A human-computer interactive method for surface cover classification without drawing polygons

Abstract

The application discloses a kind of man-machine interactive ground cover classification methods exempt from figure spot delineation, comprising the following steps: obtaining ground cover classification dataset for training;Respectively using conditional diffusion model and noise label classifier for training, the goal of training is all to carry out ground cover classification to given image;Wherein, the ground cover classification result that boundary is accurate but class contains part error is obtained after conditional diffusion model training, noise label classifier identifies class error after training;Conditional diffusion model updates model parameters according to MSE loss, and noise label classifier updates model parameters according to cross entropy loss;The result of conditional diffusion model is identified using trained noise label classifier to noise label.This application inputs remote sensing image as condition information, interacts image information with denoising information;Through controllable engineering efficiency misclassification prompter, automatically find possible error part ground cover data.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a human-computer interactive land cover classification method that eliminates the need for map patch delineation. Background Technology

[0002] Land cover classification is a core and crucial component of remote sensing image applications. High-precision land cover classification can clearly distinguish various land types, such as agricultural land, forests, water bodies, and construction land, providing policymakers with a scientific basis and having profound significance for rational resource allocation, ecological environmental protection, and urban planning. In recent years, the development of artificial intelligence technology, especially deep learning, has achieved impressive results in intelligent land cover classification, gradually shifting from traditional manual visual classification to semi-automatic classification methods assisted by deep learning.

[0003] It is worth noting that the accuracy of boundary information is crucial for practical applications in land cover classification. Taking urban planning and mapping as an example, the accuracy of boundary cell classification directly affects the determination of building edges, which in turn affects the measurement of plot area, the assessment of land use change, and the planning and construction of municipal facilities. Furthermore, accurate boundary delineation is directly related to the quantitative assessment of key environmental change parameters such as deforestation and wetland degradation, providing a scientific basis for ecological monitoring, disaster early warning, and natural resource management.

[0004] General solutions for land cover classification based on deep learning focus on learning dense representations of remote sensing images to capture contextual information within the spatial neighborhood and classify land cover for each pixel in the image. In this field, discriminative classification strategies dominate, directly mapping input pixel features to the probability distribution of the corresponding land cover category by employing a cross-entropy loss function to optimize the network model, emphasizing the learning of the posterior probability of land cover features. Currently, most research on land cover classification focuses on developing more efficient dense representation techniques that can express rich contextual information: models such as the DeepLab series effectively expand the receptive field of the network by introducing dilated convolutions, capturing broader contextual information without increasing computational complexity. Furthermore, to better model long-range spatial dependencies, some methods incorporate dense conditional random fields (CRFs) for post-processing after pixel-level classification to optimize classification boundaries and enhance the accuracy and continuity of classification. Models such as Deeplabv3+, DFN (Decoupled Feature Network), and PSPNet (Pyramid Scene Parsing Network) further enhance their ability to understand complex landform structures by fusing feature information at different scales, particularly excelling in handling land cover edges and mixed landform types. Recently, with the groundbreaking achievements of the Transformer architecture in natural language processing and computer vision, its powerful global feature encoding capabilities have attracted considerable attention from researchers. Numerous studies are actively exploring new approaches that combine the Transformer mechanism with multi-scale feature encoding strategies, aiming to enhance the model's understanding and classification performance of complex land cover scenes through this long-distance attention mechanism across pixels, thus propelling land cover classification technology to new heights.

[0005] While existing land cover classification methods have significantly improved the classification accuracy of internal region pixels by integrating long-range contextual information, these technologies generally fail to address the problem of accurate boundary classification. In the actual production of land cover products, due to the limitations of existing deep learning-based land cover classification methods, such as inaccurate boundary alignment and lack of indication of potential misclassification, the typical approach is to use the results predicted by the deep learning model as a preliminary reference base map to provide category-guided auxiliary information for manual delineation of land cover features, rather than using it directly as the final product. Specifically, in the actual operational phase, manual delineation of land cover features still occupies the majority of the time in the entire production process. Cartographers need to redefine the boundaries of various land cover features through meticulous manual delineation based on the classification results output by the model, ensuring high accuracy and reliability of land cover classification. Furthermore, deep learning technologies cannot guarantee 100% prediction accuracy, so the results still require manual review and error correction. This process involves a significant amount of manual labor, including but not limited to checking the entire area of ​​the model's prediction results, error correction, detailed differentiation of uncertain areas, and manual fine-tuning of vector land cover features based on field conditions and professional knowledge. While model predictions alleviate the workload to some extent, such as by providing suggestions on land cover categories through pre-classification, their auxiliary role is still relatively small compared to the workload of manually drawing and inspecting land cover plots.

[0006] Compared to existing land cover classification methods, generative classification methods aim to construct a joint distribution model of remote sensing data, and then use Bayes' theorem to infer the posterior probability distribution of land cover categories. Theoretical research suggests that generative classification, by directly modeling the original features of the image, has the potential to overcome the limitations of existing techniques that overemphasize overall classification accuracy while neglecting local detail identification. In recent years, diffusion probability models have achieved significant progress in image generation. These models introduce Gaussian noise into the image data through a carefully designed diffusion process until the image is almost entirely transformed into Gaussian noise. Then, in the reverse denoising process, the model gradually eliminates this noise, thus "regenerating" a clear image. Research indicates that in the early stages of reverse diffusion, especially when dealing with highly noisy images, the model consistently demonstrates effective recovery capabilities for high-frequency land cover features (such as feature edges), which is key to gradually extracting clear feature outlines from a chaotic, noisy background. Therefore, in the practical application of land cover classification, considering that the time required for manually delineating land cover patches is much longer than the time required for determining land cover categories, using generative models for land cover classification may improve the efficiency of land cover data production.

[0007] Furthermore, advances in noise learning have made it possible for machines to automatically identify mislabeled data, potentially aiding in manual review and correction processes. Confidence learning, in particular, quantifies the model's level of confidence in each prediction, providing users with immediate feedback on the reliability of classification decisions. By integrating confidence indicators into human-computer interaction, the system can automatically identify cases with low prediction confidence and a high risk of error, guiding users to focus their review and effectively reducing unnecessary comprehensive review workload. This makes interactive human-computer review and correction solutions feasible.

[0008] Based on the current technological research background, the following problems need to be solved for the production method of land cover products assisted by deep learning technology: (1) Existing deep learning models cannot accurately predict the boundaries of land features, making it difficult to meet the requirements of high-precision annotation, and manual delineation of land cover patches is required in production. (2) Deep learning technology cannot guarantee 100% accuracy in land cover classification, which requires manual inspection and error correction in the product production process. Since this process lacks indication of the possibility of misclassification, all data must be checked, resulting in low production efficiency. Summary of the Invention

[0009] In view of this, this invention is the first to use a diffusion model to model and learn human-annotated data, thereby generating data that conforms to human annotation paradigms. Furthermore, it constructs a noise label recognition based on confidence learning for prompting annotation, thus enabling land cover classification without the need for patch delineation.

[0010] This application discloses a human-computer interactive land cover classification method that eliminates the need for land parcel delineation, comprising the following steps:

[0011] Obtain the land cover classification dataset D = {(X i ,Y i )*N}, where X i Representative image, Y i The labels are represented by N, and the number of samples is represented by N.

[0012] Using the conditional diffusion model M respectively c-diff Both the conditional diffusion model and the noise label classifier are trained to classify land cover in a given image. The conditional diffusion model, after training, yields land cover classification results with accurate boundaries but containing some incorrect categories. The noise label classifier, after training, identifies the category errors. The conditional diffusion model updates its parameters based on MSE loss, and the noise label classifier updates its parameters based on cross-entropy loss.

[0013] The trained noise label classifier is used to identify noise labels on the results of the conditional diffusion model, including:

[0014] Input the image X to be classified i Engineering efficiency H;

[0015] Prediction results chart and predictive noise NL i ;

[0016] Calculate the number M of misclassified patches. blk Calculate the predicted noise NL in each patch. i Pixel ratio R nl And sort;

[0017] Select the top M based on the sorting results. blk Each patch yields NL blk ;

[0018] Output land cover classification result map And misclassification prompt results NL blk .

[0019] Furthermore, the conditional diffusion model M c-diff The structure is divided into two parts, namely condition module B. condition Noise reduction module B denoise Image X i Input Condition Module B condition The output conditional information C xi Input to noise reduction module B denoise ; Tag Y i Add Gaussian noise signal S i The noise map Y was then obtained. noise-i Noise map input denoising module B denoise Output predicted noise The predicted noise and the noise added by the labels are used to calculate the MSE loss. The overall process is as follows:

[0020]

[0021] Furthermore, condition module B condition Includes an encoder block E condition And an interaction block, which includes L convolutional layers. and L cross-attention layers L represents the number of residual layers in the encoder, Conv represents a convolutional layer, and CrossAttn represents a cross-attention layer.

[0022] Furthermore, encoder block E condition Output the residual features of each layer:

[0023]

[0024] hl, wl, and cl represent the height, width, and dimension of the feature map at layer l, respectively.

[0025] The convolutional layer outputs the residual features from each layer of the encoder after convolution.

[0026]

[0027] The cross-attention layer passes the residual of each layer through the denoising module encoder. Query the residuals of each layer of the convolutional encoder and output the query results:

[0028]

[0029] The convolution result is added to the query result and then input into the corresponding position of each layer of the denoising module decoder:

[0030]

[0031] Furthermore, the denoising module is primarily a UNET structure, consistent with the stable denoising model, and the noise map Y is used during training. noise-i Input unet, and then condition information C. xi Noise predicted by the output after input control

[0032]

[0033] The MSE loss is calculated using the noise predicted during training and the noise added by the labels.

[0034]

[0035] During prediction, the input in the first step is a pure noise map Y. t Output prediction noise Then, the pure noise graph Y t Denoising is performed to obtain the denoised image Y. t-1 Then Y t-1 This step is repeated as input, and the result image is obtained after t iterations of noise reduction.

[0036]

[0037] Furthermore, the misclassification indicator M cl It includes a discriminative semantic segmentation module B. m A noise tag recognition module B cl B m The discriminative semantic segmentation model deeplabv3+ was used.

[0038] Furthermore, the training process for the misclassification cue includes:

[0039] First, use image X. i Labels predicted by conditional diffusion model Training Module B m Using cross-entropy loss L CE as follows:

[0040]

[0041] Image X i Labels predicted by conditional diffusion model Input the trained module B m In the middle, the output gets the label. The corresponding confidence level C yi :

[0042] C yi =B m (X i )

[0043] Confidence level C yi and tags Then input the noise tag recognition module B. cl Obtain the noise label prediction result NL i ; of which B cl A confidence learning strategy is used, which identifies noise labels by constructing a noise transition matrix based on multi-class confidence and calculating outliers in the noise transition matrix.

[0044]

[0045] "Clean" indicates a correct label, and "noise" indicates a noise label.

[0046] Furthermore, for a given engineering efficiency H, the total number of patches in the dataset is N. blk The number of misclassified patches is M. blk The method for calculating engineering efficiency is as follows:

[0047]

[0048] In order to control the efficiency of the project by controlling the number of misclassified patches, based on the project efficiency H and the total number of patches N... blk Substitute into the above formula to calculate the number M of misclassified patches. blk ;Predict the noise label NL in each patch i Pixel ratio R nl Based on this, the map patches are sorted from largest to smallest, and the top M patches are selected. blkThe output result NL is a misclassification warning for each patch. blk :

[0049]

[0050] The number of pixels in the patch is P. blk .

[0051] The beneficial effects of this invention are as follows:

[0052] This invention designs a conditional diffusion model suitable for land cover classification. This model can take remote sensing imagery as input as conditional information and interact with image information and denoising information to generate land cover classification results with accurate boundaries, thereby eliminating the need for manual delineation of patches and greatly improving the production efficiency of land cover classification.

[0053] This invention designs a misclassification prompter with controllable engineering efficiency, which can provide misclassification prompts during manual inspection and correction. It automatically finds potentially erroneous land cover data based on engineering efficiency, thereby reducing the amount of data that needs to be manually checked. It can complete the production of land cover classification data according to requirements and under specified engineering efficiency. Attached Figure Description

[0054] Figure 1 Framework diagram of the present invention;

[0055] Figure 2 Conditional diffusion model M c-diff Structural diagram;

[0056] Figure 3 Condition Module B condition Structural diagram;

[0057] Figure 4 Noise reduction module B denoise Structural diagram;

[0058] Figure 5 Flowchart of misclassification alert;

[0059] Figure 6 A schematic diagram comparing the boundary conditions of the conditional diffusion model results;

[0060] Figure 7 A diagram illustrating the results of the misclassification alert. Detailed Implementation

[0061] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0062] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0063] The technical solutions provided in this application relate to machine learning technology in artificial intelligence, and are specifically described and illustrated through the following embodiments.

[0064] The framework proposed in this invention includes a conditional diffusion model M. c-diff A misclassification indicator M cl The overall process is as follows Figure 1 As shown. Given a land cover classification dataset D = {(X... i ,Y i )*N}, where X i Representative image, Y i Let N represent the number of samples and N represent the number of labels. This invention first trains both the conditional diffusion model and the noise label classifier, with the goal of classifying land cover in a given image. The conditional diffusion model, after training, yields land cover classification results with accurate boundaries but containing some incorrect categories. The noise label classifier, after training, can identify these category errors. Finally, the trained noise label classifier is used to identify noise labels on the results of the conditional diffusion model. The following sections will first explain the structure and training method of the conditional diffusion model, and then introduce the composition of the noise label classifier and the noise label identification method.

[0065] (1) Generative land cover classification with accurate boundaries

[0066] The conditional diffusion model M used in this invention c-diff Structure as Figure 2 As shown, the model is divided into two parts, namely condition module B. condition Noise reduction module B denoise Image X i Input Condition Module B condition The output conditional information C xi Input to noise reduction module B denoise ; Tag Y i Add Gaussian noise S i The noise map Y was then obtained. noise-i Noise map input denoising module B denoise Output predicted noise The predicted noise and the noise added by the labels are used to calculate the MSE loss. The overall process can be expressed as follows:

[0067]

[0068] The specific structure of each module is as follows:

[0069] Condition Module B condition Structure as Figure 3 As shown:

[0070] Condition Module B condition Includes an encoder block E condition (Same structure as the encoder part of the denoising module) and an interaction block, which includes L convolutional layers. and L cross-attention layers L represents the number of residual layers in the encoder.

[0071] encoder blockE condition Output the residual features of each layer:

[0072]

[0073] The convolutional layer outputs the residual features from each layer of the encoder after convolution.

[0074]

[0075] The cross-attention layer passes the residual of each layer through the denoising module encoder. The residuals of each layer of the convolutional encoder are queried, and the query results are output.

[0076]

[0077] Finally, the convolution result is added to the query result and input into the corresponding position of each layer of the denoising module decoder.

[0078]

[0079] Noise reduction module B denoise Structure as Figure 4 As shown:

[0080] The denoising module is primarily a UNET structure, consistent with the stable denoising model. During training, the noise map Y... noise-i Input unet, and then condition information C. xi Noise predicted by the output after input control

[0081]

[0082] The MSE loss is calculated using the noise predicted during training and the noise added by the labels.

[0083]

[0084] During prediction, the input in the first step is a pure noise map Y. t Output prediction noise Then, the pure noise graph Y t Denoising is performed to obtain the denoised image Y. t-1 Then Y t-1 This step is repeated as input, and the result image is obtained after t iterations of noise reduction.

[0085]

[0086] (2) Misclassification indicator for controllable engineering efficiency

[0087] The process of misclassification prompt in this invention is as follows: Figure 5 As shown:

[0088] Training Tips M cl

[0089] Misclassification indicator M cl It includes a discriminative semantic segmentation module B. m A noise tag recognition module B cl Among them, B m Any existing discriminative semantic segmentation model can be used; this invention employs the common deeplabv3+ model. First, image X is used... i Labels predicted by conditional diffusion model Training Module B m Using cross-entropy loss L CE as follows:

[0090]

[0091] Image X i Labels predicted by conditional diffusion model Input the trained module B m In the middle, the output gets the label. The corresponding confidence level C yi .

[0092] C yi =B m (X i (10)

[0093] Confidence level C yi and tags Then input the noise tag recognition module B. cl Obtain the noise label prediction result NL i Among them, B cl This invention employs a confidence learning strategy, which identifies noise labels by constructing a noise transition matrix based on multi-class confidence and calculating outliers in the matrix. The prediction method using pure confidence ranking in this strategy is described in this invention.

[0094]

[0095] Misclassified patch sorting and controllable engineering efficiency

[0096] For a given engineering efficiency H (times), assuming the total number of patches in the dataset is N... blk The number of misclassified patches is M. blk The method for calculating engineering efficiency is as follows:

[0097]

[0098] Therefore, this invention achieves the goal of controlling engineering efficiency by controlling the number of misclassified patches, which only requires determining the engineering efficiency H and the total number of patches N. blk Substituting the values ​​into the above formula will allow you to calculate the number of misclassified patches, M. blk Predict the noise label NL in each patch. i Pixel ratio R nl (Let the number of pixels in this patch be P) blk Sort the map patches from largest to smallest based on the following criteria, and select the top M. blk The output result NL is a misclassification warning for each patch. blk .

[0099]

[0100] Model training steps:

[0101] Input: Land cover classification dataset D = {(X i ,Y i )*N}

[0102] Output: Conditional Diffusion Model M c-diff ,

[0103] Step 1: Start iterative training of the conditional diffusion model M c-diff

[0104] Step 2: Update M using formula (7) c-diff parameters

[0105] Step 3: End the loop training M c-diff

[0106] Step 4: Start loop training of the misclassification cue M cl

[0107] Step 5: Update M using formula (9) cl parameters

[0108] Step 6: End the loop training M cl

[0109] Result prediction steps:

[0110] Input: Image X to be classified i Engineering efficiency H

[0111] Output: Land cover classification result map Misclassification result NL blk

[0112] Step 1: Predict using formula (2,3,4,5,6,8) Where t is 50

[0113] Step 2: Predict NL using formula (10,11) i

[0114] Step 3: Calculate M using formula (12) blk

[0115] Step 4: Calculate R using formula (13) nl and sort

[0116] Step 5: Select the top M based on the sorting results blk Each patch yields NL blk

[0117] The accurate generative land cover results of this invention are as follows: Figure 6 As shown in Table 1, this effect is generated by the generative conditional diffusion model in this invention. The conditional module in the conditional diffusion model can extract information from the image and guide the inference process of the denoising model; while the denoising model learns the distribution pattern of land cover classification labels, thus generating land cover classification results with accurate boundaries. To more intuitively demonstrate the accuracy of the boundaries in the land cover results, Figure 6 The boundary regions of the results from different methods are highlighted, and the boundary accuracy is calculated using the Weighted F-measure method. Figure 6 As shown in Table 1, the present invention has a significant advantage in boundary generation, meets the requirements of human annotation, and can achieve the effect of eliminating the need for manual delineation of map features.

[0118] Table 1 Comparison of Land Cover Classification Boundary Accuracy

[0119]

[0120]

[0121] The results of the misclassification indicator for controllable engineering efficiency are as follows: Figure 7 As shown: This effect is generated by the misclassification prompt in this invention, wherein the misclassification result prompted by this invention is basically consistent with the actual situation, and the efficiency of the project can be accelerated by adjusting the parameters.

[0122] The beneficial effects of this invention are as follows:

[0123] This invention designs a conditional diffusion model suitable for land cover classification. This model can take remote sensing imagery as input as conditional information and interact with image information and denoising information to generate land cover classification results with accurate boundaries, thereby eliminating the need for manual delineation of patches and greatly improving the production efficiency of land cover classification.

[0124] This invention designs a misclassification prompter with controllable engineering efficiency, which can provide misclassification prompts during manual inspection and correction. It automatically finds potentially erroneous land cover data based on engineering efficiency, thereby reducing the amount of data that needs to be manually checked. It can complete the production of land cover classification data according to requirements and under specified engineering efficiency.

[0125] As used herein, the term "preferred" is meant as an example, illustration, or illustration. Any aspect or design described herein as "preferred" need not be construed as being more advantageous than other aspects or designs. Rather, the use of the term "preferred" is intended to present the concept in a specific manner. As used in this application, the term "or" is intended to mean an inclusive "or" rather than an exclusionary "or." That is, unless otherwise specified or clear from the context, "X uses A or B" naturally includes either of the permutations. That is, if X uses A; X uses B; or X uses both A and B, then "X uses A or B" is satisfied in any of the foregoing examples.

[0126] Furthermore, although this disclosure has been shown and described with respect to one or more implementations, equivalent variations and modifications will occur to those skilled in the art based on a reading and understanding of this specification and the accompanying drawings. This disclosure includes all such modifications and variations and is limited only by the scope of the appended claims. In particular, with respect to the various functions performed by the aforementioned components (e.g., elements, etc.), the terminology used to describe such components is intended to correspond to any component (unless otherwise indicated) that performs the specified function of said component (e.g., is functionally equivalent to it), even if structurally not equivalent to the disclosed structure performing the functions in the exemplary implementations of this disclosure shown herein. Moreover, although specific features of this disclosure have been disclosed with respect to only one of several implementations, such features may be combined with one or more features of other implementations that may be desirable and advantageous for a given or particular application. Furthermore, with regard to the use of the terms “comprising,” “having,” “containing,” or variations thereof in the Detailed Description or claims, such terms are intended to be included in a manner similar to the term “including.”

[0127] The functional units in this invention embodiment can be integrated into a processing module, or each unit can exist physically separately, or multiple units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. The aforementioned devices or systems can execute the storage methods in the corresponding method embodiments.

[0128] In summary, the above embodiments are one implementation of the present invention, but the implementation of the present invention is not limited to the embodiments described above. Any changes, modifications, substitutions, combinations, or simplifications made that deviate from the spirit and principle of the present invention should be considered equivalent substitutions and are included within the protection scope of the present invention.

Claims

1. A human-computer interactive land cover classification method that eliminates the need for patch delineation, characterized in that, Includes the following steps: Obtain the land cover classification dataset for training. ,in Representative images, Represents a label. Represents the number of samples; Using conditional diffusion models respectively The conditional diffusion model and the misclassification cue are trained together, with the goal of classifying land cover for a given image. The conditional diffusion model, after training, yields land cover classification results with accurate boundaries but containing some incorrect categories. The misclassification cue identifies category errors after training. The conditional diffusion model updates its parameters based on MSE loss, and the misclassification cue updates its parameters based on cross-entropy loss. The trained misclassification cue is used to identify noise labels in the results of the conditional diffusion model, including: Input the image to be classified. Engineering efficiency ; Prediction results chart and prediction noise ; Calculate the number of misclassified patches. Calculate the prediction noise in each patch. pixel ratio And sort; Select the previous results based on the sorting. Each patch was obtained ; Output land cover classification result map And misclassification prompt results ; Wherein, the conditional diffusion model The structure is divided into two parts: condition modules. Noise reduction module Among them, images Input condition module Output condition information Input to noise reduction module ;Label Add Gaussian noise signal The noise map was then obtained. Noise denoising module for noise map input Output predicted noise The predicted noise and the noise added by the labels are used to calculate the MSE loss. The overall process is as follows: ； Conditional module Includes an encoder block And an interaction block, which includes L convolutional layers. and L cross-attention layers L is the number of residual layers in the encoder. Indicates a convolutional layer. Indicates a cross-attention layer; The denoising module is primarily based on a UET structure, consistent with the stable denoising model. For the specified engineering efficiency The total number of patches in the dataset is The number of misclassified patches is The method for calculating engineering efficiency is as follows: ； The misclassification alert It includes a discriminative semantic segmentation module. A noise tag recognition module ,in The discriminative semantic segmentation model deeplabv3+ was used.

2. The human-computer interactive land cover classification method without map patch delineation according to claim 1, characterized in that, encoder block Output the residual features of each layer: ； They represent the first The height, width, and dimension of the layer feature map; The convolutional layer outputs the residual features from each layer of the encoder after convolution. ； The cross-attention layer passes the residual of each layer through the denoising module encoder. Query the residuals of each layer of the convolutional encoder and output the query results: ； The convolution result is added to the query result and then input into the corresponding position of each layer of the denoising module decoder: 。 3. The human-computer interactive land cover classification method without map patch delineation according to claim 2, characterized in that, Training noise graph Enter unet, and then proceed with the conditional information. Noise predicted by the output after input control ; ； The MSE loss is calculated using the noise predicted during training and the noise added by the labels. ； During prediction, the input to the unet is a pure noise map in the first step. Output prediction noise Later, pure noise graph Denoising is performed to obtain a denoised image. Next This step is repeated as input, and the result image is obtained after t iterations of noise reduction. ; 。 4. The human-computer interactive land cover classification method without map patch delineation according to claim 3, characterized in that, The training process for the misclassification prompter includes: First, use images. Labels predicted by conditional diffusion model Training module Using cross-entropy loss as follows: ； image Labels predicted by conditional diffusion model Input the trained module In the middle, the output gets the label. Corresponding confidence level : ； Confidence and tags Then input the noise tag recognition module Obtain noise label prediction results ;in A confidence learning strategy is used, which identifies noise labels by constructing a noise transition matrix based on multi-class confidence and calculating outliers in the noise transition matrix. ； Indicates the correct label. Indicates a noise label.

5. The human-computer interactive land cover classification method without map patch delineation according to claim 4, characterized in that, In order to control the number of misclassified patches and thus control project efficiency, based on project efficiency... and the total number of map spots Substitute the data into the engineering efficiency calculation formula to calculate the number of misclassified patches. Predict noise labels for each patch. pixel ratio Based on this, the map patches are sorted from largest to smallest, and the top ones are selected. The output result is a misclassification warning for each image patch. : ； The number of pixels in the patch is .

Images