Coral reef whitening detection method based on PU-Bagging algorithm

Abstract

The invention relates to a coral reef whitening detection method based on a PU-Bagging algorithm, and the method comprises the steps: obtaining a plurality of sentinel 2 remote sensing images in different periods, and carrying out the geographic correction and radiation normalization processing; extracting blue and green wave bands to form a plurality of time phases, and synthesizing the extracted time phases to obtain a synthesized time sequence image; rasterization is carried out on the synthesized time sequence image, whitening coral points in the image are marked, a sample data set is formed, and hidden size processing is carried out; processing and synthesizing the time sequence image and the sample data based on a PU-Bagging algorithm, constructing a standard classifier for iterative training, recording a training score, taking a score of data processed by using a hidden size as a threshold value, finding data of which the training score exceeds the threshold value as a new whitening coral point, and generating a final whitening region graph; the data processing of the remote sensing image improves the adaptability of the remote sensing image in identifying the whitened coral reef, solves the problem of poor classification result of a traditional algorithm, and can identify the whitened coral reef from the remote sensing image more accurately.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing image detection, and in particular relates to a coral reef bleaching detection method based on a PU-Bagging algorithm. Background Art

[0002] At present, frequent coral bleaching events around the world pose a serious threat to marine ecosystems. With the rapid development of remote sensing technology, remote sensing has shown great potential in detecting coral bleaching in a low-time and high-efficiency manner. In terms of data, Sentinel-2 optical data has shown great advantages in coral bleaching detection. However, Sentinel-2 optical data does not have good applicability in common software such as ENVI5.3, and still requires more complex processing of remote sensing images; at the same time, the methods in existing software cannot directly identify bleached coral reefs from remote sensing images, and lack of systematic models and methods seriously affects the convenience of coral reef bleaching identification. Therefore, designing a coral reef bleaching detection method based on the PU-Bagging algorithm that can more conveniently process bleached coral reef remote sensing images and improve recognition efficiency has become a technical problem that needs to be solved urgently. Summary of the invention

[0003] In order to solve the above problems, the present invention provides a coral reef bleaching detection method based on PU-Bagging algorithm.

[0004] The technical solution of the present invention is a method for detecting coral reef bleaching based on the PU-Bagging algorithm. The detection method comprises the following steps: Step 1: Acquire multiple Sentinel-2 remote sensing images of different periods and pre-process the remote sensing images, including geographic correction and radiation normalization; Step 2, extracting blue and green bands according to the image time sequence to form multiple time phases, and synthesizing the extracted time phases to obtain a synthetic time series image; Step 3, rasterize the synthetic time series images, mark the bleached coral points in the images, form a sample data set, and perform size hiding processing on part of the data in the sample data set; Step 4: Based on the PU-Bagging algorithm, the synthetic time series images and sample data obtained in steps 2 and 3 are processed, and a standard classifier is constructed for iterative training. The training score is recorded, and the score of the data processed with the hidden dimension is used as the threshold. The data corresponding to the unlabeled samples whose scores reach the threshold are marked as bleached coral points to generate the final bleaching area map.

[0005] As a further improvement of the present invention, the geographic correction includes using the Sen2Cor tool to perform atmospheric correction on the remote sensing image to obtain the atmospheric bottom reflectivity of each band.

[0006] As a further improvement of the present invention, the radiation normalization processing determines pseudo-invariant feature pixels by using the PIFs algorithm, and takes the sand and deep sea areas with insignificant feature changes over time as bright pixels and dark pixels respectively; selects the first scene image as the reference image x; the remaining time image is used as the image y to be corrected; uses the linear equation fitted to the bright pixels and dark pixels in the reference image x to normalize the image y; the linear equation is , where a represents the gain value and b represents the offset.

[0007] As a further improvement of the present invention, in step 2, the blue and green bands of each phase are multiplied to obtain a composite band, and then the composite bands are synthesized to obtain a composite time series image, and each band in the composite time series image represents a time phase that can be regarded as a feature.

[0008] As a further improvement of the present invention, in step 3, gridding is performed according to the size of the Sentinel-2 remote sensing image, and a sample data set is formed by marking the known bleached coral points in the image as 1 and the remaining points as 0; the hidden size processing randomly selects some sample data with labels of 1 in the sample data set, changes the labels of these sample data to 0, and hides them in the data marked as 0.

[0009] As a further improvement of the present invention, the PU-Bagging algorithm uses synthetic time series images as training data, obtains category labels from sample data sets, and obtains positive samples and unlabeled samples according to whether the category label is 1 or 0, and randomly selects the same number of unlabeled samples as the positive samples to generate a training set; uses the positive samples and the training set to construct a classifier based on a machine learning algorithm, uses the classifier to process unlabeled samples outside the training set, and calculates the similarity scores between the image pixels of the unlabeled samples and the positive samples; iterates the steps of generating a training set and calculating the similarity scores, and then calculates the average score of the similarity scores of the unlabeled samples after iteration one by one, takes the unlabeled samples whose average scores reach a threshold as positive samples, outputs the corresponding category labels, and generates a final whitening area map; the calculation formula for the average score is , where T represents the number of iterations and x represents the score of the unlabeled samples.

[0010] As a further improvement of the present invention, the classifier is based on the CART decision tree algorithm, using a data set D containing training data and a feature set F containing all the features in the training set, using the Gini index for binary division, and calculating the Gini value to measure the purity of the data set D. The calculation formula of the Gini value is: , where n represents the number of category labels, k and k' both represent a category label, and Indicates the probability of occurrence of categories k and k'; by listing the candidate attribute set C from the feature set F for the current node partition, the data set D is divided into two subsets D1 and D2, and the difference in the Gini index after different attribute partitions is compared. The calculation formula is: ,in, represents the i-th subset, represents the number of samples in the ith subset; select the attribute f that reduces the Gini index of the data set D the most as the optimal partition attribute, and the expression is , f is a single feature in the candidate attribute set C.

[0011] After adopting the above method, the remote sensing images are geographically corrected and radiometrically normalized, and all multi-temporal images are normalized by the PIFs algorithm to eliminate the influence of water on the reflectance of benthic organisms as much as possible. The red band is excluded to reduce the impact of the low permeability of the red band in the water. The blue and green bands of each phase are multiplied to generate new synthetic bands, which increases the difference in the reflectance values ​​of bleached coral reefs on the image, highlights the changes in the reflectance values ​​of seabed organisms when the upper part is covered with water, and improves the adaptability of remote sensing images in the subsequent identification of bleached coral reefs.

[0012] A training set is generated by randomly selecting the same number of unlabeled samples as positive samples, and some unlabeled samples are distributed and sampled as negative samples, which greatly reduces the sample space. By performing hidden dimension processing on some sample data with the label 1, this data is used together with the real unlabeled data to accept the classifier to calculate the score. The score of the hidden dimension data is used as the threshold to more accurately evaluate whether the unlabeled data belongs to the data of bleached coral points. This solves the problem of poor classification results of traditional algorithms when the samples are small or unbalanced, improves the accuracy of the classification results, and makes it easier to identify bleached coral reefs from remote sensing images.

[0013] By building a classifier based on the CART classification tree algorithm, the Gini index can be used to select features to represent the impurity of the model. The Gini value is calculated to reflect the probability that two randomly selected samples in the data set D have inconsistent category labels. The smaller the calculated Gini value, the lower the impurity and the better the corresponding feature. By calculating the change in the Gini index of each feature in the candidate attribute set C when dividing the data set D, the feature that reduces the Gini index of the sample set the most is found and determined as the optimal partitioning attribute, which makes the score of the unlabeled sample more accurate and reduces the error rate of identifying bleached coral points. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Shown is a schematic flow chart of the coral reef bleaching detection method of the present invention.

[0015] Figure 2 Shown is a schematic diagram of the PIFs algorithm fitting of the present invention. DETAILED DESCRIPTION

[0016] like Figure 1-2 A method for detecting coral reef bleaching based on PU-Bagging algorithm is shown, and the detection method comprises the following steps: Step 1: Acquire multiple Sentinel-2 remote sensing images of different periods and pre-process the remote sensing images, including geographic correction and radiation normalization; Step 2, extracting blue and green bands according to the image time sequence to form multiple time phases, and synthesizing the extracted time phases to obtain a synthetic time series image; Step 3, rasterize the synthetic time series images, mark the bleached coral points in the images, form a sample data set, and perform size hiding processing on part of the data in the sample data set; Step 4: Based on the PU-Bagging algorithm, the synthetic time series images and sample data obtained in steps 2 and 3 are processed, and a standard classifier is constructed for iterative training. The training score is recorded, and the score of the data processed with the hidden dimension is used as the threshold. The data corresponding to the unlabeled samples whose scores reach the threshold are marked as bleached coral points to generate the final bleaching area map.

[0017] The geographic correction includes using the Sen2Cor tool to perform atmospheric correction on the remote sensing image to obtain the atmospheric bottom reflectance of each band.

[0018] The radiation normalization process determines pseudo-invariant feature pixels by using the PIFs algorithm, and takes the sand and deep sea areas with insignificant feature changes over time as bright pixels and dark pixels respectively; selects the first scene image as the reference image x; the remaining time image as the image y to be corrected; uses the linear equation fitted to the bright pixels and dark pixels in the reference image x to normalize the image y; the linear equation is , where a represents the gain value and b represents the offset.

[0019] In step 2, the blue and green bands of each phase are multiplied to obtain a composite band, and then the composite bands are synthesized to obtain a composite time series image. Each band in the composite time series image represents a time phase that can be regarded as a feature.

[0020] In the step 3, the image is gridded according to the size of the Sentinel-2 remote sensing image, and the known bleached coral points in the image are marked as 1 and the remaining points are marked as 0 to form a sample data set; the hidden size processing randomly selects some sample data with labels of 1 in the sample data set, changes the labels of these sample data to 0, and hides them in the data marked as 0.

[0021] The PU-Bagging algorithm uses synthetic time series images as training data, obtains category labels from sample data sets, and obtains positive samples and unlabeled samples according to whether the category label is 1 or 0. The same number of unlabeled samples as positive samples are randomly selected to generate a training set; the positive samples and the training set are used to construct a classifier based on a machine learning algorithm, and the classifier is used to process unlabeled samples outside the training set to calculate the similarity score between the image pixels of the unlabeled samples and the positive samples; iterates the steps of generating a training set and calculating the similarity score, and then calculates the average score of the similarity scores of the unlabeled samples after iteration one by one, and takes the unlabeled samples whose average score reaches the threshold as positive samples, outputs the corresponding category labels, and generates the final whitened area map; the calculation formula of the average score is , where T represents the number of iterations and x represents the score of the unlabeled samples.

[0022] The classifier is based on the CART decision tree algorithm, using a dataset D containing training data and a feature set F containing all features in the training set. The Gini index is used for binary division, and the Gini value is calculated to measure the purity of the dataset D. The calculation formula of the Gini value is: , where n represents the number of category labels, k and k' both represent a category label, and Indicates the probability of occurrence of categories k and k'; by listing the candidate attribute set C from the feature set F for the current node partition, the data set D is divided into two subsets D1 and D2, and the difference in the Gini index after different attribute partitions is compared. The calculation formula is: ,in, represents the i-th subset, represents the number of samples in the ith subset; select the attribute f that reduces the Gini index of the data set D the most as the optimal partition attribute, and the expression is , f is a single feature in the candidate attribute set C.

[0023] By performing geographic correction and radiation normalization on the remote sensing images, the PIFs algorithm is used to normalize all multi-temporal images to eliminate the influence of water on the reflectance of benthic organisms as much as possible. By excluding the red band, the impact of the low permeability of the red band in the water is reduced. The blue and green bands of each phase are multiplied to generate new synthetic bands, which increases the difference in the reflectance values ​​of bleached coral reefs on the image, highlights the changes in the reflectance values ​​of seabed organisms when the upper part is covered with water, and improves the adaptability of remote sensing images in the subsequent identification of bleached coral reefs.

[0024] A training set is generated by randomly selecting the same number of unlabeled samples as positive samples, and some unlabeled samples are distributed and sampled as negative samples, which greatly reduces the sample space. By performing hidden dimension processing on some sample data with the label 1, this data is used together with the real unlabeled data to accept the classifier to calculate the score. The score of the hidden dimension data is used as the threshold to more accurately evaluate whether the unlabeled data belongs to the data of bleached coral points. This solves the problem of poor classification results of traditional algorithms when the samples are small or unbalanced, improves the accuracy of the classification results, and makes it easier to identify bleached coral reefs from remote sensing images.

[0025] By building a classifier based on the CART classification tree algorithm, the Gini index can be used to select features to represent the impurity of the model. The Gini value is calculated to reflect the probability that two randomly selected samples in the data set D have inconsistent category labels. The smaller the calculated Gini value, the lower the impurity and the better the corresponding feature. By calculating the change in the Gini index of each feature in the candidate attribute set C when dividing the data set D, the feature that reduces the Gini index of the sample set the most is found and determined as the optimal partitioning attribute, which makes the score of the unlabeled sample more accurate and reduces the error rate of identifying bleached coral points.

Claims

1. A coral reef bleaching detection method based on PU-Bagging algorithm, characterized by: The detection method comprises the following steps: Step 1: Acquire multiple Sentinel-2 remote sensing images of different periods and pre-process the remote sensing images, including geographic correction and radiation normalization; Step 2, extracting blue and green bands according to the image time sequence to form multiple time phases, and synthesizing the extracted time phases to obtain a synthetic time series image; Step 3, rasterize the synthetic time series images, mark the bleached coral points in the images, form a sample data set, and perform size hiding processing on part of the data in the sample data set; Step 4: Based on the PU-Bagging algorithm, the synthetic time series images and sample data obtained in steps 2 and 3 are processed, and a standard classifier is constructed for iterative training. The training score is recorded, and the score of the data processed with the hidden dimension is used as the threshold. The data corresponding to the unlabeled samples whose scores reach the threshold are marked as bleached coral points to generate the final bleaching area map.

2. A coral reef bleaching detection method based on PU-Bagging algorithm according to claim 1, characterized in that: The geographic correction includes using the Sen2Cor tool to perform atmospheric correction on the remote sensing image to obtain the atmospheric bottom reflectance of each band.

3. A coral reef bleaching detection method based on PU-Bagging algorithm according to claim 2, characterized in that: The radiation normalization process determines pseudo-invariant feature pixels by using the PIFs algorithm, and regards sand and deep sea areas with insignificant feature changes over time as bright pixels and dark pixels respectively; selects the first scene image as the reference image x; and the remaining time images as the image to be corrected y; The image y is normalized using a linear equation fitted to the bright and dark pixels in the reference image x; the linear equation is , where a represents the gain value and b represents the offset.

4. A coral reef bleaching detection method based on PU-Bagging algorithm according to claim 1, characterized in that: In step 2, the blue and green bands of each phase are multiplied to obtain a composite band, and then the composite bands are synthesized to obtain a composite time series image. Each band in the composite time series image represents a time phase that can be regarded as a feature.

5. A coral reef bleaching detection method based on PU-Bagging algorithm according to claim 1, characterized in that: In the step 3, the image is gridded according to the size of the Sentinel-2 remote sensing image, and the known bleached coral points in the image are marked as 1 and the remaining points are marked as 0 to form a sample data set; the hidden size processing randomly selects some sample data with labels of 1 in the sample data set, changes the labels of these sample data to 0, and hides them in the data marked as 0.

6. A coral reef bleaching detection method based on PU-Bagging algorithm according to claim 5, characterized in that: The PU-Bagging algorithm uses synthetic time series images as training data, obtains category labels from sample data sets, and obtains positive samples and unlabeled samples according to whether the category label is 1 or 0. The same number of unlabeled samples as positive samples are randomly selected to generate a training set; the positive samples and the training set are used to construct a classifier based on a machine learning algorithm, and the classifier is used to process unlabeled samples outside the training set to calculate the similarity score between the image pixels of the unlabeled samples and the positive samples; iterates the steps of generating a training set and calculating the similarity score, and then calculates the average score of the similarity scores of the unlabeled samples after iteration one by one, and takes the unlabeled samples whose average score reaches the threshold as positive samples, outputs the corresponding category labels, and generates the final whitened area map; the calculation formula of the average score is , where T represents the number of iterations and x represents the score of the unlabeled samples.

7. A coral reef bleaching detection method based on PU-Bagging algorithm according to claim 6, characterized in that: The classifier is based on the CART decision tree algorithm, using a dataset D containing training data and a feature set F containing all features in the training set. The Gini index is used for binary division, and the Gini value is calculated to measure the purity of the dataset D. The calculation formula of the Gini value is: , where n represents the number of category labels, k and k' both represent a category label, and Indicates the probability of occurrence of categories k and k'; by listing the candidate attribute set C from the feature set F for the current node partition, the data set D is divided into two subsets D1 and D2, and the difference in the Gini index after different attribute partitions is compared. The calculation formula is: ,in, represents the i-th subset, represents the number of samples in the i-th subset; select the attribute f that reduces the Gini index of the data set D the most as the optimal partition attribute, and the expression is , f is a single feature in the candidate attribute set C.

Images