiment, as Figure 1 shown, this application provides an intelligent recognition method for Rhododendron lapponicum based on drone sensors, and the method includes:

[0015] Step S100: Obtain a drone remote sensing image sample set of the Rhododendron lapponicum planting area and a Rhododendron lapponicum recognition sample set, and perform recognition modality classification on the Rhododendron lapponicum recognition sample set to obtain Q Rhododendron recognition modality areas, where Q is a positive integer greater than 1.

[0016] Specifically, first, with the help of a drone equipped with a multispectral camera, according to the topography, area size, and distribution of the Rhododendron lapponicum planting area, plan a suitable flight altitude and route to obtain comprehensive and high-resolution multispectral images, forming a drone remote sensing image sample set. At the same time, details such as the variety, growth stage, leaf characteristics, and flowering status of Rhododendron lapponicum are recorded to construct a Rhododendron lapponicum recognition sample set. Then, the recognition sample set is carefully classified from dimensions such as variety differences, growth processes, leaf morphologies, and flowering degrees, thereby obtaining multiple (Q, Q>1) Rhododendron recognition modality areas with different characteristics. These modality areas cover various characteristics of Rhododendron lapponicum, laying a solid data foundation for the subsequent accurate recognition of Rhododendron lapponicum, and helping to improve the efficiency and accuracy of the Rhododendron lapponicum recognition work.

[0017] Step S200: Input the drone remote sensing image sample set into the Rhododendron lapponicum detection and segmentation module to obtain a Rhododendron lapponicum image sample set.

[0018] Specifically, after obtaining the UAV remote sensing image sample set, this image sample set is input into the Rhododendron lapponicum detection and segmentation module, which consists of a preprocessing unit, a rhododendron pixel detection unit, and a rhododendron background separation unit working together. First, the image is processed successively through a radiation correction model, a geometric correction model, and a Gaussian filter in the preprocessing unit to correct radiation errors, geometric deformations, and remove noise, and is transformed into a standard remote sensing image sample set. Then, images are extracted one by one from the standard sample set. Taking one of them as an example, it is input into the rhododendron pixel detection unit. The rhododendron pixel prediction model and the rhododendron pixel discrimination model in this unit cooperate. First, the prediction model processes the pixel samples in the image to obtain the probability coefficient of each pixel belonging to the Rhododendron lapponicum pixel, and then the discrimination model combines the predetermined coefficient to judge these prediction coefficients, and further determines the rhododendron pixel distribution. Finally, the rhododendron background separation unit performs background separation processing on the standard remote sensing image based on this pixel distribution, and successfully obtains a clear Rhododendron lapponicum image sample set.

[0019] Step S300: Introduce a rhododendron recognition learning channel, perform loss optimization learning on the Rhododendron lapponicum image sample set and the Q rhododendron recognition modal regions, and generate a rhododendron Q-modal recognition unit.

[0020] Specifically, after completing the acquisition of the Rhododendron lapponicum image sample set, a rhododendron recognition learning channel is introduced. This channel includes P rhododendron recognition learners and a rhododendron recognition loss calculator (P is a positive integer greater than 1). For the Q rhododendron recognition modal regions (such as the modal regions formed by classifying Rhododendron lapponicum varieties, growth stages, etc.), each time the q-th modal region (1 ≤ q ≤ Q) is extracted from them. Based on this modal region and the Rhododendron lapponicum image sample set as the basic data, the rhododendron recognition loss calculator is used to supervise and train the P rhododendron recognition learners. During the training process, every time a predetermined number of training times is reached, P rhododendron recognition loss coefficients are calculated according to a specific rhododendron recognition loss calculation formula. If these loss coefficients are less than the set rhododendron recognition loss threshold, P same-modal rhododendron recognition models are generated, and they are connected as parallel nodes to form the q-th modal rhododendron recognition model. Then, taking the output data set of the q-th modal rhododendron recognition model as the input and the q-th rhododendron recognition modal region as the output, the q-th modal rhododendron recognition fusion model is trained. Finally, the input layer of the q-th modal rhododendron recognition model is merged with the fusion model to generate the q-th modal rhododendron recognition unit. The above process is continuously repeated, and the generated q-th modal recognition units are summarized to finally form a rhododendron Q-modal recognition unit, thereby optimizing the model's recognition ability for Rhododendron lapponicum under different modalities.

[0021] Step S400: Introduce a GAN network and a predetermined remote sensing scene distribution to perform generalization enhancement learning on the rhododendron Q-modal recognition unit to obtain a rhododendron modal recognition integration module, and combine it with the Rhododendron lapponicum detection and segmentation module to build a Rhododendron lapponicum recognition module.

[0022] Specifically, after generating the rhododendron Q-modal recognition unit, to improve the generalization ability of the model, a GAN network and a predetermined remote sensing scene distribution are introduced. First, by collecting the remote sensing scene information of the rhododendron image sample set, the remote sensing scene distribution ratio is calculated and compared with the predetermined remote sensing scene distribution ratio. Taking the gap between the two as the optimization target, the remote sensing scene sample set is expanded. Based on the expanded sample set, the GAN network is used to expand the data of the rhododendron image sample set to obtain an expanded rhododendron image sample set. Then, according to the internal relationship between the rhododendron image sample set and the Q rhododendron recognition modal regions, a rhododendron image-recognition modal mapping network is established. With the help of this network, the expanded rhododendron image sample set is migrated and labeled according to the Q rhododendron recognition modal regions to generate an expanded rhododendron recognition modal sample set. Then, the expanded rhododendron image sample set and the expanded rhododendron recognition modal sample set are used for generalization enhancement learning of the rhododendron Q-modal recognition unit, thereby obtaining a rhododendron modal recognition integration module. Finally, the rhododendron modal recognition integration module is connected to the previous rhododendron detection and segmentation module, successfully building a rhododendron recognition module, enabling the module to better adapt to different remote sensing scenes and improving the recognition accuracy and stability of rhododendrons.

[0023] Step S500: According to the drone sensor, obtain the remote sensing acquisition image of the rhododendron planting area, and input the remote sensing acquisition image into the rhododendron recognition module to generate a rhododendron Q-modal recognition result.

[0024] Specifically, use a drone equipped with a specific sensor to collect data from the rhododendron planting area again. The drone collects multi-spectral remote sensing images over the planting area according to the established flight route and parameters, obtaining remote sensing acquisition images containing rich information. Subsequently, these images are input into the previously carefully built rhododendron recognition module. Each component inside the module works together to analyze and process the input images. Through deep learning algorithms and the previously trained model parameters, the rhododendrons in the images are recognized and classified, and finally a rhododendron Q-modal recognition result is generated, which can detail various aspects of the rhododendrons in the image, such as variety, growth stage, leaf characteristics, flowering status, etc., providing key data support for the research, protection, and monitoring of rhododendrons.

[0025] In a possible implementation manner, as Figure 2 shown, step S300 further includes:

[0026] Step S310: The rhododendron recognition learning channel includes P rhododendron recognition learners and a rhododendron recognition loss calculator, where P is a positive integer greater than 1.

[0027] Step S320: Extract the q-th rhododendron recognition modal area according to the Q rhododendron recognition modal areas, where q is a positive integer and 1 ≤ q ≤ Q.

[0028] Step S330: Based on the rhododendron recognition loss calculator, perform loss optimization learning on the P rhododendron recognition learners according to the alpine rhododendron image sample set and the q-th rhododendron recognition modal area to generate the q-th modal rhododendron recognition model.

[0029] Step S340: Use the output data set of the q-th modal rhododendron recognition model as input information and the q-th rhododendron recognition modal area as output information to train the q-th modal rhododendron recognition fusion model.

[0030] Step S350: Merge the input layers of the q-th modal rhododendron recognition model and the q-th modal rhododendron recognition fusion model to generate the q-th modal rhododendron recognition unit, and add the q-th modal rhododendron recognition unit to the Q-modal rhododendron recognition unit.

[0031] Specifically, the rhododendron recognition learning channel is mainly composed of P rhododendron recognition learners and a rhododendron recognition loss calculator, where P represents a positive integer greater than 1. These P rhododendron recognition learners are all machine learning models, and they possess powerful data learning and feature extraction capabilities. Different learners can analyze and learn the image data and recognition modal area data of alpine rhododendrons from different perspectives. Some learners are good at capturing the texture features of images, while others are more sensitive to spectral features. The rhododendron recognition loss calculator is used to measure the deviation in the model learning process. In the subsequent learning process, it calculates the loss values of each learner during the learning process to evaluate the fitting degree and recognition effect of the learner on the data, providing a quantitative basis for optimizing the parameters of the learner, so that the entire learning channel can efficiently learn and process data related to alpine rhododendrons, laying a solid foundation for generating a more accurate recognition model.

[0032] After completing the modal classification of the alpine rhododendron recognition sample set and obtaining Q rhododendron recognition modal areas, specific modal areas are sequentially selected for subsequent processing in order. Among them, q represents a positive integer, and its value range is limited between 1 and Q, that is, 1 ≤ q ≤ Q. In actual operation, the q-th rhododendron recognition modal area is accurately extracted from the Q rhododendron recognition modal areas. For example, if Q is 5 and q is 3, the 3rd modal area among these 5 modal areas is extracted. This extracted q-th rhododendron recognition modal area will serve as the key data input for subsequent model training and optimization, providing basic data support for constructing a more targeted q-th modal rhododendron recognition model, thereby enhancing the recognition ability of alpine rhododendrons under different modalities.

[0033] The cuckoo recognition loss calculator conducts supervised training on P cuckoo recognition learners. During the training process, every time a predetermined number of training iterations is completed, the P cuckoo recognition loss coefficients are calculated once. This coefficient is obtained through the cuckoo recognition loss calculation formula, which compares the number of cuckoo recognition prediction samples and the number of correctly predicted samples in each training, thereby measuring the deviation degree between the model prediction and the actual situation. As the training continues, each learner continuously adjusts its own parameters according to the loss coefficient, searching for the optimal parameter combination that minimizes the loss. This is the process of loss optimization learning. When all P cuckoo recognition loss coefficients are less than the pre-set cuckoo recognition loss threshold, it indicates that the training of the learner has reached a certain accuracy requirement. At this time, these P well-performing cuckoo recognition models of the same modality are connected as parallel nodes, and then the cuckoo recognition model of the q-th modality is generated, which can more effectively identify the Rhododendron lapponicum belonging to the q-th modality.

[0034] Take the output data set of the cuckoo recognition model of the q-th modality as the new input information because this output data set already contains the results of the model's preliminary extraction and analysis of the image features of Rhododendron lapponicum. At the same time, take the q-th cuckoo recognition modality area as the expected output information, and this modality area covers the accurate features and classification information of Rhododendron lapponicum in a specific modality. Using these two sets of data, the backpropagation algorithm is used to train the cuckoo recognition fusion model of the q-th modality. During the training process, the fusion model continuously adjusts its own network parameters, attempting to minimize the difference between the predicted output and the actual q-th cuckoo recognition modality area. Through multiple iterative trainings, the fusion model gradually learns how to better integrate and process the output information of the cuckoo recognition model of the q-th modality to more accurately match the q-th cuckoo recognition modality area, thereby improving the recognition accuracy and stability of Rhododendron lapponicum of the q-th modality.

[0035] After completing the training of the cuckoo recognition model of the q-th modality and the cuckoo recognition fusion model of the q-th modality, a merging operation is performed on the input layers of the cuckoo recognition model of the q-th modality and the cuckoo recognition fusion model of the q-th modality. Through the neural network structure fusion algorithm, the neuron connection relationships of the input layers of the two models are re-integrated, so that the newly generated structure can simultaneously receive the feature information of the input data from both models, and then form a more powerful cuckoo recognition unit of the q-th modality. This new unit combines the ability of the cuckoo recognition model of the q-th modality to initially extract features and the advantages of the cuckoo recognition fusion model of the q-th modality in optimizing and accurately matching features. After generating the cuckoo recognition unit of the q-th modality, according to the established data storage and management rules, it is added to the cuckoo Q-modality recognition unit. As this operation is performed for each q value (1 ≤ q ≤ Q), the cuckoo Q-modality recognition unit is continuously improved, and it covers the ability to recognize Rhododendron lapponicum for different modalities, providing the core technical support for accurately recognizing Rhododendron lapponicum subsequently.

[0036] In a possible implementation manner, step S330 further includes:

[0037] Step S331: According to the rhododendron image sample set and the q-th rhododendron recognition modality region, respectively perform supervised training on the P rhododendron recognition learners. Every time a predetermined number of training times is completed, according to the rhododendron recognition loss calculator, calculate P rhododendron recognition loss coefficients, where the rhododendron recognition loss calculator includes a rhododendron recognition loss calculation formula, and the rhododendron recognition loss calculation formula is:

[0038] ;

[0039] Among them, LOSS represents the rhododendron recognition loss coefficient, M represents the predetermined number of training times, m represents the m-th training, both M and m are positive integers, 1 ≤ m ≤ M, SUOm represents the number of rhododendron recognition prediction samples in the m-th training, and SUXm represents the number of correctly predicted rhododendron recognition samples in the m-th training.

[0040] Step S332: If the P rhododendron recognition loss coefficients are less than the rhododendron recognition loss threshold, generate P same-modality rhododendron recognition models.

[0041] Step S333: Connect the P same-modality rhododendron recognition models as parallel nodes to generate the q-th modality rhododendron recognition model.

[0042] Specifically, for the supervised training of the P rhododendron recognition learners. Taking the common deep learning framework TensorFlow as an example, first perform data preprocessing, convert the data of the rhododendron image sample set and the q-th rhododendron recognition modality region into a format acceptable to the model, perform normalization processing on the image sample set to make the pixel values between 0 and 1, and at the same time encode the data of the q-th rhododendron recognition modality region to make it correspond to the image data. During the training process, set a suitable predetermined number of training times, assumed to be 50 times. Each time during training, input the data in batches into the P rhododendron recognition learners, and each learner tries to recognize the rhododendron image samples and predict their belonging modality categories. After each training is completed, relevant data will be recorded, including the number of rhododendron recognition prediction samples SUOm and the number of correctly predicted samples SUXm in the m-th training. When the training reaches the predetermined number of training times, the rhododendron recognition loss calculator starts to play a role. It is based on the given rhododendron recognition loss calculation formula

[0043] , calculate the loss coefficients of P cuckoo recognition learners respectively. In this formula, M represents the predetermined number of training times (here it is 50), m represents the current training time (1 ≤ m ≤ M), and LOSS is the calculated cuckoo recognition loss coefficient, which reflects the deviation degree between the model prediction result and the actual situation. After the calculation, compare the P cuckoo recognition loss coefficients obtained with the pre-set cuckoo recognition loss threshold. If all the loss coefficients are less than the threshold, it means that the P cuckoo recognition learners perform well under the current training, and the learning of the alpine cuckoo images and the q-th cuckoo recognition modality area reaches a certain accuracy requirement. At this time, save the models corresponding to these P excellent learners to generate P cuckoo recognition models of the same modality. Finally, in order to further improve the recognition ability of the model, connect these P cuckoo recognition models of the same modality as parallel nodes. In actual operation, use the functions or tools in TensorFlow to reasonably connect the input layers and output layers of these models so that they can process and analyze the input data simultaneously. After such processing, a cuckoo recognition model of the q-th modality with stronger comprehensive performance is generated. It integrates the advantages of P cuckoo recognition models of the same modality and can more accurately identify the alpine cuckoos belonging to the q-th modality, providing key support for subsequent recognition work.

[0044] In a possible implementation manner, step S400 further includes:

[0045] Step S410: Based on the GAN network, expand the alpine cuckoo image sample set according to the predetermined remote sensing scene distribution to obtain an expanded cuckoo image sample set.

[0046] Step S420: Based on the mapping relationship between the alpine cuckoo image sample set and the Q cuckoo recognition modality areas, establish a cuckoo image - recognition modality mapping network.

[0047] Step S430: Based on the cuckoo image - recognition modality mapping network, perform transfer annotation on the expanded cuckoo image sample set according to the Q cuckoo recognition modality areas to generate an expanded cuckoo recognition modality sample set.

[0048] Step S440: Perform generalization enhancement learning on the cuckoo Q - modality recognition unit according to the expanded cuckoo image sample set and the expanded cuckoo recognition modality sample set to generate the cuckoo modality recognition integration module.

[0049] Step S450: Connect the cuckoo modality recognition integration module and the alpine cuckoo detection and segmentation module to obtain the alpine cuckoo recognition module.

[0050] Specifically, a generative adversarial network (GAN network) is used to expand the data. The predetermined remote sensing scene distribution ratio is used as a reference and input into the GAN network. The GAN network consists of a generator and a discriminator. The generator is responsible for generating new simulated Rhododendron lapponicum image data based on the predetermined remote sensing scene distribution ratio and the characteristics of the Rhododendron lapponicum image sample set. The discriminator then discriminates the generated image data to determine whether it is real. Through the continuous game between the generator and the discriminator, the data generated by the generator becomes closer and closer to the real Rhododendron lapponicum image, thereby realizing the data expansion of the Rhododendron lapponicum image sample set. Finally, a large number of expanded Rhododendron lapponicum image sample sets are obtained, enriching the diversity and quantity of image data.

[0051] Establishing a Rhododendron lapponicum image-recognition modality mapping network is a key link for laying the foundation for subsequent precise annotation and recognition work. First, deep feature extraction is performed on the Rhododendron lapponicum image sample set. Using a convolutional neural network (CNN), with the help of its powerful convolutional layer and pooling layer structures, multi-scale and multi-level feature extraction is performed on the image samples. The convolutional layer slides different convolutional kernels on the image to extract local features in the image, such as edges and textures. The pooling layer then performs dimensionality reduction on the output of the convolutional layer, reducing the amount of data while retaining key features. After multiple convolutional and pooling operations, a feature vector that can effectively represent the Rhododendron lapponicum image is obtained. Next, the data in Q Rhododendron lapponicum recognition modality regions is encoded. According to different modality categories, such as Rhododendron lapponicum varieties, growth stages, leaf characteristics, flowering states, etc., the data in each modality region is converted into a numerical vector representation. For example, for Rhododendron lapponicum varieties, one-hot encoding can be used to represent each variety with a unique binary vector. For continuous data such as growth stages and leaf characteristics, normalization is performed to make their values within a certain reasonable range. Then, correlation analysis is performed between the extracted Rhododendron lapponicum image feature vector and the encoded vectors in Q Rhododendron lapponicum recognition modality regions. Using a fully connected neural network (FCN) in deep learning, with the Rhododendron lapponicum image feature vector as the input and the encoded vectors in Q Rhododendron lapponicum recognition modality regions as the output, a mapping model is constructed. During the training process, the backpropagation algorithm is used to continuously adjust the weights and biases of the fully connected layer, enabling the model to learn the internal mapping relationship between the Rhododendron lapponicum image features and the recognition modality regions. Through multiple iterative trainings, the error between the prediction result of the model and the actual encoded vectors in the recognition modality regions gradually decreases until the predetermined accuracy requirement is met, thus successfully establishing a Rhododendron lapponicum image-recognition modality mapping network. This network can accurately map the features of the Rhododendron lapponicum image to the corresponding recognition modality regions, providing strong support for subsequent transfer annotation work.

[0052] With the established mapping network of rhododendron image-recognition modality, transfer annotation is performed on the expanded rhododendron image sample set. Each image in the expanded rhododendron image sample set is input into the mapping network. According to the mapping relationship learned by the network and combined with the information of Q rhododendron recognition modality regions, these expanded images are annotated to determine their corresponding rhododendron recognition modalities. For example, for a newly generated Rhododendron lapponicum image, through the analysis of the mapping network, it is determined that it belongs to a certain specific variety of Rhododendron lapponicum and is in a specific growth stage, etc., thus generating an expanded rhododendron recognition modality sample set, providing richer and more targeted annotation data for subsequent model training.

[0053] Use the expanded rhododendron image sample set and the expanded rhododendron recognition modality sample set to perform generalization enhancement learning on the rhododendron Q-modality recognition unit. Input these two sample sets into the rhododendron Q-modality recognition unit. By continuously adjusting the parameters of the model, the model can learn more characteristics of Rhododendron lapponicum and recognition modality information in different scenarios, improving the adaptability and generalization ability of the model to different situations. During the training process, optimization algorithms such as gradient descent are used to continuously reduce the error between the model prediction result and the actual annotation, enabling the model to more accurately identify the modality information of various Rhododendron lapponicum, and finally generating a rhododendron modality recognition integration module, improving the recognition performance of the model.

[0054] To build a complete Rhododendron lapponicum recognition module, it is necessary to connect the rhododendron modal recognition integration module and the Rhododendron lapponicum detection and segmentation module. Before starting the connection work, it is necessary to ensure that both modules have been trained and have stable performance. The Rhododendron lapponicum detection and segmentation module uses the Canny edge detection algorithm as the core technology to process the original remote sensing image. The Canny edge detection algorithm is a classic edge detection algorithm that achieves accurate extraction of image edges through a multi-step processing process. First, the input original remote sensing image is subjected to Gaussian filtering to remove noise interference in the image, making the image smoother and avoiding the impact of noise on subsequent edge detection. Then, the finite difference of the first-order partial derivative is used to calculate the gradient magnitude and direction of the image, thereby determining the regions with relatively drastic gray-level changes in the image, which often correspond to the edges of objects. Next, the non-maximum suppression technique is adopted to refine the gradient magnitude image, only retaining the points with the maximum local gradient as edge points and suppressing non-edge candidate points, making the edges clearer and more accurate. Finally, the double-threshold algorithm is used to detect and connect the edges, connecting the real edge points to form a complete edge contour. After being processed by the Canny edge detection algorithm, the Rhododendron lapponicum detection and segmentation module can accurately detect the position of Rhododendron lapponicum in the image and segment Rhododendron lapponicum from the complex background environment according to the extracted edge contour, generating pure image data containing only Rhododendron lapponicum. The rhododendron modal recognition integration module is responsible for performing in-depth analysis on the Rhododendron lapponicum image data after detection and segmentation processing, and identifying various modal information of Rhododendron lapponicum, such as variety, growth stage, health status, etc., based on the algorithms and models it carries. In the connection link, the output port of the Rhododendron lapponicum detection and segmentation module is accurately docked with the input port of the rhododendron modal recognition integration module. When a new remote sensing image is input into the entire system, the image will first enter the Rhododendron lapponicum detection and segmentation module, and the module will perform preliminary processing on the image according to the processing flow of the Canny edge detection algorithm, extracting effective image information about Rhododendron lapponicum from it. The processed image data will be automatically transmitted to the rhododendron modal recognition integration module according to the established data transmission path. After receiving the data, the rhododendron modal recognition integration module uses its internal algorithms and models to identify and judge the modal information of Rhododendron lapponicum. Through this connection method, the functions of the two modules are organically integrated to build a complete and coherent processing flow, and finally a Rhododendron lapponicum recognition module with functions of detection, segmentation, and modal recognition is obtained. This module can perform integrated processing on the input remote sensing image and directly output recognition results covering all aspects of the position, shape, and modality of Rhododendron lapponicum.

[0055] In a possible implementation manner, step S410 further includes:

[0056] Step S411: Collect the remote sensing scene information of the rhododendron image sample set to obtain a remote sensing scene sample set.

[0057] Step S412: Calculate the remote sensing scene distribution ratio according to the remote sensing scene sample set.

[0058] Step S413: Take the gap between the predetermined remote sensing scene distribution ratio and the remote sensing scene distribution ratio as the remote sensing scene distribution optimization target.

[0059] Step S414: Expand the remote sensing scene sample set based on the remote sensing scene distribution optimization target to obtain an expanded remote sensing scene sample set.

[0060] Step S415: Based on the expanded remote sensing scene sample set, expand the rhododendron image sample set according to the GAN network to obtain the expanded rhododendron image sample set.

[0061] Specifically, collect the remote sensing scene information of the rhododendron image sample set to obtain a remote sensing scene sample set. With the help of professional remote sensing data collection equipment and related technical means, comprehensively and meticulously collect various types of scene information involved in the rhododendron image sample set. The collection content covers multiple dimensions of the rhododendron growth environment, including but not limited to topographic and geomorphic information, such as different topographic types like mountains, plains, river valleys, etc.; climate condition information, such as temperature, humidity, light intensity, rainfall, etc.; vegetation coverage information, including the types and densities of other surrounding plants; and geographical location information, such as longitude, latitude, altitude, etc. Through precise collection equipment, these scene information are monitored and recorded in real time, and organized into a standardized data format. After the collection is completed, a large amount of acquired raw data are screened, classified, and sorted, and duplicate or invalid data are removed to ensure the accuracy and reliability of the data. Finally, a remote sensing scene sample set with clear structure and complete content is formed, laying a solid data foundation for the in-depth analysis and processing of rhododendron image data in the future.

[0062] To calculate the distribution ratio of remote sensing scenes, it is necessary to first sort out and clarify the various dimensions of scene information covered in the remote sensing scene sample set. These dimensions include aspects such as topography, climate conditions, vegetation coverage, etc. For example, in the topography dimension, it involves different types such as mountains, plains, hills, basins, etc.; the climate condition dimension covers different weather conditions such as sunny days, cloudy days, rainy days, snowy days, etc.; the vegetation coverage dimension can be divided into different ranges according to the level of vegetation coverage, such as 0% - 20%, 20 - 40%, etc., and statistical work is carried out for the specific information of each dimension. Taking the topography dimension as an example, it is necessary to count the number of occurrences of each topography type in the sample set. By manually counting or using data analysis software, traverse the entire remote sensing scene sample set, record the number of samples belonging to the mountain topography, and similarly count the number of samples of other topography types such as plains, hills, basins, etc. After completing the statistical work of the number of samples of all topography types, calculate the total number of samples in this dimension, that is, add up the number of samples of all topography types. Then, divide the number of samples of each topography type by the total number of samples to obtain the proportion of each topography type in the sample set, which is the distribution ratio of topography. Process the information of other dimensions according to the same process. For the climate condition dimension, count the number of samples corresponding to different weather conditions such as sunny days, cloudy days, rainy days, snowy days, etc., and then divide the number of samples of each weather condition by the total number of samples to obtain the distribution ratio of climate conditions. In the vegetation coverage dimension, count the number of samples corresponding to different vegetation coverage ranges, and also divide the number of samples of each range by the total number of samples to obtain the distribution ratio of vegetation coverage. In this way, for each scene information dimension in the remote sensing scene sample set, calculate the distribution ratios of different types under each dimension respectively, so as to comprehensively and detailedly obtain the distribution ratio of the remote sensing scene. These distribution ratio data can intuitively and clearly show the proportion of various scene information in the sample set, providing a crucial and valuable basis for subsequent data analysis, comparison, and optimization and adjustment.

[0063] Compare and analyze the pre-set predetermined remote sensing scene distribution ratio with the just calculated remote sensing scene distribution ratio. By calculating the gap between the two, clarify the difference between the current sample set and the expected target in terms of scene distribution, and use this gap as the optimization target for the remote sensing scene distribution. This target will guide the subsequent sample expansion work to ensure that the expanded sample set is more in line with the predetermined requirements in terms of scene distribution.

[0064] According to the determined optimization objective of remote sensing scene distribution, the remote sensing scene sample set is specifically expanded. By analyzing the deficiencies in the scene distribution of the current sample set, missing or underrepresented scene information is selectively supplemented. For example, if it is found that the scene information of a certain specific terrain and landform accounts for a small proportion in the sample set, the number of samples of this type of scene is increased through further data collection or simulation generation, so as to obtain an expanded remote sensing scene sample set. The expanded sample set is more balanced in scene distribution and can better meet the needs of subsequent data processing.

[0065] With the expanded remote sensing scene sample set, the data of the rhododendron image sample set is expanded by using the generative adversarial network (GAN network). The GAN network consists of a generator and a discriminator. The generator generates new simulated rhododendron image data according to the characteristics and distribution of the expanded remote sensing scene sample set; the discriminator then discriminates the generated image data to determine whether it is real. Through the continuous game between the generator and the discriminator, the data generated by the generator becomes closer and closer to the real rhododendron image, and finally a large number of expanded rhododendron image sample sets are obtained, providing richer and more diverse image data support for subsequent model training and recognition work.

[0066] In a possible implementation manner, step S200 further includes:

[0067] Step S210: The rhododendron detection and segmentation module includes a preprocessing unit, a rhododendron pixel detection unit, and a rhododendron background separation unit.

[0068] Step S220: Input the UAV remote sensing image sample set into the preprocessing unit to obtain a standard remote sensing image sample set, where the preprocessing unit includes a radiation correction model, a geometric correction model, and a Gaussian filter.

[0069] Step S230: Traverse the standard remote sensing image sample set and extract the first standard remote sensing image sample.

[0070] Step S240: Input the first standard remote sensing image sample into the rhododendron pixel detection unit to obtain the first image rhododendron pixel distribution.

[0071] Step S250: Based on the first image rhododendron pixel distribution, perform background separation processing on the first standard remote sensing image sample according to the rhododendron background separation unit to obtain the first rhododendron image sample.

[0072] Step S260: Add the first rhododendron image sample to the rhododendron image sample set.

[0073] Specifically, the Rhododendron lapponicum detection and segmentation module is internally composed of three sub-units with different functions but closely collaborating, namely the preprocessing unit, the Rhododendron pixel detection unit, and the Rhododendron background separation unit. The preprocessing unit is the starting point of the entire detection and segmentation process and shoulders the important task of improving the quality of the input image. It is equipped with professional tools such as a radiation correction model, a geometric correction model, and a Gaussian filter. The radiation correction model can effectively eliminate the image radiation error caused by factors such as sensor performance differences, atmospheric scattering and absorption, and lighting condition changes, ensuring that the information such as the brightness and color of the image can truly reflect the actual scene; the geometric correction model focuses on correcting the geometric deformation of the image caused by factors such as unstable UAV flight attitude, terrain undulation, and shooting angle, so that the position and shape of the objects in the image are restored to an accurate state; the Gaussian filter is mainly used to smooth the image, effectively filtering out the noise introduced during the image acquisition process and improving the clarity and recognizability of the image. The Rhododendron pixel detection unit deeply analyzes the preprocessed image and judges whether each pixel belongs to Rhododendron lapponicum pixel by pixel, so as to accurately determine the pixel distribution of Rhododendron lapponicum in the image. The Rhododendron background separation unit accurately separates Rhododendron lapponicum from the complex background environment based on the pixel distribution information provided by the Rhododendron pixel detection unit, removes the background part, and only retains the pure Rhododendron lapponicum image information, providing high-quality image data for the subsequent Rhododendron lapponicum recognition work.

[0074] The UAV remote sensing image sample set is sent to the preprocessing unit to obtain a standard remote sensing image sample set. The preprocessing unit is a key link in improving the image quality, and it contains a variety of processing models and filters. Among them, the radiation correction model corrects the radiation brightness of the image through the formula L=(DN - b)×K (L is the radiation brightness, DN is the digital quantization value of each pixel, b is the offset value, and K is the gain coefficient), eliminating the radiation error caused by factors such as sensor characteristics and atmospheric conditions, making the image more truly reflect the ground object information; the geometric correction model is used to correct the geometric deformation of the image caused by the UAV flight attitude, terrain undulation, etc., making the position of the ground objects in the image more accurate; the Gaussian filter can smooth the image, remove noise interference, and improve the image quality. After this series of processes, the UAV remote sensing image sample set is transformed into a standard remote sensing image sample set, providing a high-quality data basis for subsequent Rhododendron lapponicum detection and segmentation and other work.

[0075] Perform a traversal operation on the standard remote sensing image sample set. During the traversal process, extract the image samples one by one, and name the currently extracted sample the first standard remote sensing image sample. This operation is to process each image in the sample set separately to ensure that the subsequent detection and separation work can be accurately applied to each image sample.

[0076] To obtain the pixel distribution of rhododendron in the first image, the first standard remote sensing image sample needs to be input into the rhododendron pixel detection unit. The following is the specific implementation process. First is the data preparation work. It is necessary to ensure that the format of the first standard remote sensing image sample can be effectively processed by the detection unit, and convert the image into a specific data format. For example, in common image analysis, image data is converted into a suitable array form, which is convenient for subsequent feature extraction and model processing. Next, enter the feature extraction stage. In this stage, image features are extracted from multiple dimensions. On the one hand, color feature extraction is carried out. Different species of alpine rhododendrons have unique colors. For example, their flowers are brightly colored and the leaf colors are relatively fixed. To better capture these color information, the image is converted from the common RGB color space to the HSV color space. By setting specific color ranges, pixel regions that match the color characteristics of alpine rhododendrons can be screened out. For example, if the flowers of alpine rhododendrons are red, the color threshold range of red can be set, and image processing techniques are used to generate a mask of the corresponding color to mark the pixel regions that may belong to alpine rhododendrons. On the other hand, texture feature extraction is carried out. The leaves and flowers of alpine rhododendrons have unique textures. The gray-level co-occurrence matrix (GLCM) can be used to effectively extract the texture features of the image. First, the image is converted into a grayscale image, and then the gray-level co-occurrence matrix is calculated. Through this matrix, texture features such as contrast and correlation can be further extracted. These texture features help to more accurately distinguish rhododendron pixels from other background pixels. Then is the model prediction stage. A pre-trained pixel classification model, such as a convolutional neural network (CNN), is used for prediction. This model has learned a large number of features of alpine rhododendron images and corresponding pixel classification labels during the training process. Therefore, it can perform pixel classification prediction based on the input first standard remote sensing image sample. Before inputting the image, the image needs to be adjusted to the size required by the model to ensure that the model can process it normally. After being processed by the model, the probability that each pixel belongs to different categories will be output. By selecting the category with the highest probability as the classification result of the pixel, the pixel classification situation of the entire image can be obtained. Finally, it is the stage of determining the rhododendron pixel distribution. According to the pixel classification result obtained by the model prediction, it is determined which pixels belong to alpine rhododendrons. Alpine rhododendron pixels can be screened out by setting specific thresholds or based on category labels. After the screening is completed, there will be some small isolated pixel regions. These regions may be the result of noise or misclassification. In order to make the rhododendron pixel distribution more accurate and clear, morphological processing methods, such as removing small isolated regions, are used to finally obtain the accurate pixel distribution of rhododendron in the first image. This distribution information provides a key basis for subsequent processing such as background separation.

[0077] The GrabCut algorithm is used to combine the pixel distribution of rhododendrons in the first image to perform background separation on the first standard remote sensing image sample to obtain the first rhododendron image sample. First, an initial mask is constructed based on the pixel distribution of rhododendrons in the first image. The pixels determined to be rhododendrons are marked as foreground, the pixels determined to be background are marked as background, and the uncertain pixels are marked as unknown regions. Then, using this mask and the first standard remote sensing image sample as inputs, the GrabCut algorithm is started. This algorithm models the color distributions of the foreground and background based on the Gaussian Mixture Model (GMM), and continuously updates the pixel labels of the foreground and background through iterative optimization. In each iteration, the algorithm re-estimates the parameters of the GMM based on the current pixel labels, and then re-classifies the pixels in the unknown regions according to the new GMM to determine whether they are more likely to belong to the foreground or the background. After multiple iterations, until the pixel labels no longer change significantly or reach the preset number of iterations, finally, the first rhododendron image sample is extracted from the first standard remote sensing image sample according to the determined foreground pixels, achieving background separation.

[0078] The first rhododendron image sample obtained after background separation is added to the rhododendron image sample set. This operation continuously enriches the content of the rhododendron image sample set, accumulates more data resources for subsequent model training, data analysis, etc., and helps to improve the performance and accuracy of the entire rhododendron recognition.

[0079] In a possible implementation manner, step S240 further includes:

[0080] Step S241: The rhododendron pixel detection unit includes a rhododendron pixel prediction model and a rhododendron pixel discrimination model.

[0081] Step S242: Obtain multiple pixel samples of the first standard remote sensing image sample.

[0082] Step S243: Input the multiple pixel samples into the rhododendron pixel prediction model to obtain multiple rhododendron pixel prediction coefficients, where each rhododendron pixel prediction coefficient is used to represent the probability that each pixel sample belongs to a rhododendron pixel.

[0083] Step S244: Input the multiple rhododendron pixel prediction coefficients into the rhododendron pixel discrimination model to obtain multiple rhododendron pixel discrimination results, where the rhododendron pixel discrimination model includes a predetermined rhododendron pixel prediction coefficient.

[0084] Step S245: Based on the multiple rhododendron pixel discrimination results, perform rhododendron pixel extraction on the first standard remote sensing image sample to generate the pixel distribution of rhododendrons in the first image.

[0085] Specifically, in the Rhododendron detection and segmentation module, the Rhododendron pixel prediction model in the Rhododendron pixel detection unit adopts the Convolutional Neural Network (CNN) algorithm. First, a large number of remote sensing images containing Rhododendron are collected and the pixel categories are labeled, and the training set, validation set, and test set are divided. A CNN with a U-Net architecture is constructed, using convolutional layers to extract local pixel features, pooling layers to reduce dimensions, and deconvolutional layers to restore the size. The model is trained with the cross-entropy loss function and the Adam optimizer, and the validation set is used to prevent overfitting. Multiple pixel samples of the first standard remote sensing image sample are input into the trained model, and multiple Rhododendron pixel prediction coefficients are obtained through forward propagation. The Rhododendron pixel discrimination model adopts a threshold-based discrimination algorithm. Based on the data of the training set and validation set, the predetermined Rhododendron pixel prediction coefficient is determined through experiments or optimization methods, and multiple Rhododendron pixel prediction coefficients are compared with it. The corresponding pixel samples greater than or equal to the threshold are determined as Rhododendron pixels, and those less than the threshold are determined as non-Rhododendron pixels, realizing the accurate discrimination of pixel samples.

[0086] To further analyze the first standard remote sensing image sample to detect Rhododendron pixels, multiple pixel samples of this image sample need to be obtained. First, considering that the image itself is a two-dimensional matrix structure composed of numerous pixels arranged in a specific row and column, directly traverse each row and column of the image, and extract each pixel as an individual sample. For a color image, each pixel is usually represented by the values of three channels: red (R), green (G), and blue (B) to represent its color information. Therefore, when extracting pixel samples, the specific values of these three channels will be obtained simultaneously. If the image has additional band information, such as the near-infrared band, the pixel values of these bands will also be extracted. During the extraction process, a block sampling strategy is also adopted. The first standard remote sensing image sample is divided into multiple sub-blocks of the same size, and some pixels are selected from each sub-block as samples. This can not only ensure obtaining pixel features in different regions but also reduce the computational amount to a certain extent and improve the efficiency of subsequent analysis. Finally, multiple pixel samples are obtained through the above method, providing basic data for subsequent Rhododendron pixel prediction and discrimination.

[0087] After obtaining multiple pixel samples of the first standard remote sensing image sample, these samples are input into the rhododendron pixel prediction model. This model is constructed based on deep learning algorithms such as convolutional neural networks (CNNs) and is trained with a large number of labeled alpine rhododendron image data. Inside the model, each pixel sample first passes through a series of convolutional layers, where the convolutional kernels in the convolutional layers extract features from the pixel sample, capturing its color, texture, and other feature information. Then, the feature map is downsampled through the pooling layer to reduce the data volume and enhance the robustness of the features. Subsequently, the extracted features are integrated and mapped through the fully connected layer. Finally, the model uses the softmax activation function to output the rhododendron pixel prediction coefficient corresponding to each pixel sample. The value range of this coefficient is between 0 and 1, which is used to accurately characterize the probability that the pixel sample belongs to the alpine rhododendron pixel, providing a key basis for the subsequent discrimination of whether a pixel is an alpine rhododendron pixel.

[0088] After obtaining multiple rhododendron pixel prediction coefficients, they are input into the rhododendron pixel discrimination model. This model has a preset rhododendron pixel prediction coefficient, which is a key threshold determined through training and verification on a large number of alpine rhododendron image samples. The model will compare each rhododendron pixel prediction coefficient with the preset rhododendron pixel prediction coefficient in turn. When a certain rhododendron pixel prediction coefficient is greater than or equal to the preset rhododendron pixel prediction coefficient, it indicates that the probability that the pixel sample corresponding to this coefficient belongs to the alpine rhododendron pixel reaches or exceeds the preset standard. At this time, the pixel sample is marked as an alpine rhododendron pixel; while when the rhododendron pixel prediction coefficient is less than the preset rhododendron pixel prediction coefficient, it means that the probability that the pixel sample belongs to the alpine rhododendron pixel does not reach the preset standard, and it is marked as a non-alpine rhododendron pixel. Through such comparison and marking operations, the model finally outputs multiple rhododendron pixel discrimination results, providing a clear basis for accurately extracting alpine rhododendron pixels from the first standard remote sensing image sample subsequently.

[0089] Based on the precise rhododendron pixel extraction operation on the first standard remote sensing image sample, the first image rhododendron pixel distribution is generated. At this time, traverse the entire first standard remote sensing image sample, and correspond each pixel with the previously obtained rhododendron pixel discrimination results one by one. For those pixel points marked as alpine rhododendron pixels in the discrimination results, select them from the image sample and retain the pixels related to the alpine rhododendron. While the pixel points marked as non-alpine rhododendron pixels will be ignored. With the completion of the traversal, all the selected alpine rhododendron pixels form a new set. The position and distribution information of these pixels in the new set are the first image rhododendron pixel distribution. This distribution information intuitively shows the specific positions of the alpine rhododendron pixels in the first standard remote sensing image sample, clearly presenting the outline and form of the alpine rhododendron in the image, and further improving the accuracy and efficiency of the analysis of alpine rhododendron images.

[0090] In a possible implementation, step S250 further includes:

[0091] Step S251: Perform contour recognition on the first rhododendron delavayi image sample to obtain a first rhododendron contour recognition result.

[0092] Step S252: Evaluate the smoothness based on the first rhododendron contour recognition result to obtain the first rhododendron contour smoothness.

[0093] Step S253: Determine whether the first rhododendron contour smoothness is less than a predetermined rhododendron contour smoothness.

[0094] Step S254: If the first rhododendron contour smoothness is less than the predetermined rhododendron contour smoothness, generate a first rhododendron contour enhancement instruction.

[0095] Step S255: Perform rhododendron contour recognition on the first standard remote sensing image sample based on the first rhododendron contour enhancement instruction to determine a first standard contour recognition result.

[0096] Step S256: Perform deviation detection on the first rhododendron contour recognition result according to the first standard contour recognition result to determine a first rhododendron contour deviation detection result.

[0097] Step S257: Perform contour enhancement processing on the first rhododendron delavayi image sample according to the first rhododendron contour deviation detection result.

[0098] Specifically, carry out contour recognition work based on the already obtained first rhododendron delavayi image sample. By using an edge detection algorithm, such as the Canny algorithm, detect the contour of the rhododendron delavayi part in the image, and outline the edge contour of the rhododendron delavayi, so as to obtain the first rhododendron contour recognition result, which initially presents the contour shape of the rhododendron delavayi in the image.

[0099] Evaluate the smoothness of the first rhododendron contour based on the recognition result, and then obtain the smoothness of the first rhododendron contour. First, the recognition result of the first rhododendron contour presents as a set of contour points, and each point carries coordinate information in the two-dimensional plane. Then, starting from these contour points, calculate the vectors between adjacent points, that is, subtract the coordinates of the previous point from the coordinates of the next point to obtain the direction and distance between two adjacent points. Considering the closedness of the contour, the vector between the last point and the first point also needs to be calculated. After completing the vector calculation, use the principle of vector dot product to calculate the angle between adjacent vectors. First, calculate the modulus of each vector, that is, the length of the vector, and then according to the vector dot product formula, obtain the cosine value of the angle by dividing the dot product value by the product of the moduli of the two vectors, and then use the inverse cosine function to get the actual angle of the angle. After calculating the angles of all adjacent vectors, calculate the average value of these angles to obtain the average angle. Finally, based on the relationship between the average angle and pi, obtain a value between 0 and 1 by subtracting the ratio of the average angle to pi from 1. The larger the value, the smoother the contour. This value is the smoothness of the first rhododendron contour, which can intuitively reflect the smoothness of the current alpine rhododendron contour and provide a key quantitative basis for subsequent judgments and processing.

[0100] Compare and judge the smoothness of the first rhododendron contour with the pre-set smoothness of the predetermined rhododendron contour. The smoothness of the predetermined rhododendron contour is a reference standard determined according to a large amount of alpine rhododendron image data and actual application requirements, used to judge whether the smoothness of the current contour meets the requirements.

[0101] If the smoothness of the first rhododendron contour is less than the smoothness of the predetermined rhododendron contour, it indicates that the current alpine rhododendron contour is not smooth enough, with more serrated edges or discontinuous parts. At this time, a first rhododendron contour enhancement instruction will be generated to further optimize the contour to make it more in line with the actual situation.

[0102] Based on the generated first rhododendron contour enhancement instruction, perform the rhododendron contour recognition operation on the first standard remote sensing image sample again. By adjusting the parameters of the recognition algorithm and applying image enhancement techniques, re-recognize the alpine rhododendron contour in the image to obtain more accurate and clear contour information, and finally determine the first standard contour recognition result.

[0103] Taking the first standard contour recognition result as the benchmark, perform deviation detection on the first rhododendron contour recognition result. By comparing the differences between the two contours, calculate indicators such as the position deviation and shape difference of each point on the contour to determine the first rhododendron contour deviation detection result, so as to understand the deviation degree and specific deviation situation between the first rhododendron contour recognition result and the standard contour.

[0104] According to the first rhododendron contour deviation detection result, perform contour enhancement processing on the first alpine rhododendron image sample. According to the specific situation of the deviation, adopt corresponding image restoration and edge smoothing techniques to optimize and adjust the contour of the alpine rhododendron image, making the contour more accurate and smooth, thereby improving the quality and usability of the alpine rhododendron image.

[0105] It should be noted that the above-mentioned order of the embodiments of the present application is only for description and does not represent the superiority or inferiority of the embodiments. And the above describes specific embodiments of this specification. In addition, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0106] The above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included within the protection scope of the present application.

[0107] This specification and the drawings are only exemplary descriptions of the present application and are considered to have covered any and all modifications, variations, combinations, or equivalents within the scope of the present application. Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalent technologies, the present application is intended to include these changes and modifications.

Claims

1. An intelligent identification method of alpine rhododendron based on drone sensors, characterized in that: The method comprises: Obtaining a sample set of UAV remote sensing images of the alpine rhododendron planting area and an alpine rhododendron identification sample set, and performing identification modality classification on the alpine rhododendron identification sample set to obtain Q rhododendron identification modal areas, where Q is a positive integer greater than 1; Inputting the UAV remote sensing image sample set into the alpine rhododendron detection and segmentation module to obtain the alpine rhododendron image sample set; A cuckoo recognition learning channel is introduced to perform loss optimization learning on the alpine cuckoo image sample set and the Q cuckoo recognition modal areas to generate a cuckoo Q modal recognition unit; the cuckoo recognition learning channel includes P cuckoo recognition learners and a cuckoo recognition loss calculator, wherein P is a positive integer greater than 1; Extracting the qth cuckoo identification modal area according to the Q cuckoo identification modal areas, wherein q is a positive integer, 1≤q≤Q; Based on the azalea recognition loss calculator, according to the alpine azalea image sample set and the qth azalea recognition modal area, the P azalea recognition learners are subjected to loss optimization learning to generate the qth modal azalea recognition model; Taking the output data set of the qth modal cuckoo recognition model as input information and the qth cuckoo recognition modal area as output information, training the qth modal cuckoo recognition fusion model; Merging the qth modal cuckoo recognition model with the input layer of the qth modal cuckoo recognition fusion model to generate a cuckoo qth modal recognition unit, and adding the cuckoo qth modal recognition unit to the cuckoo Q modal recognition unit; The GAN network and the predetermined remote sensing scene distribution are introduced to perform generalized reinforcement learning on the cuckoo Q modal recognition unit to obtain a cuckoo modal recognition integrated module, and the alpine cuckoo detection and segmentation module is combined to build an alpine cuckoo recognition module; including: Based on the GAN network, data expansion is performed on the alpine rhododendron image sample set according to the predetermined remote sensing scene distribution ratio to obtain an expanded rhododendron image sample set; Based on the mapping relationship between the alpine rhododendron image sample set and the Q rhododendron recognition modal areas, a rhododendron image-recognition modal mapping network is established; Based on the cuckoo image-recognition modality mapping network, the expanded cuckoo image sample set is migrated and annotated according to the Q cuckoo recognition modality areas to generate an expanded cuckoo recognition modality sample set; Performing generalized reinforcement learning on the cuckoo Q modality recognition unit according to the expanded cuckoo image sample set and the expanded cuckoo recognition modality sample set to generate the cuckoo modality recognition integrated module; Connecting the azalea modal recognition integration module and the alpine azalea detection and segmentation module to obtain the alpine azalea recognition module; According to the drone sensor, a remote sensing image of the alpine rhododendron planting area is obtained, and the remote sensing image is input into the alpine rhododendron recognition module to generate a rhododendron Q modal recognition result.

2. The intelligent identification method of alpine rhododendron based on drone sensor according to claim 1 is characterized in that: Based on the cuckoo recognition loss calculator, according to the alpine cuckoo image sample set and the qth cuckoo recognition modal area, loss optimization learning is performed on the P cuckoo recognition learners to generate the qth modal cuckoo recognition model, including: According to the alpine rhododendron image sample set and the qth rhododendron recognition modal area, supervised training is performed on the P rhododendron recognition learners respectively, and each time a predetermined number of training times is trained, P rhododendron recognition loss coefficients are calculated according to the rhododendron recognition loss calculator, wherein the rhododendron recognition loss calculator includes a rhododendron recognition loss calculation formula, and the rhododendron recognition loss calculation formula is: ； Among them, LOSS represents the cuckoo recognition loss coefficient, M represents the predetermined number of training times, m represents the mth training, M and m are both positive integers, 1≤m≤M, SUO m Represents the number of cuckoo recognition prediction samples in the mth training, SUX m Represents the number of correct samples for cuckoo recognition prediction in the mth training; If the P cuckoo recognition loss coefficients are less than the cuckoo recognition loss threshold, generating P same-modal cuckoo recognition models; The P same-modal cuckoo recognition models are connected as parallel nodes to generate the q-th modal cuckoo recognition model.

3. The intelligent identification method of alpine rhododendron based on drone sensor as claimed in claim 1 is characterized in that: Based on the GAN network, data expansion is performed on the alpine rhododendron image sample set according to the predetermined remote sensing scene distribution ratio to obtain an expanded rhododendron image sample set, including: Collecting remote sensing scene information of the alpine rhododendron image sample set to obtain a remote sensing scene sample set; Calculating a remote sensing scene distribution ratio according to the remote sensing scene sample set; The remote sensing scene distribution optimization target is determined according to the gap between the predetermined remote sensing scene distribution ratio and the remote sensing scene distribution ratio; Based on the remote sensing scene distribution optimization goal, the remote sensing scene sample set is expanded to obtain an expanded remote sensing scene sample set; Based on the expanded remote sensing scene sample set, data expansion is performed on the alpine rhododendron image sample set according to the GAN network to obtain the expanded rhododendron image sample set.

4. The intelligent identification method of alpine rhododendron based on drone sensor as claimed in claim 1 is characterized in that: The UAV remote sensing image sample set is input into the alpine rhododendron detection and segmentation module to obtain the alpine rhododendron image sample set, including: The alpine rhododendron detection and segmentation module includes a pre-processing unit, a rhododendron pixel detection unit and a rhododendron background separation unit; Inputting the UAV remote sensing image sample set into the preprocessing unit to obtain a standard remote sensing image sample set, wherein the preprocessing unit includes a radiation correction model, a geometric correction model and a Gaussian filter; Traversing the standard remote sensing image sample set, extracting a first standard remote sensing image sample; Inputting the first standard remote sensing image sample into the azalea pixel detection unit to obtain the azalea pixel distribution of the first image; Based on the rhododendron pixel distribution of the first image, performing background separation processing on the first standard remote sensing image sample according to the rhododendron background separation unit to obtain a first alpine rhododendron image sample; The first alpine rhododendron image sample is added to the alpine rhododendron image sample set.

5. The intelligent identification method of alpine rhododendron based on drone sensor as claimed in claim 4 is characterized in that: Inputting the first standard remote sensing image sample into the cuckoo pixel detection unit to obtain the cuckoo pixel distribution of the first image includes: The cuckoo pixel detection unit includes a cuckoo pixel prediction model and a cuckoo pixel discrimination model; Obtaining a plurality of pixel samples of the first standard remote sensing image sample; Inputting the plurality of pixel samples into the azalea pixel prediction model to obtain a plurality of azalea pixel prediction coefficients, wherein each azalea pixel prediction coefficient is used to characterize the probability that each pixel sample belongs to an alpine azalea pixel; Inputting the plurality of cuckoo pixel prediction coefficients into the cuckoo pixel discrimination model to obtain a plurality of cuckoo pixel discrimination results, wherein the cuckoo pixel discrimination model includes predetermined cuckoo pixel prediction coefficients; Based on the multiple cuckoo pixel discrimination results, cuckoo pixels are extracted from the first standard remote sensing image sample to generate the cuckoo pixel distribution of the first image.

6. The intelligent identification method of alpine rhododendron based on drone sensor as claimed in claim 4 is characterized in that: Based on the rhododendron pixel distribution of the first image, the first standard remote sensing image sample is subjected to background separation processing by the rhododendron background separation unit to obtain a first alpine rhododendron image sample, further comprising: Performing contour recognition according to the first alpine rhododendron image sample to obtain a first rhododendron contour recognition result; Performing a smoothness evaluation according to the first cuckoo contour recognition result to obtain the first cuckoo contour smoothness; Determining whether the first cuckoo contour smoothness is less than a predetermined cuckoo contour smoothness; If the first cuckoo contour smoothness is less than the predetermined cuckoo contour smoothness, generating a first cuckoo contour enhancement instruction; Performing cuckoo contour recognition on the first standard remote sensing image sample based on the first cuckoo contour enhancement instruction to determine a first standard contour recognition result; Performing deviation detection on the first cuckoo contour recognition result according to the first standard contour recognition result to determine a first cuckoo contour deviation detection result; The first alpine rhododendron image sample is subjected to contour enhancement processing according to the first rhododendron contour deviation detection result.

Images