Remote sensing extraction method of fish raft in culture area based on GEE and machine learning

Abstract

The invention discloses a remote sensing extraction method for fish rows in a culture area based on GEE and machine learning, and the method comprises the steps: obtaining a remote sensing image on line based on a GEE platform, carrying out the preprocessing, constructing a multi-dimensional data set for the preprocessed remote sensing image through employing Sentinel-2 spectral features and a plurality of indexes, and carrying out the selection of the multi-dimensional data set, and constructing a training sample and a verification sample; inputting the training sample into a random forest classification model for training, determining an optimal parameter, obtaining a trained random forest classification model, and carrying out fish raft remote sensing image classification by adopting the trained random forest classification model; and inputting a remote sensing image to be detected into the trained random forest classification model to output a classification result, and post-processing and visualizing the classification result. According to the method, efficient batch online processing of multi-time-sequence remote sensing images, multi-dimensional data set construction, remote sensing classification based on machine learning, classification post-processing, classification result precision evaluation and aquaculture area statistics can be realized, and the remote sensing extraction precision of aquaculture area fish raft distribution information can be greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of marine remote sensing information, and in particular to a remote sensing extraction method for fish cages in aquaculture areas based on GEE and machine learning. Background Art

[0002] Nearshore marine fishery management is of great significance, and the type and location information of fish cages in aquaculture areas is a key factor. Remote sensing observation has the characteristics of wide monitoring range and good timeliness. Through remote sensing technology, we can quickly and widely obtain information such as the temporal and spatial distribution of fish cages in aquaculture areas.

[0003] The traditional algorithm for extracting fish cages in aquaculture areas based on single-view remote sensing images and single indexes has less information, while the comprehensive use of multiple time series and multiple indexes for fish cages in aquaculture areas can effectively improve the accuracy and efficiency of extraction. Multi-time series and multi-index remote sensing images often require a lot of downloading and preprocessing, which is time-consuming and labor-intensive. Summary of the invention

[0004] The main purpose of the present invention is to overcome the above-mentioned defects in the prior art, and propose a remote sensing extraction method for fish rafts in aquaculture areas based on GEE and machine learning, which can realize efficient batch online processing of multi-time series remote sensing images, construction of multidimensional data sets, remote sensing classification based on machine learning, post-classification processing, classification result accuracy evaluation and area statistics of fish rafts in aquaculture areas, which can greatly improve the accuracy of remote sensing extraction of fish raft distribution information in aquaculture areas.

[0005] The present invention adopts the following technical solution:

[0006] A remote sensing extraction method for fish rafts in aquaculture areas based on GEE and machine learning, comprising:

[0007] Based on the GEE platform, remote sensing images are obtained online and preprocessed. The preprocessed remote sensing images are used to construct multidimensional data sets using Sentinel-2 spectral features and multiple indexes. The multidimensional data sets are selected to construct training samples and verification samples.

[0008] Input the training samples into the random forest classification model for training and determine the optimal parameters to obtain a trained random forest classification model, and use the trained random forest classification model to classify the remote sensing images of the fish cage;

[0009] Inputting the remote sensing image to be tested into the trained random forest classification model to output the classification result, and post-processing and visualizing the classification result;

[0010] Based on the post-processing and visualized classification results, the distribution of fish rafts is extracted, and the areas of cage-type fish rafts and longline fish rafts are counted.

[0011] The preprocessing includes using the QA60 band to identify and mask the thick cloud and cirrus cloud pixels of the remote sensing image; secondly, selecting the false color band for synthetic display; and finally, cutting and mosaicking the remote sensing image according to the study area.

[0012] The Sentinel-2 spectral features include band B2, band B3, band B4, band B5, band B6, band B7, band B8, band B8A, band B11 and band B12.

[0013] The multiple indexes include normalized vegetation index, normalized building index, normalized water index, improved normalized difference water index and enhanced water index; the normalized vegetation index is used to monitor the important parameters of vegetation growth status and its coverage, and NDVI is used to enhance vegetation information and reduce the impact of vegetation coverage on the identification of fish cages in aquaculture areas. The calculation formula is as follows:

[0014]

[0015] In the formula, NIR represents the reflectivity of the near-infrared band, and R represents the reflectivity of the red light band;

[0016] The normalized building index is used to eliminate building information and to reduce the impact of building land on the identification of fish cages in aquaculture areas. The calculation formula is as follows:

[0017]

[0018] In the formula, SWIR1 represents the reflectivity of the short-wave infrared band, and NIR represents the reflectivity of the near-infrared band;

[0019] The normalized water index is used to highlight the water information in the image and reduce the probability of mixing the water with the fish cage in the aquaculture area. The calculation formula is as follows:

[0020]

[0021] In the formula, G represents the reflectivity of the green light band, and NIR represents the reflectivity of the near infrared band;

[0022] The improved normalized difference water index significantly enhances the contrast between water and buildings and reduces background errors. The calculation formula is as follows:

[0023]

[0024] In the formula, G represents the reflectivity of the green light band, and SWIR1 represents the reflectivity of the short-wave infrared band;

[0025] The enhanced water index is used to suppress noise from residential areas, soil, and vegetation, and is suitable for identifying water bodies in semi-arid areas, reducing the probability of mixing water bodies with fish cages in aquaculture areas. The calculation formula is as follows:

[0026]

[0027] Where G represents the reflectivity of the green light band, NIR represents the reflectivity of the near infrared band, and SWIR1 represents the reflectivity of the short-wave infrared band.

[0028] The parameters required for the random forest classification model include the number of decision trees, the number of features required to be input when splitting a node, the minimum number of samples contained in a leaf node, the number of features in each tree bag fraction, and a random seed.

[0029] The classification results are post-processed, specifically including: converting the classified images into vector files, smoothing and filtering the sporadic category noises that appear in the classification, and merging the smaller patches; the visualization is to partition the post-processed images according to the required areas of the study area.

[0030] It also includes performing accuracy evaluation on the classification results output by the random forest classification model and outputting the evaluation results.

[0031] The accuracy evaluation includes performing a quantitative accuracy analysis on the classification result and the validation sample based on a confusion matrix, respectively calculating the overall accuracy and the Kappa coefficient, and determining the classification quality of the classification result according to the corresponding relationship between the Kappa coefficient and the classification quality.

[0032] 9. A remote sensing extraction method for fish rafts in aquaculture areas based on GEE and machine learning as claimed in claim 8, characterized in that the overall accuracy calculation formula is as follows:

[0033]

[0034] Where: n is the number of rows and columns; X ij is the element in the i-th row and j-th column in the classification result; N is the sum of all elements.

[0035] The Kappa coefficient calculation formula is as follows:

[0036]

[0037] Where: n is the number of rows and columns; X ij is the element in the i-th row and j-th column in the classification result; X i+ is the sum of the rows; X +i is the sum of the columns; N is the sum of all elements.

[0038] It can be seen from the above description of the present invention that, compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. The method of the present invention can realize efficient batch online processing of multi-time series remote sensing images, construction of multidimensional data sets, remote sensing classification based on machine learning, post-classification processing, classification result accuracy evaluation and area statistics of fish cages in aquaculture areas, which can greatly improve the remote sensing products of fish cage distribution information in aquaculture areas.

[0040] 2. The present invention processes remote sensing images online quickly and efficiently based on the GEE platform, avoiding the download of large amounts of heavy remote sensing images, and selecting samples by comparing the GEE platform and Google images, thereby increasing the accuracy of sample selection interpretation and improving the automation of the entire process.

[0041] 3. The present invention utilizes Sentinel-2 spectral features and various indexes to construct a multidimensional data set, thereby reducing the impact of other land objects on the identification of fish cages in aquaculture areas and improving the accuracy of remote sensing extraction of fish cages in aquaculture areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is the principle diagram of the method of the present invention;

[0043] Figure 2 Select training samples for the GEE visualization interface;

[0044] Figure 3 This is a distribution map of fish farming cages in a certain place.

[0045] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0046] The present invention is further described below through specific implementation modes.

[0047] See also Figure 1 , a remote sensing extraction method for fish rafts in aquaculture areas based on GEE and machine learning, including:

[0048] 1) Based on the GEE platform, remote sensing images are obtained online and preprocessed. The preprocessed remote sensing images are used to construct a multidimensional data set using Sentinel-2 spectral features and multiple indices. The multidimensional data set is then selected to construct training samples and validation samples.

[0049] To build a high-quality Sentinel-2 dataset, cloud pixels in the image must be removed to generate a minimum cloud image. The present invention uses the QA60 band (Table 1 Sentinel-2MSI band parameter information) to identify and mask thick cloud and cirrus pixels to obtain high-quality observations; secondly, the false color band (the red, green, and blue bands are set to correspond to B8, B3, and B2) is conducive to improving the display of fish cages in the aquaculture area, and the false color band is selected for synthetic display; then the remote sensing image is cropped and mosaicked according to the study area.

[0050] Table 1

[0051]

[0052] Specifically, a multidimensional data set is constructed using Sentinel-2 spectral features and various indices as input features of the classifier. Sentinel-2 spectral features include B2 band, B3 band, B4 band, B5 band, B6 band, B7 band, B8 band, B8A band, B11 band and B12 band. Multiple indices include normalized vegetation index, normalized building index, normalized water index, improved normalized difference water index and enhanced water index.

[0053] In addition, the ALOSDSM30m data product is used to obtain two terrain features: elevation and slope. Finally, a multidimensional dataset with the above 17 bands is formed as the classification feature set of the classifier.

[0054] Typical ground object recognition indices such as NDVI (Normalized Difference Vegetation Index), NDBI (Normalized Difference Building Index), NDWI (Normalized Difference Water Index), MNDWI (Modified Normalized Difference Water Index), and EWI (Enhanced Water Index) are selected to improve the separability of different ground object categories in remote sensing images. Among them, different water body indices have their advantages and disadvantages. Three typical water body indices are selected as input features to better extract fish cages in aquaculture areas from other ground objects.

[0055] (1) NDVI (Normalized Difference Vegetation Index)

[0056] Normalized Difference Vegetation Index (NDVI) is an important parameter used to monitor vegetation growth and coverage. Generally, NDVI>0 indicates vegetation coverage, and increases with coverage. NDVI is used to enhance vegetation information and reduce the impact of vegetation coverage on fish cage identification in aquaculture areas. The calculation formula is as follows:

[0057]

[0058] Where NIR represents the reflectivity in the near-infrared band, and R represents the reflectivity in the red band.

[0059] (2) NDBI (Normalized Difference Building Index)

[0060] Normalized Difference Build-up Index (NDBI) is mainly used to remove building information and use the normalized building index to reduce the impact of building land on the identification of fish cages in aquaculture areas. In general, the NDBI value is between [-1,1], NDBI>0 is building land, and the larger the value, the higher the probability that the pixel is a building. The calculation formula is as follows:

[0061]

[0062] Where SWIR1 represents the reflectivity of the shortwave infrared band, and NIR represents the reflectivity of the near infrared band.

[0063] (3) NDWI (Normalized Difference Water Index)

[0064] Normalized Difference Water Index (NDWI) is used to highlight the water body information in the image and reduce the probability of mixing the water body with the fish cages in the aquaculture area. Generally, the part with NDWI ≥ 0 is the water body area. The calculation formula is as follows:

[0065]

[0066] Where G represents the reflectivity of the green light band, and NIR represents the reflectivity of the near-infrared band.

[0067] (4) MNDWI (Modified Normalized Difference Water Index)

[0068] The modified normalized difference water index (MNDWI) was proposed to improve the limitation of NDWI in extracting water information (the extraction effect of water bodies with building shadows in the image is poor). MNDWI significantly enhances the contrast between water bodies and buildings and reduces background errors. Therefore, it can well distinguish between water bodies and buildings. The calculation formula is as follows:

[0069]

[0070] Where G represents the reflectivity of the green light band, and SWIR1 represents the reflectivity of the shortwave infrared band.

[0071] (5) EWI (Enhanced Water Index)

[0072] The Enhanced Water Index (EWI) can suppress noise from residential areas, soil, and vegetation and is suitable for identifying water bodies in semi-arid areas, reducing the probability of mixing water bodies with fish cages in aquaculture areas. The calculation formula is as follows:

[0073]

[0074] Where G represents the reflectivity of the green light band, NIR represents the reflectivity of the near infrared band, and SWIR1 represents the reflectivity of the short-wave infrared band.

[0075] The training samples and validation samples can be selected by overlaying the selected remote sensing image and high-resolution Google map in the map display interface of the code editor of the GEE platform, and selecting samples on the GEE platform. Figure 2 As shown, the sample data is stored in the cloud. You can also select samples locally, such as selecting samples on Google Earth, converting the data format to shapefile, and then uploading it to a specific account on the GEE platform. A total of 5,250 training samples and 4,250 verification samples were selected. The specific types of samples include fish cages in aquaculture areas (divided into cage-type fish cages and longline fish cages), bare land, water bodies, vegetation, buildings, mangroves, salt marshes, and others.

[0076] 2) Input the training samples into the random forest classification model for training and determine the optimal parameters to obtain a trained random forest classification model, and use the trained random forest classification model to classify fish cage remote sensing images.

[0077] The Random Forest (RF) algorithm is a multi-classifier ensemble algorithm. The GEE platform provides the function ee.Classifier.smileRandomForest() to implement the Random Forest algorithm. The required parameters are: 1) The number of decision trees

[0078] (number of trees); 2) the number of features that need to be input when splitting a node (variables persplit); 3) the minimum number of samples contained in a leaf node (min leaf population); 4) the number of features in each tree bag (bag

[0079] fraction); 5) random seed (seed). After experiments, the optimal parameters of this project are shown in Table 2.

[0080] Table 2

[0081]

[0082]

[0083] 3) Input the remote sensing image to be tested into the trained random forest classification model to output the classification results, and post-process and visualize the classification results.

[0084] In order to further improve the accuracy of remote sensing image classification results, the classified images are usually converted into vector files, the sporadic category noise in the classification is smoothed and filtered, and the small patches are merged. The classified vector results are exported from the GEE platform and post-processed and visualized in ArcGIS Pro. The post-processing is based on the comparison of the superimposed Sentinel-2 remote sensing images to adjust the vector results. The visualization is based on the required zoning of the study area. See Figure 3 .

[0085] The present invention also includes: 4) performing accuracy evaluation on the classification results output by the random forest classification model and outputting the evaluation results.

[0086] The accuracy evaluation includes quantitative accuracy analysis of the classification results and validation samples based on the confusion matrix, calculating the overall accuracy and Kappa coefficient respectively, and determining the classification quality of the classification results according to the corresponding relationship between the Kappa coefficient and the classification quality.

[0087] Specifically, the overall accuracy calculation formula is as follows:

[0088]

[0089] Where: n is the number of rows and columns; X ij is the element in the i-th row and j-th column of the classification result; N is the sum of all elements.

[0090] The Kappa coefficient calculation formula is as follows:

[0091]

[0092] Where: n is the number of rows and columns; X ij is the element in the i-th row and j-th column of the classification result; X i+ is the sum of the rows; X +i is the sum of the columns; N is the sum of all elements.

[0093] The corresponding relationship between the Kappa coefficient and the classification quality is shown in Table 3 below.

[0094] Table 3

[0095]

[0096]

[0097] The present invention also includes 5) extracting the distribution of fish rafts according to the post-processing and visualized classification results, and counting the areas of cage-type fish rafts and longline-type fish rafts, thereby achieving refined statistics of fish rafts in aquaculture areas.

[0098] The method of the present invention can process remote sensing images online quickly and efficiently based on the GEE platform, avoid downloading a large amount of heavy remote sensing images, and select samples by comparing the GEE platform and Google images to increase the accuracy of sample selection interpretation, thereby improving the automation of the whole process processing.

[0099] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A remote sensing extraction method for fish rafts in aquaculture areas based on GEE and machine learning, characterized in that: include: Based on the GEE platform, remote sensing images are obtained online and preprocessed. The preprocessed remote sensing images are used to construct multidimensional data sets using Sentinel-2 spectral features and multiple indexes. The multidimensional data sets are selected to construct training samples and verification samples. Input the training samples into the random forest classification model for training and determine the optimal parameters to obtain a trained random forest classification model, and use the trained random forest classification model to classify the remote sensing images of the fish cage; Inputting the remote sensing image to be tested into the trained random forest classification model to output the classification result, and post-processing and visualizing the classification result; Based on the post-processing and visualized classification results, the distribution of fish rafts is extracted, and the areas of cage-type fish rafts and longline fish rafts are counted.

2. The remote sensing extraction method of fish rafts in aquaculture areas based on GEE and machine learning as claimed in claim 1, characterized in that: The preprocessing includes using the QA60 band to identify and mask the thick cloud and cirrus cloud pixels of the remote sensing image; secondly, selecting the false color band for synthetic display; and finally, cutting and mosaicking the remote sensing image according to the study area.

3. The remote sensing extraction method of fish rafts in aquaculture areas based on GEE and machine learning as claimed in claim 1, characterized in that: The Sentinel-2 spectral features include band B2, band B3, band B4, band B5, band B6, band B7, band B8, band B8A, band B11 and band B12.

4. The remote sensing extraction method of fish rafts in aquaculture areas based on GEE and machine learning as claimed in claim 1, characterized in that: The multiple indexes include normalized vegetation index, normalized building index, normalized water index, improved normalized difference water index and enhanced water index; the normalized vegetation index is used to monitor the important parameters of vegetation growth status and its coverage, and NDVI is used to enhance vegetation information and reduce the impact of vegetation coverage on the identification of fish cages in aquaculture areas. The calculation formula is as follows: In the formula, NIR represents the reflectivity of the near-infrared band, and R represents the reflectivity of the red light band; The normalized building index is used to eliminate building information and to reduce the impact of building land on the identification of fish cages in aquaculture areas. The calculation formula is as follows: In the formula, SWIR1 represents the reflectivity of the short-wave infrared band, and NIR represents the reflectivity of the near-infrared band; The normalized water index is used to highlight the water information in the image and reduce the probability of mixing the water with the fish cage in the aquaculture area. The calculation formula is as follows: In the formula, G represents the reflectivity of the green light band, and NIR represents the reflectivity of the near infrared band; The improved normalized difference water index significantly enhances the contrast between water and buildings and reduces background errors. The calculation formula is as follows: In the formula, G represents the reflectivity of the green light band, and SWIR1 represents the reflectivity of the short-wave infrared band; The enhanced water index is used to suppress noise from residential areas, soil, and vegetation, and is suitable for identifying water bodies in semi-arid areas, reducing the probability of mixing water bodies with fish cages in aquaculture areas. The calculation formula is as follows: Where G represents the reflectivity of the green light band, NIR represents the reflectivity of the near infrared band, and SWIR1 represents the reflectivity of the short-wave infrared band.

5. The remote sensing extraction method of fish rafts in aquaculture areas based on GEE and machine learning as claimed in claim 1, characterized in that: The parameters required for the random forest classification model include the number of decision trees, the number of features required to be input when splitting a node, the minimum number of samples contained in a leaf node, the number of features in each tree bag fraction, and a random seed.

6. The remote sensing extraction method of fish rafts in aquaculture areas based on GEE and machine learning as claimed in claim 1, characterized in that: The classification results are post-processed, specifically including: converting the classified images into vector files, smoothing and filtering the sporadic category noises that appear in the classification, and merging the smaller patches; the visualization is to partition the post-processed images according to the required areas of the study area.

7. The remote sensing extraction method of fish rafts in aquaculture areas based on GEE and machine learning as claimed in claim 1, characterized in that: It also includes performing accuracy evaluation on the classification results output by the random forest classification model and outputting the evaluation results.

8. The remote sensing extraction method of fish rafts in aquaculture areas based on GEE and machine learning as claimed in claim 1, characterized in that: The accuracy evaluation includes performing a quantitative accuracy analysis on the classification result and the validation sample based on a confusion matrix, respectively calculating the overall accuracy and the Kappa coefficient, and determining the classification quality of the classification result according to the corresponding relationship between the Kappa coefficient and the classification quality.

9. The remote sensing extraction method of fish rafts in aquaculture areas based on GEE and machine learning as claimed in claim 8, characterized in that: The overall accuracy calculation formula is as follows: Where: n is the number of rows and columns; X ij is the element in the i-th row and j-th column in the classification result; N is the sum of all elements.

10. The remote sensing extraction method of fish rafts in aquaculture areas based on GEE and machine learning as claimed in claim 9, characterized in that: The Kappa coefficient calculation formula is as follows: Where: n is the number of rows and columns; X ij is the element in the i-th row and j-th column in the classification result; X i+ is the sum of the rows; X +i is the sum of the columns; N is the sum of all elements.

Images