Method for semantic segmentation of multi-source hyperspectral point clouds of unmanned aerial vehicles based on integrated prediction weak supervision

Abstract

The application discloses a kind of weakly supervised unmanned aerial vehicle multi-source hyperspectral point cloud semantic segmentation methods based on integrated prediction. It is based on unmanned aerial vehicle multi-source hyperspectral point cloud data as the research object, construct weakly supervised semantic segmentation framework based on integrated prediction, on the basis of incomplete supervision, embedding consistency constraint of integrated prediction, entropy regularization guided by integrated prediction result and self-adaptive soft pseudo label method, make full use of the encoding information of unlabeled data, provide various spatial and spectral constraints for weakly supervised training, enhance the semantic segmentation ability of network model. The application solves the problems of high cost of forestry vegetation data sampling, rare and difficult to obtain forest sample and time-consuming and laborious artificial high-precision label large scene forest vegetation data in the prior art; It has the advantages of guaranteeing the training efficiency of forest sample, increasing the supervision source for weakly supervised forestry information extraction network model and improving the forestry vegetation information extraction ability of learning model.

Description

Technical Field

[0001] This invention relates to a semantic segmentation method for multi-source hyperspectral point clouds from unmanned aerial vehicles (UAVs) based on ensemble prediction. More specifically, it is a method within a deep learning framework to address the difficulties in vegetation sample collection and the lack of labeled data, thereby promoting forestry informatization using UAV hyperspectral point clouds. Background Technology

[0002] Forest vegetation is the main body of terrestrial ecosystems, undertaking ecological service functions such as climate regulation and water conservation. It plays an irreplaceable role in maintaining regional ecology and environment, global carbon balance, and mitigating global climate change. Its own characteristics and changes have significant impacts on the terrestrial biosphere and other surface processes. Globally, numerous national and regional institutions are seeking more limited spatial coverage, more accurate forest inventories, and more precise data on forest vegetation attributes and spatial distribution to monitor changes in forest ecosystems caused by natural and human activities. This forest vegetation resource monitoring provides policymakers with more information to support their carbon resource, regional biodiversity, and sustainable resource development strategies. As the supervision of forest vegetation resources continues to strengthen, users are demanding higher accuracy and resolution in the extraction of forest vegetation parameters. Forest vegetation monitoring is constantly developing towards real-time, multi-dimensional, and refined methods, placing higher demands on the precision and breadth of data fusion and processing.

[0003] Hyperspectral data can capture biophysical and chemical characteristics across various temporal and spatial scales, offering fine spectral resolution bands and rich spectral information. Its extracted narrow-band vegetation indices can mitigate the effects of atmospheric and water absorption and broadband vegetation index saturation. Therefore, hyperspectral data has significant advantages in estimating tree canopy parameters such as chlorophyll content, forest type (type group), and stock volume, as well as in obtaining quantitative and qualitative information on crops and vegetation. In recent years, many researchers have utilized hyperspectral technology for tree species classification and identification. Studies have shown that using only airborne hyperspectral data to classify some tree species achieves an accuracy of 60%-90%, demonstrating the tree species classification capabilities of hyperspectral data.

[0004] Although hyperspectral data has a narrow spectral range, it can accurately detect various land cover types with subtle spectral differences, improving the accuracy of tree vegetation classification methods based on hyperspectral data. However, it remains limited in distinguishing tree species with similar spectral characteristics, and optical data can only detect canopy surface information, restricting the accuracy of tree vegetation identification. LiDAR, on the other hand, can acquire detailed three-dimensional information on tree vegetation structure, which has significant advantages for forest vegetation type identification, forest vegetation structural characteristics, and canopy physicochemical features. However, it lacks corresponding spectral information, making it difficult to classify complex forest vegetation types. Combining hyperspectral and LiDAR data from multiple sources to achieve complementary advantages and applying them to forestry vegetation research has become a new research hotspot, and some scholars have already conducted related research in this field. Studies have shown that fusing LiDAR data and hyperspectral imagery data can better improve the accuracy of forest vegetation parameter estimation and tree species type identification than single-source data. Tree species mapping research based on the fusion of hyperspectral and LiDAR data covers a very wide range of areas, including urban areas, subtropical forests, natural temperate forests, and natural forests. In addition, with the rapid development of UAV remote sensing technology and communication technology, the fusion of optical and LiDAR data from the same UAV remote sensing platform for single tree species identification and classification research has also shown that its identification accuracy is more outstanding than that based on single data type identification and classification.

[0005] Current research on land cover classification, forestry vegetation information extraction, tree species identification and classification, etc., generated from multi-source hyperspectral point cloud data fusion of LiDAR and hyperspectral data typically involves converting LiDAR data into image data, fusing it with hyperspectral image data, and employing deep learning methods such as convolutional neural networks. However, training deep learning models requires not only massive amounts of supervised samples, but also precise labeling of these training samples. Therefore, for multi-source hyperspectral point cloud data with complex labeled land cover, large data volume, disorder, and discreteness, especially in forest areas, it is a very time-consuming and labor-intensive task. Weakly supervised semantic segmentation methods can complete point cloud segmentation and classification tasks using only a portion of labeled samples, which is an effective way to solve the above problems. Traditional weakly supervised point cloud segmentation frameworks suffer from high computational cost consistency constraints, overconfidence (overfitting) due to minimizing entropy values, and problems such as training efficiency and uncertain weights in pseudo-label learning methods, resulting in poor performance of weakly supervised point cloud segmentation.

[0006] Therefore, it is necessary to develop a method that can effectively process large amounts of unsupervised information from UAV multi-source hyperspectral point cloud data using only a small number of labeled samples. Summary of the Invention

[0007] The purpose of this invention is to provide a semantic segmentation method for multi-source hyperspectral point clouds from unmanned aerial vehicles (UAVs) based on ensemble prediction and weak supervision. This method utilizes the latent spatial and spectral information of unlabeled data within an incomplete supervised network to develop an ensemble constraint method. The focus is on consistency constraints based on ensemble prediction, generating more stable semantic predictions for forestry vegetation targets without increasing computational costs. The ensemble prediction results are introduced to assist in entropy regularization of unlabeled points, reducing inter-class overlap, improving network classification reliability, and suppressing overfitting in minimizing entropy values. Based on the integrated prediction consistency constraint, this study investigates an adaptive pseudo-label learning strategy. This strategy ensures efficient sample training while adding a source of supervision to the weakly supervised network model, thereby improving the model's ability to extract forestry vegetation information. It addresses the problems of high sampling costs, scarce and difficult-to-obtain samples, time-consuming and labor-intensive manual labeling of large-scale forest vegetation data, high computational costs associated with consistency constraints in traditional weakly supervised point cloud segmentation frameworks, overfitting due to entropy minimization, and the uncertainty of training efficiency and weights in pseudo-label learning methods, all of which contribute to poor weakly supervised point cloud segmentation performance.

[0008] To achieve the above objectives, the technical solution of this invention is as follows: a semantic segmentation method for multi-source hyperspectral point clouds of UAVs based on ensemble prediction and weak supervision, characterized in that: based on the multi-source hyperspectral point cloud data of UAVs fused with UAV lidar and hyperspectral imagery, more useful feature information is extracted, and a weakly supervised semantic segmentation framework based on ensemble prediction is constructed. On the basis of incomplete supervision, consistency constraints of ensemble prediction, entropy regularization guided by ensemble prediction results, and an adaptive soft pseudo-labeling method are embedded to fully utilize the encoding information of unlabeled data space and spectrum, providing multiple constraints for weakly supervised training and enhancing the semantic segmentation capability of the weakly supervised semantic segmentation network model framework based on ensemble prediction.

[0009] The above technical solution, specifically the semantic segmentation method for multi-source hyperspectral point clouds of unmanned aerial vehicles based on ensemble prediction, includes the following steps:

[0010] Step 1, incomplete supervised learning;

[0011] Step 2, Consistency constraints based on integrated predictions;

[0012] Step 3, entropy regularization guided by prediction results;

[0013] Step 4, Adaptive pseudo-label learning.

[0014] In the above technical solution, the specific method for incomplete supervised learning in step 1 is as follows:

[0015] With a sufficient amount of sampled data, and assuming that the sampled label data is close to the independent and identically distributed (IID) assumption, the model can achieve semantic segmentation results similar to those of fully supervised learning through weakly supervised learning. This invention first employs a random sampling strategy to select label sample data for incomplete supervised learning of UAV multi-source hyperspectral point clouds; then, it calculates the square root weighted cross-entropy loss based on the number of categories of the selected label sample points to construct an incomplete supervised learning network framework.

[0016] In the above technical solution, the specific method for consistency constraints based on integrated prediction in step 2 is as follows:

[0017] To address the issues of randomness and overlapping input samples in the backbone network, an ensemble prediction iterative update is employed during the training phase. This ensures that each training step requires only one forward propagation, maintaining efficient sample training. Furthermore, the overlapping training regions of different input samples contain varying global information, which can be viewed as point-level data augmentation. For each point-level data point, an exponential moving average is used to calculate the ensemble value. Perform a correlation update; that is, the ensemble prediction value updated at the t-th time. The calculation is shown in equation (1).

[0018]

[0019] Where: α is the update weight. p is the ensemble prediction value updated by the model at the (t-1)th time. t This is the current predicted value;

[0020] In each training iteration, the updated ensemble prediction distribution will be... Compared with the current predicted distribution p i The consistency cost is described by the Kullback-Leibler divergence (KLD), and consistency constraints are applied; the consistency cost V(P) i ) and consistency loss L epc The calculation is shown in equations (2) and (3);

[0021]

[0022]

[0023] in: It is an integrated prediction P i Point p represents the posterior probability of class c. ic Is the current prediction P i The point represents the posterior probability of category c, K is the number of categories in the UAV hyperspectral point cloud dataset, and N is the number of points in the training dataset.

[0024] In the above technical solution, in step 3, the entropy regularization guided by the prediction results adopts the entropy regularization (ER) method to improve the semantic segmentation performance of the model by utilizing the posterior probability of multi-source hyperspectral points of unlabeled UAVs; and the entropy regularization processing of unlabeled points is guided by comparing the current prediction with the ensemble prediction.

[0025] The specific method for entropy regularization guided by prediction results is as follows:

[0026] For unlabeled points P that match the current prediction and the integrated prediction ic By minimizing prediction entropy, we can reduce class overlap and obtain significant semantic features.

[0027] Unlabeled point P with inconsistent predictions iu This means that the prediction result at this point is unstable. Therefore, the method of maximizing prediction entropy is adopted to encourage high uncertainty in network predictions and suppress overfitting. Maximizing the prediction entropy value is equivalent to minimizing the negative prediction entropy value; the entropy value H(P) i and entropy regularization loss L er The calculation is shown in equations (4) and (5):

[0028]

[0029]

[0030] Where, p ic For P i Point prediction is the posterior probability of category c, K is the number of categories appearing in the UAV multi-source hyperspectral point cloud dataset, and || is the number of points in the point set.

[0031] In the above technical solution, the specific method for adaptive pseudo-label learning in step 4 is as follows:

[0032] The integrated prediction results of unlabeled UAV multi-source hyperspectral points are directly used as pseudo-label data to maintain the original model training speed to the greatest extent; with consistency cost V(P) i As a metric for the weight of pseudo-labels, the weights of pseudo-labels are adaptively calculated; the weight of each pseudo-label point... and pseudo-label learning loss L ps The calculation is shown in equations (6) and (7);

[0033]

[0034]

[0035] Among them, y icThe pseudo-labels are obtained by integrating the predicted values ​​through the argmax function, p ic For P i The point prediction is the posterior probability of class c, K is the number of classes appearing in the UAV multi-source hyperspectral point cloud dataset, and || is the number of points in the point set;

[0036] Combining all the above-mentioned constraints and losses, the final loss function L of the network is... all Calculated using equation (8):

[0037] L all =L se +L epc +L er +λL ps (8)

[0038] Where λ is the weighting factor; L se This represents the loss function.

[0039] In this invention, "multi-source" refers to the result of fusing hyperspectral point cloud data, laser data, and hyperspectral data from UAVs.

[0040] The present invention has the following advantages:

[0041] (1) The present invention is based on the semantic segmentation framework of weakly supervised UAV hyperspectral point cloud based on ensemble prediction. It makes full use of the encoded information of unlabeled points and provides a variety of spatial and spectral constraints for the training of weakly supervised network in this project by adopting incomplete supervised learning, consistency constraints based on ensemble prediction, entropy regularization based on prediction comparison guidance and adaptive pseudo-label learning methods, thereby improving the semantic segmentation capability of UAV multi-source hyperspectral point cloud.

[0042] (2) This invention has the characteristics of ensuring the training efficiency of forest tree samples while adding a supervision source to the weakly supervised forestry information extraction network model, thereby improving the learning model's ability to extract forestry vegetation information; it solves the problems of high sampling cost of forestry vegetation data, scarce and difficult-to-obtain forest tree samples, and time-consuming and laborious manual high-precision labeling of large-scale forest vegetation data under the existing technology.

[0043] (3) The present invention adds integrated prediction and consistency constraints, which can effectively utilize the features of unlabeled points and provide positive self-supervised signals for model training; the entropy regularization method effectively alleviates the problem of model overfitting caused by entropy minimization, and more fully explores the features of point clouds themselves, further improving the model classification performance. This allows the present invention to effectively process large amounts of data (i.e., data results of large-scale laser point clouds and hyperspectral image fusion) and unsupervised information using only a small number of labeled samples (about 0.1% of the number of training point cloud samples in supervised classification methods). Attached Figure Description

[0044] Figure 1This is a framework diagram of the semantic segmentation method for hyperspectral point clouds of UAVs based on integrated prediction according to the present invention. Detailed Implementation

[0045] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, these descriptions do not constitute a limitation of the present invention and are merely illustrative. The advantages of the present invention will become clearer and easier to understand through this description.

[0046] This invention proposes a weakly supervised semantic segmentation framework for UAV multi-source hyperspectral point cloud data. It introduces ensemble learning to improve training efficiency and create more representative comparative samples, enhancing the stability of the semantic segmentation model. An entropy regularization method assisted by ensemble prediction results reduces class overlap and suppresses overfitting caused by minimizing entropy values. An adaptive pseudo-label learning method is used to adjust weights and increase the supervision sources for model training. This proposed model is expected to solve the problems of weak intelligent interpretation and insufficient sample sampling of UAV multi-source hyperspectral point cloud data, providing a general framework for extracting forestry vegetation information from UAV hyperspectral point cloud data.

[0047] Referring to the attached figures, a weakly supervised semantic segmentation method for UAV multi-source hyperspectral point clouds based on ensemble prediction is proposed. This method uses multi-source hyperspectral point cloud data from UAVs, fused from UAV lidar and hyperspectral imagery, as the research object. It constructs a weakly supervised semantic segmentation framework based on ensemble prediction. On the basis of incomplete supervision, it embeds consistency constraints of ensemble prediction, entropy regularization guided by ensemble prediction results, and an adaptive soft pseudo-labeling method. This fully utilizes the spatial and spectral encoding information of the unlabeled data, providing multiple spatial and spectral constraints for weakly supervised training and enhancing the semantic segmentation capability of the weakly supervised semantic segmentation network model framework based on ensemble prediction.

[0048] Furthermore, the semantic segmentation method for weakly supervised UAV multi-source hyperspectral point clouds based on ensemble prediction specifically includes the following steps:

[0049] Step 1, incomplete supervised learning;

[0050] Step 2, Consistency constraints based on integrated predictions;

[0051] Step 3, entropy regularization guided by prediction results;

[0052] Step 4, Adaptive pseudo-label learning (e.g.) Figure 1 As shown, in Figure 1 In this context, the encoder and decoder are the encoding and decoding structures, i.e., the algorithm network structure; the input data is the fused UAV hyperspectral point cloud data, and the label data is also UAV hyperspectral point cloud data.

[0053] Furthermore, in step 1, the specific method for incomplete supervised learning is as follows:

[0054] With a sufficient amount of sampled data, and assuming that the sampled label data is approximately independently and identically distributed, the model can achieve semantic segmentation results similar to those obtained through fully supervised learning using a weakly supervised learning approach. This invention first employs a random sampling strategy to select label sample data for incomplete supervised learning of UAV multi-source hyperspectral point clouds; then, based on the number of categories of the selected label sample points, it calculates the square root weighted cross-entropy loss to construct the incomplete supervised learning network framework for UAV multi-source hyperspectral point clouds.

[0055] Furthermore, in step 2, the specific method for consistency constraints based on integrated prediction is as follows:

[0056] To address the issues of randomness and overlapping input samples in the backbone network, an ensemble prediction iterative update is employed during the training phase. This ensures that each training step requires only one forward propagation, maintaining efficient sample training. Furthermore, the overlapping training regions of different input samples contain varying global information, which can be viewed as point-level data augmentation. For each point-level data point, an exponential moving average is used to calculate the ensemble value. Perform a correlation update; that is, the ensemble prediction value updated at the t-th time. The calculation is shown in equation (1).

[0057]

[0058] Where: α is the update weight. p is the ensemble prediction value updated by the model at the (t-1)th time. t This is the current predicted value;

[0059] In each training iteration, the updated ensemble prediction distribution will be... Compared with the current predicted distribution p i The consistency cost is described by the Kullback-Leibler divergence (KLD), and consistency constraints are applied; the consistency cost V(P) i ) and consistency loss L epc The calculation is shown in equations (2) and (3);

[0060]

[0061]

[0062] in: It is an integrated prediction P i Point p represents the posterior probability of class c. ic Is the current prediction P iThe point represents the posterior probability of class c, K is the number of classes in the dataset, and N is the number of points in the training dataset.

[0063] Furthermore, in step 3, the entropy regularization guided by the prediction results adopts the entropy regularization (ER) method to improve the semantic segmentation performance of the model by utilizing the posterior probability of unlabeled UAV multi-source hyperspectral points; and the entropy regularization processing of unlabeled points is guided by comparing the current prediction with the ensemble prediction.

[0064] The specific method for entropy regularization guided by prediction results is as follows:

[0065] For unlabeled points P that match the current prediction and the integrated prediction ic By minimizing prediction entropy, we can reduce class overlap and obtain significant semantic features of UAV multi-source hyperspectral point clouds.

[0066] Unlabeled point P with inconsistent predictions iu This means that the prediction result at this point is unstable. Therefore, the method of maximizing prediction entropy is adopted to encourage high uncertainty in network predictions and suppress overfitting. Maximizing the prediction entropy value is equivalent to minimizing the negative prediction entropy value; the entropy value H(P) i and entropy regularization loss L er The calculation is shown in equations (4) and (5):

[0067]

[0068]

[0069] Where, p ic For P i Point prediction is the posterior probability of category c, K is the number of categories appearing in the UAV multi-source hyperspectral point cloud dataset, and || is the number of points in the point set.

[0070] Furthermore, in step 4, the specific method for adaptive pseudo-label learning is as follows:

[0071] The integrated prediction results of unlabeled UAV multi-source hyperspectral points are directly used as pseudo-label data to maintain the original network model training speed to the greatest extent; with consistency cost V(P) i As a metric for the weight of pseudo-labels, the weights of pseudo-labels are adaptively calculated; the weight of each pseudo-label point... and pseudo-label learning loss L ps The calculation is shown in equations (6) and (7);

[0072]

[0073]

[0074] Among them, y ic The pseudo-labels are obtained by integrating the predicted values ​​through the argmax function, p ic For P i The point prediction is the posterior probability of class c, K is the number of classes appearing in the UAV multi-source hyperspectral point cloud dataset, and || is the number of points in the point set;

[0075] Combining all the above-mentioned constraints and losses, the final loss function L of the network is... all Calculated using equation (8):

[0076] L all =L se +L epc +L er +λL ps (8)

[0077] Where λ is the weighting factor.

[0078] To verify the accuracy of this application, the following experiments were conducted:

[0079] A semantic segmentation method for multi-source hyperspectral point clouds based on ensemble prediction was evaluated on an airborne multispectral LiDAR point cloud dataset. Each laser point in this dataset contains geometric coordinate information and information in three bands: 532nm, 1062nm, and 1550nm. Ground features in the entire scene were classified into six categories—roads, grass, trees, buildings, bare ground, and power lines—through manual interpretation.

[0080] The airborne multispectral LiDAR point cloud dataset consists of 12,137, 1,870, and 2,874 point cloud samples from the training, validation, and test datasets, respectively. Each sample contains N = 4,096 points.

[0081] The Adam optimizer was selected for the experiment. The initial learning rate, number of nearest neighbors K, and grid downsampling size were set to 0.01, 16, and 0.05m, respectively. A learning rate decay exponent of 0.98 was used to train the model for 200 epochs. Based on the available GPU memory, the number of input points was chosen to be 65,536, and the batch size was 3. One-thousandth of the points in the original dataset were randomly selected as markers for model training.

[0082] To reduce the impact of model randomness on the results, fixed sparse markers and test data were selected for model training and testing. Furthermore, to more accurately evaluate the model's testing accuracy, this paper uses overall accuracy (OA), mean intersection over union (mIoU), precision, recall, and F1 score (F1) as evaluation metrics to analyze the experimental results. Coordinate and spectral information were used as input features for network training and testing, and the test results are shown in Table 1 below.

[0083] Table 1 Test Results

[0084]

[0085] As shown in Table 1, compared with the fully supervised methods (SE-PointNet++ and FR-GCNet) that use coordinate information, the method of this invention, when using only one-thousandth of the marked points, achieves OA, mIoU, and IoU of roads, buildings, vegetation, bare land, and grassland that are equivalent to the fully supervised method. The supervision coverage of OA and mIoU reaches 90.19% and 64.12%, respectively.

[0086] The above analysis results show that, compared with the fully supervised method, the semantic segmentation method for multi-source hyperspectral point clouds of UAVs based on integrated prediction proposed in this invention can achieve competitive performance in the case of sparse labeled points. That is, this invention can effectively process large amounts of unsupervised information using only a small number of labeled samples.

[0087] In summary, the weakly supervised semantic segmentation method based on ensemble prediction point clouds provided by this invention has the following technical effects: adding ensemble prediction and consistency constraints can effectively utilize the features of unlabeled points, providing positive self-supervised signals for model training; the entropy regularization method can effectively alleviate the problem of model overfitting caused by entropy minimization, more fully explore the features of the point cloud itself, and further improve the model classification performance, enabling this invention to effectively process large amounts of unsupervised information using only a small number of labeled samples (approximately 0.1% of the number of training point cloud samples in supervised classification methods).

[0088] All other unspecified parts belong to the prior art.

Claims

1. A semantic segmentation method for multi-source hyperspectral point clouds based on integrated predictive weakly supervised UAVs, characterized in that: Taking UAV multi-source hyperspectral point cloud data fused with UAV lidar and hyperspectral imagery as the research object, this paper constructs a weakly supervised semantic segmentation framework based on ensemble prediction. On the basis of incomplete supervision, it embeds consistency constraints of ensemble prediction, entropy regularization guided by ensemble prediction results, and an adaptive soft pseudo-label method. It makes full use of the spatial and spectral encoding information of unlabeled data to provide geometric and spectral constraints for weakly supervised training, thereby enhancing the semantic segmentation capability of the weakly supervised semantic segmentation network model framework based on ensemble prediction. A semantic segmentation method for multi-source hyperspectral point clouds from unmanned aerial vehicles based on ensemble prediction includes the following steps: Step 1, incomplete supervised learning; Step 2, Consistency constraints based on integrated predictions; Step 3, entropy regularization guided by prediction results; Step 4, Adaptive pseudo-label learning; In step 2, the specific method for consistency constraints based on ensemble prediction is as follows: During the training phase, ensemble prediction iterative updates are employed. The overlapping training regions of hyperspectral point cloud samples from different UAVs contain different global information, which is considered as a point-level data augmentation. For each point-level data, an exponential moving average is used to calculate the ensemble value. Perform a correlation update; that is, the ensemble prediction value updated at the t-th time. The calculation is shown in equation (1): Where: α is the update weight. p is the ensemble prediction value updated by the model at the (t-1)th time. t This is the current predicted value; In each training iteration, the updated ensemble prediction distribution will be... Compared with the current predicted distribution p i Consistency cost is described by KL divergence, and consistency constraints are applied; consistency cost V(P) i ) and consistency loss L epc The calculation is shown in equations (2) and (3): in: It is an integrated prediction P i Point p represents the posterior probability of class c. ic Is the current prediction P i The point represents the posterior probability of class c, K is the number of classes in the UAV hyperspectral point cloud dataset, and N is the number of points in the UAV hyperspectral point cloud training dataset. In step 3, the entropy regularization guided by the prediction results adopts the entropy regularization method to improve the semantic segmentation performance of the learning model by utilizing the posterior probability of the hyperspectral point cloud of the unlabeled UAV; and the entropy regularization processing of unlabeled points is guided by comparing the current prediction with the ensemble prediction. The specific method for entropy regularization guided by prediction results is as follows: For unlabeled points P that match the current prediction and the integrated prediction ic By minimizing prediction entropy, we can reduce class overlap and obtain significant semantic features. Unlabeled point P with inconsistent predictions iu This means the prediction result at this point is unstable. Therefore, the method of maximizing prediction entropy is adopted to encourage high uncertainty in network predictions and suppress overfitting. Maximizing the prediction entropy value is equivalent to minimizing the negative prediction entropy value; the entropy value H(P) i and entropy regularization loss L er The calculation is shown in equations (4) and (5): Where, p ic For P i The point prediction is the posterior probability of category c, K is the number of categories appearing in the UAV hyperspectral point cloud dataset, and || is the number of points in the point set; In step 4, the specific method for adaptive pseudo-label learning is as follows: The ensemble prediction results of unlabeled UAV multi-source hyperspectral point clouds are directly used as pseudo-label data to maintain the original model training speed to the greatest extent; with consistency cost V(P) i As a metric for the weight of pseudo-labels, the weights of pseudo-labels are adaptively calculated; the weight of each pseudo-label point... and pseudo-label learning loss L ps The calculation is shown in equations (6) and (7); Among them, y ic The pseudo-labels are obtained by integrating the predicted values ​​through the argmax function, p ic For P i The point prediction is the posterior probability of class c, K is the number of classes appearing in the UAV hyperspectral point cloud dataset, and || is the number of points in the point set; Combining all constraint losses, the final loss function L of the network is... all Calculated using equation (8): THE all =L se +L epc +L er +λL ps (8) Where λ is the weighting factor.

2. The semantic segmentation method for multi-source hyperspectral point clouds based on integrated predictive weakly supervised UAVs according to claim 1, characterized in that: In step 1, the specific method for incomplete supervised learning is as follows: First, a random sampling strategy is used to select labeled sample data for incomplete supervised learning of UAV multi-source hyperspectral point clouds; then, the square root weighted cross-entropy loss is calculated based on the number of categories of the selected labeled sample points to construct the UAV hyperspectral point cloud incomplete supervised learning network framework.

Images