Weakly supervised semantic segmentation method for remote sensing images based on dual learning mechanism

Abstract

The application discloses a weakly annotated remote sensing image semantic segmentation method based on a double learning mechanism and belongs to the field of remote sensing image processing. The implementation process comprises the following steps: 1) a semi-supervised classification network is trained by using a weakly annotated remote sensing image dataset with limited annotation quantity and precision, and a pixel-level pseudo segmentation mask is generated; 2) a dataset redivision strategy guided by a mask disturbance mechanism is designed, and the training dataset is redivided into two subsets, namely an annotated subset and an unannotated subset; 3) a semi-supervised semantic segmentation network is trained by using the redivided dataset; 4) a dynamic threshold strategy is proposed for the training of the semi-supervised semantic segmentation network, and the confidence bias of the pseudo label is reduced; and 5) in the test stage, the segmentation result of a remote sensing image to be tested is output by using the semantic segmentation network. The method can effectively utilize rich ground feature semantic information in a large number of unannotated samples, and realizes accurate and efficient segmentation of multi-class ground feature targets of remote sensing images under the condition of lower annotation cost than traditional weak annotation.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, specifically relating to a semantic segmentation method for weakly labeled remote sensing images based on a dual learning mechanism. Background Technology

[0002] With the rapid development of Earth observation technology, massive amounts of high-resolution remote sensing images have become readily available, meeting the basic requirements for fine-grained land surface classification. Semantic segmentation of remote sensing images aims to assign corresponding semantic category labels to each pixel of the image, and is a widely used and important technology in land surface classification, greatly promoting the development of intelligent interpretation of remote sensing images. This task has been applied in many fields, such as urban planning, environmental monitoring, and resource exploration. In recent years, weakly labeled semantic segmentation based on weakly supervised learning has attracted widespread attention and achieved significant research progress due to its low labeling cost. However, in real-world open scenarios for remote sensing image applications, image-level weak labeling of massive multi-category remote sensing image data still incurs considerable labeling costs. Simultaneously, due to data perturbations and the high labeling costs resulting from domain expertise, a large amount of relatively easily obtainable unlabeled remote sensing image sample data exists. However, most existing weak labeling methods ignore the enormous potential application value hidden within massive amounts of unlabeled remote sensing image data. Therefore, how to fully utilize the large amount of unlabeled data in open environments and effectively suppress the negative impact of anomalous samples is a crucial problem that urgently needs to be solved in practical applications.

[0003] Currently, only a few studies combine weakly labeled and unlabeled data for semantic segmentation. Li et al., in their paper "Weaklier Supervised Semantic Segmentation With Only One Image Level Annotation per Category," proposed a semantic segmentation method based on a weaker supervision mechanism. Their training data for each semantic category contains only one image-level weakly labeled data. To obtain valuable semantic representations from unlabeled samples, they primarily utilize unsupervised representation learning techniques to introduce prior information into the training of the weakly supervised network. However, the response activation clustering technique used to learn image features relies on high-dimensional clustering calculations, leading to significant computational overhead. In complex remote sensing scenarios, this further increases the training complexity of the model, thus limiting its practical application. How to efficiently improve the performance of semantic segmentation models in complex remote sensing scenarios while minimizing manual annotation costs (i.e., under limited annotation conditions) is one of the crucial problems urgently needing to be solved in the field of remote sensing image processing.

[0004] Unlike the traditional weak annotation methods used in existing weakly supervised learning, this invention constructs a novel weak annotation method for the first time. This can be viewed as a further weakened version of traditional weak annotation, comprising only a small number of real image-level weakly annotated samples and a large amount of unlabeled data. The aim of this invention is to generate pixel-level semantic segmentation prediction results with far less supervisory information than traditional image-level weak annotation, while minimizing annotation costs to the greatest extent possible while ensuring model segmentation accuracy.

[0005] Learning to map image-level target semantic category labels to pixel-level semantic predictions is an extremely challenging task when a large amount of training data lacks real image-level category labels. Therefore, the core of this invention is to fully mine and utilize valuable semantic knowledge from a large number of unlabeled samples, thereby promoting the learning of labeled sample branches. Based on the above analysis, the novel weakly labeled semantic segmentation task proposed in this invention differs significantly from existing weakly labeled semantic segmentation methods. Therefore, this invention proposes a novel dual-learning framework from the perspective of network structure, which can effectively utilize the rich semantic information of ground features in a large number of unlabeled samples through image-level and pixel-level semantic learning. With the assistance of semi-supervised learning, it effectively reduces the interference of noisy pseudo-supervisory signals, thereby achieving accurate and efficient segmentation of multi-category ground features in remote sensing images while further reducing labeling costs. Summary of the Invention

[0006] The purpose of this invention is to provide a semantic segmentation method for weakly labeled remote sensing images based on a dual learning mechanism. Compared with natural scene images, multi-class remote sensing images have rich ground feature information and complex background interference, thus their annotation relies heavily on expert knowledge. Meanwhile, in practical remote sensing applications, there is a large amount of relatively easily accessible unlabeled remote sensing image data. However, most existing methods ignore the potential of this abundant unlabeled remote sensing data to improve the performance of weakly labeled semantic segmentation models based on weakly supervised learning. Therefore, this invention focuses on how to fully utilize the rich semantic information in unlabeled samples in weakly supervised learning, that is, combining extremely limited samples with high-quality image-level annotations with a large number of unlabeled samples to further reduce the annotation cost of remote sensing image semantic segmentation tasks. Under the condition that annotation accuracy and annotation quantity are simultaneously limited, this effectively improves the performance of semantic segmentation models in multi-class remote sensing scenarios.

[0007] This invention mainly consists of three parts: image-level semantic learning based on semi-supervised classification training, a dataset re-segmentation strategy guided by a mask perturbation mechanism, and pseudo-supervised relearning based on dynamic thresholding. Specifically, in the training phase, firstly, a classification convolutional neural network based on semi-supervised learning is trained using a weakly labeled remote sensing image dataset consisting of only a small number of weakly labeled images and a large number of unlabeled images. A pixel-level pseudo-segmentation mask is generated through image-level semantic learning, thus achieving the first layer of semantic learning. Secondly, a dataset re-segmentation strategy guided by a mask perturbation mechanism is designed to re-segment the remote sensing image dataset into two subsets: labeled and unlabeled, based on the reliability of the pseudo-segmentation mask. Thirdly, the re-segmented dataset is used as input to train the semi-supervised semantic segmentation network for pixel-level semantic learning, thereby achieving the second layer of semantic learning. Finally, a dynamic thresholding strategy is proposed to adaptively adjust the confidence threshold during training, thereby reducing the confidence bias caused by noise interference in the artificial pseudo-labels generated by the network itself. In the testing phase, the semi-supervised semantic segmentation network is used to generate the final segmentation result of the remote sensing image to be tested, specifically including the following steps:

[0008] Step 1: Using a weakly labeled remote sensing image dataset consisting of a small amount of image-level labeled data and a large amount of unlabeled data, train a classification network based on semi-supervised learning to learn the image-level target semantic information in the remote sensing image dataset. Then, based on the classification prediction results of the trained classification network, assign image-level pseudo-class labels to each unlabeled sample, and output pixel-level pseudo-segmentation masks based on the pseudo-class labels and post-processing operations.

[0009] Step 2: Use the pseudo-segmentation mask obtained in Step 1 to perform pixel-level mask perturbation on the input image. Input the samples before and after perturbation into the trained classification convolutional neural network, calculate the similarity between the classification prediction results output by the two, and calculate the division threshold of each semantic category according to the preset ratio of reliable and unreliable samples. Combine the confidence score of the classification results to divide the segmentation pseudo-mask containing a large amount of semantic noise interference into two subsets: reliable pseudo-mask and unreliable pseudo-mask.

[0010] Step 3: Input the dataset obtained after the second partitioning in step 2 into the semantic segmentation network based on semi-supervised learning for training. Use a combination of supervised loss and unsupervised loss as the overall training loss for pixel-level semantic learning. Calculate the cross-entropy loss between the prediction results of labeled data and the labels as the supervised loss term. At the same time, calculate the cross-entropy loss between the strong perturbation prediction results and the weak perturbation prediction results of unlabeled data as the unsupervised loss term, and output the semantic segmentation results.

[0011] Step 4: Calculate the pixel-level confidence map of the segmentation result output by the semi-supervised semantic segmentation network in Step 3, and binarize the pixel-level confidence map by setting a dynamic threshold according to the training rounds. Use the obtained dynamic weight map as the pixel-level weight of the unsupervised loss term.

[0012] Step 5: After the semantic segmentation network based on semi-supervised learning is trained, in the testing phase, the remote sensing image to be tested is input into the trained network to obtain the final segmentation result of the remote sensing image to be tested. Attached Figure Description

[0013] Figure 1 This is a flowchart of the present invention.

[0014] Figure 2 Examples of remote sensing images used in this invention and their corresponding ground truths are shown. The first column is the original imagery, with the three images on the left from the ORSI-4199 test set and the three images on the right from the EORSSD test set. The second column is the ground truth.

[0015] Figure 3 and Figure 4 Examples of semantic segmentation results obtained using the method of this invention and other methods on the ORSI-4199 and EORSSD test sets are shown below. In the examples, (a) represents the original image, (b) represents the Ground Truth, and (c)-(f) represent the results obtained using SEAM, Puzzle-CAM, AMR, ReCAM, CPN, RCA, and the method of this invention, respectively. Specifically, SEAM is a weakly labeled semantic segmentation method based on the Self-Supervised Equivariant Attention Mechanism (SEAM), Puzzle-CAM is a class activation mapping method combined with the Puzzle module, AMR is a weakly supervised semantic segmentation method based on Activation Modulation and Recalibration (AMR), ReCAM is a weakly supervised segmentation method based on Class Re-Activation Map (ReCAM), CPN is a weakly supervised segmentation method based on Complementary Patch Network (CPN), and RCA is a method based on Regional Semantic Contrast and Aggregation (RCA). Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. The overall framework of the present invention is as follows: Figure 1 As shown, the implementation details of each step will now be introduced.

[0017] Step 1: Using a weakly labeled remote sensing image dataset consisting of a small amount of image-level labeled data and a large amount of unlabeled data, train a classification convolutional neural network based on semi-supervised learning, and calculate the pixel-level pseudo-segmentation mask of the training data.

[0018] The optical remote sensing image training dataset used in this invention This represents a dataset of N remote sensing images with dimensions H×W and 3 channels, where m represents the length of the image and n represents its width. The dataset contains N... s The images have weak image-level annotations, and the image-level category labels can be represented as y = [y1, y2, ..., y]. C ]∈{0,1} C Where C represents the total number of ground object categories in the training set; the dataset also contains N m The images are unlabeled samples, and N = N s +N m N m >>N s .

[0019] First, construct a classification convolutional neural network, denoted as . The system then trains a classification convolutional neural network and performs image-level semantic learning. The classification convolutional neural network, based on semi-supervised learning, is trained using loss functions designed for both labeled and unlabeled samples. Specifically, for samples with image-level annotations, the classification network... The standard cross-entropy loss between the predicted score and the true class label is the supervised loss term. Its form is:

[0020]

[0021] Where b = 1, ..., B, and B represents the batch size of the labeled training sample set.

[0022] For the unlabeled samples in the current training batch Batch Weak and strong data augmentation operations are performed separately to obtain two different transformed views of the same input sample. Specifically, geometric transformation operations such as flipping and rotating are used as weak augmentation operations (Aug). w (·), while strongly enhanced operation Aug s (·) includes color transformation, contrast enhancement, etc. For weakly enhanced views, if the model prediction result... If the highest predicted probability is greater than the threshold τ, then the corresponding sample is considered a reliable sample, and its prediction result is used as the pseudo-label of the strongly augmented view. Therefore, the cross-entropy loss is calculated between the model prediction result of the strongly augmented view and the reliable label obtained from the corresponding weakly augmented view, and used as the training loss for unlabeled samples. The formula is expressed as:

[0023]

[0024] Where μB represents the batch size of the unlabeled training set, and μ represents the ratio of unlabeled data to labeled data in the training set. This indicates an indicator function.

[0025] Then, the prediction results of the trained classification network are used as image-level pseudo-class labels for unlabeled samples, and pixel-level pseudo-segmentation masks are output through post-processing operations. Specifically, pixel-level foreground attention maps are generated using the Class Activation Mapping (CAM) algorithm. The classification feature maps output from the last convolutional layer are weighted and combined using the weights of the fully connected layer. Then, the corresponding background activation map is estimated. The specific calculation process is as follows:

[0026]

[0027] Where (x,y) represents the spatial coordinates of the pixel, and β is the attenuation factor for suppressing background label pixels and β>1.

[0028] The multi-channel foreground attention map and background activation map are concatenated, and then processed using Dense Conditional Random Field (DenseCRF) and Argmax operations to generate the final pixel-level segmentation pseudomask.

[0029] Y pseudo =Argmax(CRF([Bg(M :,h,w ),M :,h,w ]))

[0030] Step 2: Use the pseudo-segmentation mask obtained in Step 1 to perform pixel-level mask perturbation on the input image. Input the samples before and after perturbation into the trained classification convolutional neural network, calculate the similarity between the classification prediction results output by the two, and calculate the division threshold of each semantic category according to the preset ratio of reliable and unreliable samples. Combine the confidence score of the classification results to divide the segmentation pseudo-mask containing a large amount of semantic noise interference into two subsets: reliable pseudo-mask and unreliable pseudo-mask.

[0031] The class-aware dataset repartitioning strategy in this invention comprises three parts: 1) image-level similarity modeling based on mask perturbation, 2) class-aware confidence threshold calculation, and 3) dataset partitioning based on dual criteria of similarity and confidence.

[0032] According to pixel-level pseudomask Y Pseudo Obtain the binarized foreground mask M Fore This involves assigning a value of 1 to all regions in the pseudo-mask with non-zero pixel values, while assigning a value of 0 to all other regions. The calculation process is shown in the following formula:

[0033]

[0034] Next, M Fore A pixel-level multiplication operation is performed with the original input image I to obtain the pixel-level mask-perturbed sample, thereby eliminating the target region detected by the false mask in the input image. Finally, the images before and after perturbation are respectively input into the trained classification network. In the process, the similarity between the classification prediction scores output by the two is calculated:

[0035]

[0036] Where δ(·) represents the Softmax activation function and sim(·) represents the cosine similarity metric function.

[0037] Based on the similarity score of the pseudomask as the partitioning criterion, those with a similarity ρ greater than the upper bound threshold σ are considered. high or less than the lower bound threshold σ low The samples were initially assigned to the unlabeled dataset because they had a higher probability of being noise samples. Specifically, the indicator variable g was used as shown in the following formula. i To represent the input sample I i Whether it is assigned to an unlabeled subset: when g i A value of 1 indicates that the input sample is assigned to the unlabeled subset; conversely, if g i A value of 0 indicates that the input sample belongs to the labeled subset.

[0038]

[0039] The values ​​of the two threshold terms are set based on the ratio between reliable and unreliable subsets. Considering the issue of class balance, this invention proposes a class-aware partitioning threshold calculation strategy, which sets partitioning thresholds for reliable and unreliable pseudomasks for each target class, thereby ensuring that the proportion of labeled and unlabeled samples in each class remains relatively balanced in the repartitioned dataset.

[0040] Specifically, for semantic category c, the classification network in the dataset... The similarity scores ρ of all samples classified into this category are aggregated to obtain a similarity vector. Where N c This represents the total number of images that the classification network has determined to be category c. After sorting in descending and ascending order respectively, we get and Based on the reliable-unreliable ratio θ, the quantiles of the similarity vector are calculated and used as the upper and lower bound thresholds, respectively. The calculation process is shown in the following formula:

[0041]

[0042] Where Percentile(·) represents the quantile calculation function, and θ represents the ratio between labeled and unlabeled data after partitioning. Accordingly, the image samples with the highest unreliability percentage in the generated pixel-level pseudomask are assigned to the unlabeled dataset, while the remaining image samples are assigned to the labeled dataset.

[0043] After the initial segmentation based on the mask perturbation strategy, classification confidence scores are used to measure the certainty of the classification results. The segmentation results are further refined by dividing the segmentation pseudomasks containing significant semantic noise into two subsets: reliable pseudomasks and unreliable pseudomasks. In semi-supervised learning, the numerical vector (logits) output by the classification network is typically transformed using a softmax transform into a probability distribution for each class, and the highest probability value is used as the confidence score of the image classification result. Therefore, the classification confidence score is defined as follows:

[0044]

[0045] For g i For a sample with a value of 1, if the classification confidence s i If the confidence level is greater than the upper bound threshold, it indicates that the semantic category attribute has high reliability, and therefore it is classified into the reliable pseudo-mask subset and included in the labeled dataset; for g i The value is 0 and s i Samples with confidence scores below the lower bound threshold are likely to have inaccurate semantic category classifications, and therefore are classified into the unreliable pseudomask subset and included in the unlabeled dataset. In this chapter, based on experience, the upper bound threshold for confidence is set to 0.95, and the lower bound threshold for confidence is set to 0.75.

[0046] Step 3: Construct a semi-supervised semantic segmentation network. Input the re-segmented dataset into the semi-supervised semantic segmentation network for training. Use a combination of supervised and unsupervised loss as the overall training loss to perform pixel-level semantic learning and output the semantic segmentation results.

[0047] After the dataset is repartitioned, a semi-supervised segmentation network based on consistency learning is constructed. This semi-supervised semantic segmentation further extracts semantic information from the pseudo-supervised information. A combination of supervised and unsupervised losses is used as the overall training loss for pixel-level semantic learning, as shown in the following equation:

[0048] L = L s +λ u ·L u

[0049] Where, λ u This represents the weight of the unsupervised loss term.

[0050] Specifically, the prediction results M of the labeled data l With label P l The cross-entropy loss between the two is used as the supervision loss term, and the calculation process is shown in the following formula:

[0051]

[0052] Among them, B l This indicates the training batch size of the labeled dataset.

[0053] Simultaneously, calculate the strong perturbation prediction results for unlabeled data. With weak disturbance prediction results The cross-entropy loss between them is used as an unsupervised loss term, as shown in the following equation:

[0054]

[0055] Among them, B u represents the training batch size of the unlabeled dataset, and W represents the dynamic weight graph used for pseudo-label filtering. The calculation process will be explained in detail in step four.

[0056] Step 4: Using the pixel-level confidence map of the segmentation results, binarize the pixel-level confidence map by setting a dynamic threshold according to the training rounds, and use the resulting dynamic weight map as the pixel-level weights of the unsupervised loss term.

[0057] To minimize the impact of noisy pseudo-labels on the training of the semi-supervised segmentation network, a dynamic thresholding strategy is designed. By adaptively setting a dynamic threshold, more reliable predictions are selected from the output predictions of the weakly perturbed view as pseudo-supervisory signals for the strongly perturbed view. First, pixel-level confidence maps of the prediction results under the weakly perturbed view are calculated based on entropy.

[0058]

[0059] in, This represents the predicted output of a weakly perturbated view from a semi-supervised segmentation model. This represents the prediction result for the c-th channel. (i,j) represents the spatial coordinates of the pixel.

[0060] Then, the masks corresponding to pixel positions with confidence values ​​higher than the threshold in the segmentation masking results are defined as reliable pseudo-labels, while the masking results for the remaining pixel positions are considered unreliable masks and therefore cannot participate in network training. Based on the above analysis, as shown in the following formula, according to the dynamic threshold γ... t Binarize the pixel-level confidence map to obtain the dynamic weight map. This allows for pixel-level reweighting of the unsupervised loss term:

[0061]

[0062] Where, γ t This represents the dynamic threshold corresponding to the t-th training epoch, and its value gradually increases as training progresses. Specifically, in this invention, γ... t The settings are as follows:

[0063] γ t =Percentile(Flatten(Conf),α t )

[0064] Where Flatten(·) represents the operation of flattening a two-dimensional image into a one-dimensional vector, and α t = (100-η)·(1-t / N) epoch ), where η represents the initial proportion of pseudo-label pixels from unlabeled samples participating in network training. In this invention, η is set to 70.

[0065] Step 5: After training the network using the overall loss function L of the semantic segmentation network based on semi-supervised learning obtained in Step 4, input the remote sensing image to be tested into the trained network during the testing phase to obtain the final segmentation result of the remote sensing image to be tested.

[0066] The effectiveness of this invention can be further illustrated by the following experimental results and analysis:

[0067] 1. Experimental Data

[0068] This invention selects the publicly available multi-class remote sensing image geographic target datasets ORSI-4199 and EORSSD as experimental data. In the preprocessing stage, we resize these datasets to 256×256.

[0069] 2. Comparative Experiment

[0070] To demonstrate the necessity of the three-stage algorithm framework proposed in this invention, we compared it with two variants that have the same baseline model architecture and supervision conditions: a) directly using the pixel-level pseudomask generated in the first stage as the final result, denoted as "Ours-Stage1"; b) following the traditional weakly labeled segmentation method framework, using the pixel-level pseudomask obtained in the first stage to train the fully supervised segmentation network, denoted as "Ours-Stage1 w / Seg". Figure 2 A visual comparison of the segmentation results of the proposed method and the aforementioned variants is presented. In comparison, the segmentation results of the proposed method have clearer boundaries and better suppress complex background interference, demonstrating superior performance in accurate segmentation of small-scale targets and suppression of shadow interference.

[0071] To evaluate the performance of the method of this invention, we designed the following comparative experiment, selecting six existing representative semantic segmentation methods under the condition of weak annotation at the complete image level for performance comparison with the method of this invention. These include the Self-Supervised Equivariant Attention Mechanism (SEAM), the Class Activation Mapping (Puzzle-CAM) method combined with Puzzle modules, the Activation Modulation and Recalibration (AMR) method, the Class Reactivation Mapping (ReCAM) method, the Complementary Block Network (CPN) method, and the Region Contrast and Aggregation (RCA) method. The semantic segmentation results obtained by different methods were subjectively compared, such as... Figure 4 As shown.

[0072] The results demonstrate that the proposed method achieves superior semantic segmentation performance in remote sensing imagery scenes containing numerous multi-scale ground objects, accurately locating these objects while suppressing background interference. This method achieves high-quality segmentation results comparable to traditional weakly labeled methods with lower annotation costs. This superior performance is attributed to modules such as the class-aware dataset re-partitioning strategy and dynamic threshold-guided pseudo-supervised relearning, which fully utilize the rich semantic information in a large number of unlabeled remote sensing image samples, thereby further reducing the need for large amounts of precise pixel-level labeled data in deep learning-based remote sensing image semantic segmentation algorithms.

Claims

1. A semantic segmentation method for weakly labeled remote sensing images based on a dual learning mechanism. In the training phase, this method first trains a classification convolutional neural network based on semi-supervised learning using a weakly labeled remote sensing image dataset consisting of a small number of image-level weakly labeled samples and a large number of unlabeled samples. Pixel-level pseudo-segmentation masks are generated through image-level semantic learning, thus achieving the first layer of semantic learning. Second, a dataset re-partitioning strategy guided by a mask perturbation mechanism is designed. Based on the reliability of the pseudo-segmentation masks, the remote sensing image dataset is re-partitioned into two subsets: labeled and unlabeled. Third, the re-partitioned dataset is used as input to train the semi-supervised semantic segmentation network for pixel-level semantic learning, thus achieving the second layer of semantic learning. Finally, a dynamic thresholding strategy is proposed to adaptively adjust the confidence threshold during training, thereby reducing the confidence bias caused by noise interference in the artificial pseudo-labels generated by the network itself. In the testing phase, the semi-supervised semantic segmentation network is used to generate the final segmentation results of the remote sensing image to be tested. The method is characterized by… Includes the following steps: Step 1: Using a weakly labeled remote sensing image dataset consisting of a small amount of image-level labeled data and a large amount of unlabeled data, train a classification network based on semi-supervised learning to learn the image-level target semantic information in the remote sensing image dataset. Then, based on the classification prediction results of the trained classification network, assign image-level pseudo-class labels to each unlabeled sample, and output pixel-level pseudo-segmentation masks based on the pseudo-class labels and post-processing operations. Step 2: Use the pseudo-segmentation mask obtained in Step 1 to perform pixel-level mask perturbation on the input image. Input the samples before and after perturbation into the trained classification convolutional neural network, calculate the similarity between the classification prediction results output by the two, and calculate the division threshold of each semantic category according to the preset ratio of reliable and unreliable samples. Combine the confidence score of the classification results to divide the segmentation pseudo-mask containing a large amount of semantic noise interference into two subsets: reliable pseudo-mask and unreliable pseudo-mask. Step 3: Input the dataset obtained after the second partitioning in step 2 into the semantic segmentation network based on semi-supervised learning for training. Use a combination of supervised loss and unsupervised loss as the overall training loss for pixel-level semantic learning. Calculate the cross-entropy loss between the prediction results of labeled data and the labels as the supervised loss term. At the same time, calculate the cross-entropy loss between the strong perturbation prediction results and the weak perturbation prediction results of unlabeled data as the unsupervised loss term, and output the semantic segmentation results. Step 4: Calculate the pixel-level confidence map of the segmentation result output by the semi-supervised semantic segmentation network in Step 3, and binarize the pixel-level confidence map by setting a dynamic threshold according to the training rounds. Use the obtained dynamic weight map as the pixel-level weight of the unsupervised loss term. Step 5: After the semantic segmentation network based on semi-supervised learning is trained, in the testing phase, the remote sensing image to be tested is input into the trained network to obtain the final semantic segmentation result of the remote sensing image to be tested.

Images