A method for generating land cover classification samples

Abstract

A method for generating land cover samples comprises: determining i prior products of a remote sensing image; orthogonalizing and fusing the i prior products to obtain comprehensive probabilities of each category of each pixel of the remote sensing image; determining the maximum value of the comprehensive probability of each category of each pixel, and determining the land cover type of each pixel of the remote sensing image according to the maximum value of the comprehensive probability; combining the maximum value of the comprehensive probability of each pixel of the remote sensing image to obtain a confidence map, and combining the land cover type of each pixel of the remote sensing image to obtain a land cover map; screening out and removing the land cover types corresponding to the pixels whose values ​​in the land cover map are lower than a set first threshold according to the confidence map, and obtaining a land cover map with a credibility that meets the requirements; the land cover map with a credibility that meets the requirements is the initial label of the remote sensing image; and cutting the remote sensing image and the initial label into a plurality of slice images and corresponding slice labels to obtain a land cover classification sample.

Description

Technical Field

[0001] The present application relates to the field of geographic information system application technology, and in particular to a method for generating land cover classification samples. Background Art

[0002] With the continuous advancement of global remote sensing technology, land cover maps have become core data for national land resource surveys, spatial planning, and many other areas closely related to sustainable development. These images not only provide detailed information on various types of surface cover, but also provide key data support for government decision-making, environmental protection, urban planning, and agricultural management. Especially when modern remote sensing platforms can provide high-spatial-resolution images, the monitoring and classification accuracy of land cover has been unprecedentedly improved, so that the categories of surface objects can be captured more finely. Traditional land cover classification methods, such as pixel-based or object-based classification, often rely on manually designed features and the selection of appropriate classifiers. This not only requires a high degree of expertise, but is also time-consuming and labor-intensive, especially in large-scale land cover classification tasks, where its efficiency and accuracy are challenged. With the rise of deep learning technology, especially the widespread application of convolutional neural networks, new opportunities have been brought to land cover classification. Convolutional neural networks have the ability to automatically learn and extract features, and can automatically identify key features in complex remote sensing images, thereby greatly improving the accuracy and efficiency of classification.

[0003] However, the training of deep learning models requires a large amount of labeled data, especially pixel-by-pixel annotation. This level of annotation not only requires a lot of time and manpower, but also requires a high degree of expertise. Therefore, in a large range of land cover classification, high-quality labeled data is often very scarce. This has become the main bottleneck limiting the widespread application of deep learning in land cover classification.

[0004] Some studies have begun to explore the use of existing public land cover products to automatically generate labeled data. These products are usually obtained by converting publicly available land cover classification maps, which, although there are certain errors, can still be used as preliminary annotations for training deep learning models. These match low-resolution product data with high-resolution remote sensing images to generate a large number of training labels. This method alleviates the problem of insufficient labeled data to a certain extent, but there are still many potential risks and challenges.

[0005] First, due to the resolution mismatch, objects in low-resolution products may not be continuous or completely missing in high-resolution images. This will result in a large amount of inaccurate information in the generated annotation data, which will affect the training effect of the model. Secondly, due to time and classification errors, there are often semantic errors in existing public land cover products. These errors will be learned by the model during the training process, resulting in difficulties in the convergence of the classifier and even affecting the generalization ability of the model. In addition, due to the time differences and inconsistent classification standards between different products, directly using these products as training labels will also introduce additional noise and bias.

[0006] How to more reasonably utilize existing products across resolutions and time periods to generate low-cost, high-quality annotated data has become an urgent problem to be solved in the current land cover classification field. This is not only related to the effect of model training, but also directly affects the accuracy of updating and application of large-scale land cover maps. Summary of the invention

[0007] In order to solve the above problems, an embodiment of the present application provides a method for generating land cover samples, the method comprising: determining i prior products of a remote sensing image; the prior products comprising a classification map obtained by a publicly available classification method and an accuracy evaluation result of the classification map; orthogonalizing and fusing the i prior products to obtain a comprehensive probability of each category for each pixel in the remote sensing image; determining the maximum value of the comprehensive probability of each category for each pixel, and determining the land cover type of each pixel of the remote sensing image according to the maximum value of the comprehensive probability of each category for each pixel; combining the maximum value p(T p ) to obtain a confidence map, and combine the land cover type of each pixel of the remote sensing image to obtain a land cover map; each pixel of the confidence map, each pixel of the land cover map and each pixel of the remote sensing image correspond one to one in spatial position; according to the confidence map, p(T p ) values ​​lower than the set first threshold value corresponding to the land cover types of pixels are screened out and removed to obtain a land cover map with credibility that meets the requirements; the land cover map with credibility that meets the requirements is the initial label of the remote sensing image; the remote sensing image and the initial label are cropped into multiple slice images and corresponding slice labels to obtain the initial sample of land cover classification.

[0008] In some embodiments, i prior products are orthogonalized and fused to obtain the comprehensive probability of each category of each pixel of the remote sensing image, including: assigning the basic probability of the target land cover type of corresponding product 1, product 2, ... product i as p1(T j )、p2(T j )……p i (Tj ), where T j is the jth land cover type, j is a natural number; i target land cover type T j The basic probabilities are orthogonalized and fused to obtain the comprehensive probability p of each pixel in each category. i (T j ), the formula is as follows:

[0009]

[0010]

[0011] In the formula, represents the orthogonal sum, p1(T j )、p2(T j )……p i (T j ) are i prior products for the target land cover type T j The basic probability assignment is, k is the conflict coefficient, the larger the value of the conflict coefficient k is, the more difficult it is to combine the prior products, and the smaller the value of k is, the easier it is to combine the products.

[0012] In some embodiments, the labels in the initial sample of land cover classification contain noise, and the method further includes: using two probability-based uncertainty principles to filter the labels in the land cover classification samples to contain noise, including: predicting the class probability distribution of each pixel based on each slice image in the initial sample of land cover classification and the deep learning model; determining the maximum probability value and the second largest probability value of each pixel based on the class probability distribution of each pixel; calculating the confidence based on the maximum probability value and the second largest probability value; the confidence includes a credibility U1 and a credibility difference U2:

[0013]

[0014]

[0015] In the formula, is the category with the largest confidence probability for pixel X, and U1 is the category corresponding to the maximum confidence probability for pixel X The larger the U1, the greater the credibility of the category corresponding to the maximum confidence probability of the pixel X; is the category with the second largest confidence probability of pixel X; U2 is and The difference in credibility between the two categories, U2, reflects the similarity between the two categories; the noise of the label in the initial sample is filtered according to the credibility U1, the credibility difference U2 and the set second threshold, and the original label corresponding to the pixel is updated; the second threshold includes a credibility threshold and a similarity threshold.

[0016] In some embodiments, filtering the noise of the label in the initial sample and updating the original label corresponding to the pixel are determined according to the credibility U1, the credibility difference U2 and the set second threshold, including: taking the average value of U1 and U2 of each batch of slice images as the updated second threshold; truncating the second threshold to obtain two thresholds for each batch of training samples:

[0017]

[0018]

[0019] If U1 of each pixel in the current slice image is greater than φ1 and U2 is greater than φ2, the category of the position corresponding to the predicted label predicted by the model under the current slice image is regarded as the correct category; if the slice label corresponding to the current slice image is inconsistent with the correct predicted label, the position corresponding to the current slice label is regarded as noise, and the category of the position corresponding to the current slice label is corrected to the category of the position corresponding to the correct predicted label.

[0020] In some embodiments, the method further includes: filtering the noise of the labels in the initial sample of the land cover classification based on the principle of the normalized vegetation index NDVI, including: calculating the value of the normalized vegetation index NDVI corresponding to each pixel in the slice image, retaining the current slice label when the NDVI value of the current pixel in the slice image meets the set third threshold, and changing the value of the current pixel in the slice image to the background value when the NDVI value of the current pixel in the slice image does not meet the third threshold.

[0021] In some embodiments, the method further includes: for the initial training set, updating and correcting the noise of the labels in the samples of the land cover classification through deep learning model training, including: in the first stage, using the slice images and corresponding slice labels of the initial samples to train the deep learning model, iteratively training the deep learning model with a larger initial learning rate; when the minimum training loss does not decrease for M consecutive rounds, scaling the learning rate to 0.1 times the previous one; selecting the weight of the M+1th round before the current round as the weight of the deep learning model in the second stage, and optimizing the deep learning model; in the second stage, using the initial samples to adopt the joint loss function L noise After training and optimization, the deep learning model is trained to iteratively correct and update the noise of the labels in the initial samples based on the confidence probability predicted by the deep learning model during the training process; where L noise is the calculation result of the loss of the initial label and the updated label with the model prediction probability.

[0022] In some embodiments, the joint loss function L noise for:

[0023]

[0024] Among them, Y is the initial label, represents the label after the previous round of training correction, p is the predicted probability of the deep learning model; L initial and L update They are the initial loss function and the updated loss function respectively; α is used to balance the two loss terms during training, and α changes dynamically with training:

[0025]

[0026] Among them, λ is a constant value between 0 and 1, E is the total training rounds, and current_epoch is the current training round.

[0027] In a second aspect, an embodiment of the present application provides an electronic device, comprising: at least one memory for storing programs; and at least one processor for executing the programs stored in the memory. When the programs stored in the memory are executed, the processor is used to execute the method provided in the first aspect.

[0028] In a third aspect, an embodiment of the present application provides a computer storage medium, in which instructions are stored. When the instructions are executed on a computer, the computer executes the method provided in the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the multiple embodiments disclosed in this specification, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only the multiple embodiments disclosed in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0030] The following is a brief introduction to the drawings required for describing the embodiments or prior art.

[0031] Figure 1 A flow chart of the method for generating land cover classification samples proposed in an embodiment of the present application;

[0032] Figure 2 A flowchart of noise screening of labels in the land cover classification sample generation method proposed in an embodiment of the present application;

[0033] Figure 3 A flow chart of the land cover classification sample generation method proposed in the embodiment of the present application using NDVI pixel-by-pixel screening labels;

[0034] Figure 4aA training flow chart of a deep learning model in a land cover classification sample generation method proposed in an embodiment of the present application;

[0035] Figure 4b A schematic diagram of an application scenario of updating and correcting land cover classification samples through deep learning model training proposed in an embodiment of the present application;

[0036] Figure 5 A schematic diagram of an electronic device proposed in an embodiment of the present application. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0038] In the description of the embodiments of the present application, words such as "exemplary", "for example" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary", "for example" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary", "for example" or "for example" is intended to present related concepts in a concrete way.

[0039] In the description of the embodiments of the present application, the term "and / or" is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, B exists alone, and A and B exist at the same time. In addition, unless otherwise specified, the term "multiple" means two or more. For example, multiple systems refers to two or more systems, and multiple terminals refers to two or more terminals.

[0040] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "include", "comprises", "has" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0041] In the description of the embodiments of the present application, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it can be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0042] In the description of the embodiments of the present application, the terms "first\second\third, etc." or die cut A, die cut B, die cut C, etc. are only used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that the specific order or sequence can be interchanged where permitted so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0043] In the description of the embodiments of the present application, the numbers representing the steps, such as S110, S120, etc., do not necessarily mean that the steps will be executed in this manner. If permitted, the order of the previous and next steps can be interchanged, or they can be executed simultaneously.

[0044] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0045] The embodiment of the present application proposes a land cover classification sample generation method, which is based on the Dempster-Shafer (DS) evidence theory and prior products, and generates low-cost, high-quality label data by more reasonably utilizing existing products across resolutions and time to improve the accuracy and efficiency of large-scale land cover classification. The quality of label data is improved by optimizing the training process of the deep learning model, providing a new solution for the rapid update and application of large-scale land cover classification maps.

[0046] Figure 1 This is a flow chart of the method for generating land cover classification samples proposed in the embodiment of the present application. Figure 1 As shown, the method includes S11, determining i prior products of remote sensing images; the prior products include land cover classification maps and accuracy evaluation results of the classification maps; i is a natural number; S12, orthogonalizing and fusing the i prior products to obtain the comprehensive probability p(T j ); S13, determine the maximum value of the comprehensive probability of each pixel in each category, and determine the land cover type of each pixel in the remote sensing image according to the maximum value of the comprehensive probability of each pixel in each category. S14, combine the maximum value p(T p ) to obtain a confidence map, and to obtain a land cover map by combining the land cover type of each pixel of the remote sensing image; each pixel of the confidence map, each pixel of the land cover map and each pixel of the remote sensing image correspond one to one in spatial position; S15, according to the confidence map, p(T p) value is lower than the set first threshold value, and the land cover type value corresponding to the pixel is set as the background value to obtain a land cover map with higher credibility, and the screened land cover map is used as the initial label of the remote sensing image; S16, the remote sensing image and the initial label (that is, the land cover map with credibility meeting the requirements) are cropped into multiple slice images and corresponding slice labels to obtain the initial sample of land cover classification.

[0047] Steps S11 to S16 are described in detail below.

[0048] S11, determine i prior products of the remote sensing image; the prior products include available classification maps and accuracy evaluation results of the classification maps; i is a natural number.

[0049] Figure 1 The prior products 1, 2, ..., i shown in the figure can be cross-resolution land cover classifications obtained by traditional classification methods. Figure 1 , land cover classification Figure 2 , …land cover classification maps i, and the accuracy evaluation results of these land cover classification maps.

[0050] The i prior products allow for spatially consistent product data across resolutions, across time, and from multiple sources with different classification standards.

[0051] The cross-resolution product data includes low-resolution product data, medium-resolution product data, and high-resolution product data that are consistent in spatial position.

[0052] Cross-temporal product data includes products that are inconsistent with the time of remote sensing images. For example, products from 2020 can be used as labels for training image classification models in 2021, and such products are cross-temporal.

[0053] S12, based on the Dempster-Shafer (DS) evidence theory, the i prior products are orthogonalized and fused to obtain the comprehensive probability p (T j ), T j is the j-th land cover type, where j is a natural number.

[0054] In some embodiments, the basic probability of the land cover type of the corresponding product 1, product 2, ... product i can be assigned as p1(T j )、p2(T j )……p i (T j ).

[0055] The essence of DS evidence theory is to calculate the orthogonal sum of multiple evidences. Based on this, the target land cover type T in i prior products can be j The basic probabilities of the remote sensing image are orthogonalized and fused to obtain the comprehensive probability p(T j ), the formula is as follows:

[0056]

[0057] In formula (1), represents the orthogonal sum, p1(T j )、p2(T j )……p i (T j ) are i prior products for the target land cover type T j The basic probability assignment of. In formula (2), k is the conflict coefficient. The larger the value of the conflict coefficient k, the more difficult it is to combine the prior products. The smaller the value of k, the easier it is to combine the products.

[0058] In order to determine the land cover type of each pixel of the remote sensing image, it is necessary to use a decision rule to make a decision based on the result of the fusion of the evidence combination rule. In the embodiment of the present application, the maximum comprehensive probability principle can be selected as the decision rule.

[0059] S13, determine the maximum value p(T p ), according to the maximum value p(T p ) Determine the land cover type T of each pixel in the remote sensing image.

[0060] In some embodiments, based on the maximum comprehensive probability principle, the comprehensive probability p(T j ) and take the maximum value as the final land cover type T of the pixel. The decision rules are as follows:

[0061] p(T p )=max j∈{0，1，2……7} p(T j ) (3)

[0062] T=T p (4)

[0063] In formula (3), p(T p ) is the maximum comprehensive probability value, T p is the land cover type corresponding to the maximum value, p(T j ) is the comprehensive probability of each type, j is the encoding of the type value, j∈{0, 1, 2…7}.

[0064] In some embodiments, since a single pixel has only one category, when orthogonalizing and combining pixels, only the comprehensive probability that all products except the background value have the same category needs to be calculated. Therefore, the maximum comprehensive probability value is used to replace the trust function and likelihood function to evaluate the credibility of the orthogonal and result.

[0065] S14, the maximum value p(T p ) to obtain a confidence map, and to obtain a land cover map by combining the land cover type of each pixel of the remote sensing image; each pixel of the confidence map, each pixel of the land cover map and each pixel of the remote sensing image correspond one-to-one in spatial position.

[0066] S15, according to the confidence map, the land cover type value corresponding to the pixel whose value is lower than the set first threshold in the land cover map is set as the background value, so as to obtain the land cover map with the required credibility, and use the land cover map with the required credibility as the initial label of the remote sensing image. The background value is a value that is not within the category code range.

[0067] S16, cutting the remote sensing image and the initial label (ie, the land cover map with the required reliability) into a plurality of slice images and corresponding slice labels to obtain an initial sample for land cover classification.

[0068] The land cover classification samples generated by the method proposed in the embodiment of the present application are used as the initial training set of the deep learning model, which has low-cost and high-quality label data and can improve the accuracy and efficiency of large-scale land cover classification.

[0069] The land cover classification sample generation method proposed in the embodiment of the present application generates large-scale land cover classification samples based on DS evidence theory and prior products, and improves the accuracy and efficiency of large-scale land cover classification by more reasonably utilizing existing products across resolutions and time to generate low-cost, high-quality label data.

[0070] Since multi-source prior products may include product data across resolutions and time and different classification standards, the initial training set may contain noisy labels. In order to reduce the noise in the initial training set of initial land cover labels, thereby reducing the impact of label noise on subsequent land cover classification tasks, the land cover classification sample generation method proposed in the embodiment of the present application can also use the deep learning model's ability to adapt to noise to filter label noise without relying on a clean data set or prior knowledge.

[0071] Figure 2 This is a flow chart of noise screening of labels in the land cover classification sample generation method proposed in the embodiment of the present application. Figure 2As shown in the figure, since too many class probability distributions may introduce redundant information, if the class with the highest probability in the class probability distribution predicted by the deep learning model is sometimes not much different from the probabilities corresponding to other classes, the deep learning model cannot classify well. Two probability-based uncertainty principles and one principle based on the normalized vegetation index can be used to filter the noise of labels to ensure that the retained labels are as accurate as possible.

[0072] The two probability-based uncertainty principles are used to filter the noise of the labels in the land cover classification sample, which specifically includes the following steps:

[0073] S21, for each slice image in the initial training set, the class probability distribution of each pixel can be predicted using the output of the deep learning model. The class probability distribution indicates the distribution of the probabilities of multiple categories on the graph.

[0074] S22, determining the maximum probability value and the second largest probability value of each pixel according to the class probability distribution of each pixel.

[0075] S23, calculating the confidence according to the maximum probability value and the second maximum probability value; the confidence includes the credibility U1 and the credibility difference U2 of the two categories, including the following steps:

[0076] S231, determine the category with the highest confidence probability for pixel X according to The confidence level U1 for determining the maximum confidence probability of X is:

[0077]

[0078] In formula (5), represents the category with the highest confidence probability of X, and U1 represents the category corresponding to the maximum confidence probability of the pixel The larger the U1, the higher the confidence level of the pixel. The greater the credibility.

[0079] S232, determine the category with the second highest probability of X confidence probability According to the category with the highest probability and the second most likely category Determine the difference U2 between the two categories of confidence:

[0080]

[0081] In formula (6), Represents the category with the second largest confidence probability of pixel X. and The difference in the credibility of the two categories, U2, reflects the similarity of the two categories predicted by the model. The larger the difference in U2, the smaller the similarity, indicating that and There is no confusion between the two categories. U2 can also be recorded as similarity.

[0082] S24, determining whether to update the original slice label corresponding to the pixel according to the credibility U1, the credibility difference U2 and a set second threshold; the second threshold includes a credibility threshold and a similarity threshold.

[0083] However, using a fixed second threshold to determine the pixels whose labels need to be updated so as to update the slice labels will result in different updating degrees for different slice labels.

[0084] Therefore, the land cover classification sample generation method proposed in the embodiment of the present application adopts an adaptive segmentation method to obtain the second threshold value for each batch of slice images, and S24 includes the following steps:

[0085] S241, the average values ​​of U1 and U2 of each batch of slice images are calculated. B×1×H×W (U1) and mean B×1×H×W (U2) is used as the updated second threshold.

[0086] Exemplarily, the size of each slice image is 256*256, and each batch represents the calculation of values ​​of multiple slices at one time.

[0087] S242, in order to avoid the restriction of the threshold of the easy-to-identify category and the relaxation of the threshold of the difficult-to-identify category, the second threshold is truncated according to the empirical value to obtain two thresholds for slicing each batch of training samples:

[0088]

[0089] If U1 of each pixel in the current slice image is greater than φ1 and U2 is greater than φ2, the category of the position corresponding to the predicted label predicted by the model under the current slice image is regarded as the correct category; if the slice label corresponding to the current slice image is inconsistent with the correct predicted label, the position corresponding to the current slice label is regarded as noise, and the category of the position corresponding to the current slice label is corrected to the category of the position corresponding to the correct predicted label.

[0090] Since the second threshold is affected by the training of the deep learning model and is not stable, even after the confidence probability screening predicted by the deep learning model such as steps S21-S23, the label noise area cannot be completely removed. In addition, due to the limitation of the initial labels, the classification boundaries of the deep learning model are often smoothed, making it difficult to identify some small and fragmented land object categories.

[0091] The Normalized Difference Vegetation Index (NDVI) can separate vegetation from water and soil, accurately reflect the information of land objects, and objectively reflect the changes in vegetation coverage. It is an index that distinguishes different land objects and reflects the growth status and coverage of vegetation.

[0092] Therefore, the land cover classification sample generation method proposed in the embodiment of the present application can also use NDVI to screen out the pixels that are misclassified in the labels of the initial training set pixel by pixel.

[0093] Figure 3 The land cover classification sample generation method proposed in the embodiment of the present application uses a flowchart of the NDVI pixel-by-pixel screening label. The screening process is as follows Figure 3 As shown, the following steps are included:

[0094] S31, determining the normalized difference vegetation index NDVI corresponding to each pixel in the slice image as:

[0095]

[0096] In formula (9), NIR is the reflectance value of the near-infrared band, and RED is the reflectance value of the red band. The NDVI value is distributed between -1 and 1. Since vegetation has strong reflectance values ​​for the red and near-infrared bands, the more lush the vegetation, the higher the NDVI value. The NDVI value of water bodies that are insensitive to these two bands is generally less than 0, and the NDVI value of bare land and buildings is generally distributed between 0 and 0.2.

[0097] S32, setting a corresponding NDVI index threshold for each pixel category in the slice image corresponding to the slice label, and the NDVI index threshold can be recorded as a third threshold.

[0098] In some implementations, a corresponding NDVI index threshold may be set for each category in the slice NDVI image corresponding to the slice label according to the empirical NDVI range of different categories.

[0099] S33, when the NDVI value of the current pixel in the slice image meets the corresponding third threshold, retain the corresponding current slice label; when the NDVI value of the current pixel in the slice image does not meet the third threshold, change the value of the current pixel to the background value.

[0100] Label noise correction is to reduce the potential noise in the initial labels, thereby alleviating the impact of label noise on deep learning models. Although deep learning models have strong feature learning capabilities, they can easily fit random noise, which will greatly reduce the performance of the model. However, an interesting phenomenon is that deep learning models tend to learn samples with correct labels in the early stages, and only start to learn samples with incorrect labels in the later stages. In addition, when keeping a high learning rate, deep networks are not prone to fitting incorrect samples.

[0101] Based on this, the land cover classification sample generation method proposed in the embodiment of the present application performs two-stage training on the deep learning model, and updates and corrects the noise of the labels in the land cover classification samples through deep learning model training.

[0102] Figure 4a This is a training flow chart of a deep learning model in the land cover classification sample generation method proposed in an embodiment of the present application.

[0103] Figure 4b Schematic diagram of the application scenario of updating and correcting land cover classification samples through deep learning model training. Figure 4a and 4b As shown, in the first stage, a large learning rate is used to train the deep learning model, including the following steps:

[0104] S41, using the slice images of the initial sample and the corresponding slice labels to train a deep learning model, and using the deep learning model as an initial classification model.

[0105] S42, with a larger initial learning rate, when the minimum training loss does not decrease for M consecutive rounds, the learning rate is scaled to 0.1 times the previous value, and the training is stopped when the learning rate starts to change for the first time. Exemplarily, the value M=10 can be taken.

[0106] S43, selecting the weight of the M+1th round before the current round as the initial weight of the deep learning model in the second stage.

[0107] The second stage of training the deep learning model includes the following steps: S44, using the slice images of the initial sample and the corresponding slice label data to train the deep learning model optimized in S43; and updating the label based on the confidence probability predicted by the optimized deep learning model.

[0108] Specifically, for each pixel, if the confidence probability predicted by the deep learning model is greater than the credibility threshold, the current label is corrected to the model prediction result, otherwise it remains unchanged.

[0109] In an embodiment of the present application, the credibility threshold is adaptively and dynamically changed, and in order to avoid label update errors caused by overfitting of the deep learning model, NDVI is used to further remove erroneous areas in the label.

[0110] Considering that the initial labels are mostly correct, if the initial labels are not involved in the training, the model's prediction results may deviate from the initial labels.

[0111] S45, use the joint loss function to calculate the loss L of the initial label and the updated label and the model prediction probability noise ,as follows:

[0112]

[0113] In formula (10), Y represents the initial label, L represents the label after the previous round of training correction, and p is the probability predicted by the model. initial and L update They represent the initial loss function and the updated loss function respectively, and λ changes dynamically with training:

[0114]

[0115] In formula (11), α is used to balance the two loss terms during the training process, where λ is a constant value between 0 and 1, E is the total training rounds, and current_epoch is the current training round. In this way, the deep learning model is trained using the loss functions of formulas (10) and (11), and the labels are iteratively updated and corrected during the training process.

[0116] The parameters of the deep learning model and the noise of the labels are dynamically optimized during the training process. Through iterative joint optimization, samples with incorrect labels are corrected, the noise rate of the dataset is reduced, and the performance of the model is improved.

[0117] The embodiment of the present application adopts DS evidence theory for the generation of land cover mapping labels, and combines the deep learning model prediction map and the normalized vegetation index to dynamically correct the land cover classification samples.

[0118] The present application embodiment provides an electronic device, such as Figure 5 As shown, it includes: at least one memory 1102, used to store programs; at least one processor 1101, used to execute the programs stored in the memory. When the programs stored in the memory are executed, the processor is used to execute the above-mentioned method for predicting ecological indicators.

[0119] The embodiment of the present application provides a computer storage medium 1103, in which instructions are stored. When the instructions are executed on a computer, the computer executes the above-mentioned method for predicting ecological indicators.

[0120] It is understood that the processor in the embodiments of the present application may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0121] The method steps in the embodiments of the present application can be implemented by hardware or by a processor executing software instructions. The software instructions can be composed of corresponding software modules, and the software modules can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, mobile hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an ASIC.

[0122] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When software is used for implementation, it can be implemented in whole or in part in the form of a computer program model. The computer program model includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the process or function according to the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted by a computer-readable storage medium. The computer instructions can be transmitted from a website site, a computer, a server or a data center to another website site, a computer, a server or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or a data center that includes one or more available media integration. The available medium can be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid state drive (SSD)), etc.

[0123] It should be understood that the various numerical numbers involved in the embodiments of the present application are only used for the convenience of description and are not used to limit the scope of the embodiments of the present application.

Claims

1. A method for generating land cover samples, characterized in that: The method comprises: Determine i prior products of the remote sensing image; the prior products include available classification maps and accuracy evaluation results of the classification maps; Orthogonalizing and fusing the i prior products to obtain the comprehensive probability of each category of each pixel in the remote sensing image; Determine the maximum value of the comprehensive probability of each category of each pixel, and determine the land cover type of each pixel of the remote sensing image according to the maximum value of the comprehensive probability of each category of each pixel; The maximum value p(T p ) to obtain a confidence map, and to obtain a land cover map by combining the land cover type of each pixel of the remote sensing image; each pixel of the confidence map, each pixel of the land cover map and each pixel of the remote sensing image correspond one-to-one in spatial position; According to the confidence map, p(T p ) the land cover type value corresponding to the pixel whose value is lower than the set first threshold value is set as the background value, and a land cover map with a credibility that meets the requirements is obtained; the land cover map with the credibility that meets the requirements is the initial label of the remote sensing image; Cutting the remote sensing image and the initial label into a plurality of slice images and corresponding slice labels to obtain an initial sample of land cover classification; Two probability-based uncertainty principles are used to filter the noise of labels in the initial sample of land cover classification, including: Training a deep learning model based on the initial samples of the land cover classification to predict the class probability distribution of each pixel of each batch of slice images; Determine the maximum probability value and the second maximum probability value of each pixel according to the class probability distribution of each pixel; The confidence is calculated according to the maximum probability value and the second maximum probability value; the confidence includes the credibility U1 and the credibility difference U2: In the formula, is the category with the largest confidence probability for pixel X, and U1 is the category corresponding to the maximum confidence probability for pixel X The larger the U1, the greater the credibility of the category corresponding to the maximum confidence probability of the pixel X; is the category with the second largest confidence probability of the pixel X; U2 is and The difference in the credibility of the two categories, U2 reflects the similarity of the two categories; According to the credibility U1, the credibility difference U2 and the set second threshold, the noise of the label in the initial sample is filtered, and the original label corresponding to the pixel is updated, including: The average values ​​of the credibility U1 and the credibility difference U2 of each batch of slice images are used as the updated second threshold value; The second threshold is truncated to obtain two thresholds for each batch of training samples: If U1 of each pixel in the current slice image is greater than φ1 and U2 is greater than φ2, the category of the position corresponding to the predicted label predicted by the model under the current slice image is regarded as the correct category; if the slice label corresponding to the current slice image is inconsistent with the correct predicted label, the position corresponding to the current slice label is regarded as noise, and the category of the position corresponding to the current slice label is corrected to the category of the position corresponding to the correct predicted label.

2. The method according to claim 1, characterized in that: The i prior products are orthogonalized and fused to obtain the comprehensive probability of each category of each pixel in the remote sensing image, including: According to the accuracy evaluation results of the i prior products, the basic probability of the target land cover type of the corresponding product 1, product 2, ... product i is assigned as p1(T j )、p2(T j )……p i (T j ), where T j is the jth land cover type, j is a natural number; The target land cover type T of the i prior products j The basic probabilities of each pixel are orthogonalized and fused to obtain the comprehensive probability p(T j ), the formula is as follows: In the formula, represents the orthogonal sum, p1(T j )、p2(T j )……p i (T j ) are the i prior products for the target land cover type T j The basic probability assignment is, k is the conflict coefficient, the larger the value of the conflict coefficient k is, the more difficult it is to combine the prior products, and the smaller the value of k is, the easier it is to combine the products.

3. The method according to claim 1 or 2, characterized in that: The labels in the initial sample of the land cover classification contain noise, and the method further includes: filtering the noise of the labels in the initial sample based on the principle of Normalized Difference Vegetation Index (NDVI), including: Calculate the value of the normalized vegetation index NDVI corresponding to each pixel in the slice image; When the NDVI value of the current pixel in the slice image meets the set third threshold, the current slice label is retained; when the NDVI value of the current pixel in the slice image does not meet the third threshold, the value of the current pixel is changed to a background value.

4. The method according to claim 1 or 2, characterized in that: The method further includes: for the initial training set, correcting and updating the noise of the labels in the initial samples of the land cover classification through deep learning model training, including: In the first stage, the deep learning model is trained using the slice images and corresponding slice labels of the initial sample, and the deep learning model is iteratively trained at the initial learning rate; when the minimum training loss does not decrease after M consecutive rounds, the learning rate is scaled to 0.1 times the previous one; the weight of the M+1th round before the current round is selected as the initial weight of the deep learning model in the second stage, and the deep learning model is optimized; In the second stage, the initial sample is used to adopt the joint loss function L noise The deep learning model is trained and optimized, and during the training process, the labels containing noise in the initial samples are iteratively corrected and updated based on the confidence probability predicted by the deep learning model; wherein the L noise is the calculation result of the loss of the initial label and the updated label with the model prediction probability.

5. The method according to claim 4, characterized in that The joint loss function L noise for: Among them, Y is the initial label, represents the updated label after the initial label correction in the last round of training, p is the predicted probability of the deep learning model; L initial and L update are the initial loss function and the updated loss function respectively; α is used to balance the two loss terms during training, and the α changes dynamically with training: Among them, λ is a constant value between 0 and 1, E is the total training rounds, and current_epoch is the current training round.

6. An electronic device comprising: at least one memory for storing a program; At least one processor is used to execute the program stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method according to any one of claims 1 to 5.

7. A computer storage medium storing instructions, which, when executed on a computer, enable the computer to execute the method according to any one of claims 1 to 5.

Images