Dataset of Blue Algae in Taihu Lake Based on Random Forest Algorithm and Satellite Monitoring (2019)

Yang, Z.^1,2  Pan, X.^2,3*  Yuan, J. ^1,2  Song ,H. ^1,2  Xu, K. ^1,2  Wu, Y. H. ^1,2  Yang, Y. B.^2,3*

1. School of Earth Sciences and Engineering, Hohai University, Nanjing 211100, China;

2. Jiangsu Province Engineering Research Center of Water Resources and Environment Assessment Using Remote Sensing, Hohai University, Nanjing 211100;

3. School of Geography and Remote Sensing, Hohai University, Nanjing 210098, China

Abstract: The cyanobacteria data of Taihu Lake represent essential and important data for the water resource management of Taihu Lake. In this paper, a GF-6 satellite image of Taihu Lake in 2019 is selected, and the random forest method based on multiple remote sensing factors (Normalized Differential Vegetation Index and Normalized Differential Water Index) is used to extract cyanobacteria to obtain a cyanobacteria dataset for the western part of Taihu Lake in 2019. The dataset was validated using the overall classification accuracy, Kappa coefficient, producer accuracy, user accuracy, misclassification error, and omission error. The validation results showed that the mean overall classification accuracy and Kappa coefficient for this dataset reached 0.97 and 0.95, respectively. The dataset includes cyanobacteria distribution data from May to December 2019 for six periods. The spatial resolution of the dataset is 20 m. The dataset is archived in the .tif format, and it consists of six data files with a data size of 0.98 MB (compressed into one file, 601 KB).

Keywords: Taihu Lake; cyanobacteria; random forest; 2019

DOI: https://doi.org/10.3974/geodp.2023.03.11

CSTR: https://cstr.escience.org.cn/CSTR:20146.14.2023.03.11

Dataset Availability Statement:

The dataset supporting this paper was published and is accessible through the Digital Journal of Global Change Data Repository at: https://doi.org/10.3974/geodb.2023.12.01.V1 or https://cstr.escience.org.cn/CSTR:20146.11.2023.12.01.V1.

1 Introduction                               

Taihu Lake is the second-largest freshwater lake in China. With the rapid development of the economy and the expansion of industry, the phytoplankton in Taihu Lake have multiplied, resulting in an outbreak of cyanobacteria. This has damaged the ecosystem of Taihu Lake and affected the domestic water supply of the surrounding cities. Therefore, it is important to monitor the outbreak of cyanobacteria in Taihu Lake^[1,2]. The traditional monitoring method for blue-green algae is to collect water samples on site, but this method requires high manpower and material resources, and the number of sampling points that can be used is limited. Remote sensing technology allows large-scale monitoring, and it is low cost and provides a fast response. Therefore, in recent years, monitoring blue-green algae based on remote sensing images has been a research direction of concern for domestic and foreign scholars. GF-6 is the first optical imaging satellite with a red edge band in China; it can better monitor vegetation, and cyanobacteria and vegetation have similar spectral characteristics. Therefore, it is reasonable to generate the Taihu Lake cyanobacteria dataset based on GF-6 satellite data.

The methods for extracting blue-green algae locations based on remote sensing data mainly include empirical modelling, threshold, and machine learning methods^[3?C7]. The empirical model method establishes a nonlinear or linear model of an index and measured chlorophyll concentration, and then it monitors blue-green algae through the chlorophyll concentration; however, this method has regional limitations^{[8, 9]}. The threshold method extracts blue-green algae by setting a single band or exponential threshold, but the accurate selection of the threshold is a challenge^{[10, 11]}. Machine learning methods are based on feature indicators to classify and extract blue-green algae; these methods can mine big data patterns, and they include support vector machines and random forest methods^[12]. The random forest method is a popular method in machine learning and has been successfully applied to parameter inversion.

Therefore, random forest extraction based on GF-6 data can better meet the requirements of the long-term and high-precision dynamic monitoring of cyanobacteria in Taihu Lake in the future. This dataset is based on a 2019 GF-6 image of Taihu Lake of good quality, and it uses the random forest method to achieve the extraction of blue-green algae in Taihu Lake. This paper describes the generation of Taihu Lake blue-green algae data products for 2019.

2 Metadata of the Dataset

The metadata of the Cyanobacteria dataset of random forest algorithm for satellite monitoring in Taihu Lake (2019)^[13] is summarized in Table 1.

3 Methods

3.1 Algorithm Principle

This dataset mainly uses the random forest (RF) algorithm, which is a special bagging algorithm proposed by Leo^[15] for classification or regression. The difference between the RF method and the original bagging algorithm is that the RF method uses a decision tree as the model. In this paper, the RF classification algorithm is used, and the class to which the current object belongs is obtained by voting through the decision tree. The establishment of an RF model requires representative values of the input variables and classification results; finally, predictions of the results are obtained based on the RF model^[16].

The input variables selected in this dataset are the Normalized Differential Vegetation Index (NDVI) and Normalized Differential Water Index (NDWI) because ideally, Taihu Lake only has two kinds of surface features: water and cyanobacteria. The NDVI and NDWI can identify vegetation and water bodies, respectively. Therefore, the NDVI and NDWI are used as input variables for the RF method:

                                                                                                             (1)

                                                                                                           (2)

Table 1  Metadata summary of the Cyanobacteria dataset of random forest algorithm for satellite monitoring in Taihu Lake (2019)

where, , , and  represent the reflectance of the red, near-red, and green bands, respectively. According to the band requirements of the NDVI, the GF-6 satellite uses the third and fourth bands to calculate the NDVI. According to the band requirements of the NDWI, the GF-6 satellite uses the second and fourth bands to calculate the NDWI.

3.2 Data Collection and Processing

The technical roadmap of this dataset is shown in Figure 1, which mainly consists of three steps: data preparation, data pre-processing, and the establishment and prediction of RF models. The main purpose of data preparation is to find high-quality GF-6 data from 2019 with low cloud cover. The data pre- processing mainly involves the radi­ometric correction, atmospheric cor­rection, and geometric correction of the original high-resolution GF-6 data. Because the original high-resolution GF-6 data are in the form of digital number (DN) values obtained by satellites, these data need to be pre-processed and converted into surface reflectance values. The establishment and prediction process of the RF model mainly involves calculating the values of the NDVI and NDWI based on the pre-treated GF-6 data, using these two factors as the RF training data to establish the model, and finally predicting the blue-green algae in Taihu Lake.

4 Data Results and Validation

4.1 Data Composition

The dataset of blue algae in Taihu Lake in 2019 based on the random forest algorithm and satellite monitoring consists of a total of one folder. The folder contains six periods of blue-green algae data in the format of labelled image files (.tif).

4.2 Data Result

Figure 2 is a spatial distribution map of the blue-green algae in Taihu Lake in 2019, where the colours blue and green represent water and blue-green algae, respectively. As there is aquatic vegetation in the eastern part of Taihu Lake, the spatial distribution of this area will not be shown. On 5 May 2019, the blue-green algae in Taihu Lake were mainly concentrated in the southern and north-west coastal areas of Meiliang Lake. On 29 July 2019, the outbreak of blue-green algae in Taihu Lake was significantly weakened. The outbreak of blue-green algae was concentrated in Zhushan Lake and Meiliang Lake, and there was a small amount of blue-green algae in Gonghu Lake. Compared with 29 July 2019, on 13 September 2019, blue-green algae began to erupt in the north-west, south-west, and southern coastal areas of Taihu Lake, and they continued to erupt in Meiliang Lake. On 19 October 2019, blue-green algae broke out in Taihu Lake, mainly in the middle of Taihu Lake and Meiliang Lake. On 5 November 2019, the blue-green algae outbreak in Taihu Lake weakened rapidly, and there was almost no blue-green algae in Gonghu Lake. The blue-green algae outbreak was mainly in Meiliang Lake and along the north-west coast. On 12 December 2019, there were almost no cyanobacteria in Taihu Lake, and sporadic cyanobacteria were mainly concentrated in the south-west coastal area. This shows that in 2019, the blue-green algae in Taihu Lake mainly broke out in autumn, spring, and summer; there were almost no outbreaks of blue-green algae in winter.

4.3 Data Validation

4.3.1 Verification Method

The confusion matrix was used to calculate six indicators, the overall classification accuracy, Kappa coefficient, producer accuracy, user accuracy, misclassification error, and missed classification error^[17], to evaluate the extracted results. This dataset was validated using pseudocolour images synthesised from the near-red, red, and green bands as a reference. Through visual interpretation, 300 sample points (150 blue-green algae samples and 150 water samples) were uniformly selected for validation.

4.3.2 Verification Results

The accuracy evaluation of this dataset is shown in Table 2. It can be seen that the average overall classification accuracy of this dataset reaches 0.97, and the average Kappa coefficient also reaches 0.95. This indicates that the accuracy of this dataset is relatively high. In particular, the overall classification accuracy, producer accuracy, and user accuracyof the Taihu Lake cyanobacteria results on 12 December 2019 reached 0.99. The accuracy of the Taihu Lake cyanobacteria results on 12 December 2019 was highest in the six Taihu Lake images. On 5 May 2019, the accuracy of the cyanobacteria results for Taihu Lake was relatively low, and the overall classification accuracy, Kappa coefficient, and

Figure 2  Maps of blue algae distribution in Taihu Lake in 2019

producer accuracy reached 0.95, 0.91, and 1.00, respectively. However, on 5 May 2019, the cyanobacteria results for Taihu Lake wrongly classified water as cyanobacteria, which may be due to the influence of thin clouds in the Taihu Lake image from 5 May 2019. In conclusion, it can be seen from the table that the RF method based on the NDVI and NDWI calculated from GF-6 data can better extract the locations of cyanobacteria in Taihu Lake.

Table 2  The precision evaluation of the dataset

5 Discussion and Conclusion

Since previous research on the extraction of cyanobacteria from Taihu Lake rarely used domestic GF series satellite images as the data source, this design uses the RF method to obtain a high-quality Taihu Lake cyanobacteria dataset for 2019 based on the GF-6 data. This dataset has undergone radiation correction, atmospheric correction, and geometric correction, and the RF method trained on the original high-resolution data was used to obtain predictions. The data product contains the 2019 results of the Taihu Lake cyanobacteria, with a relatively good original image quality, a spatial resolution of 20 m, and a tagged image file format (.tif).

This dataset provides data support for the environmental governance of Taihu Lake, and it is of great significance for the dynamic monitoring of cyanobacteria in Taihu Lake. Compared to the traditional threshold method for extracting blue-green algae, the algorithm used to create this product avoids the uncertainty of threshold values and is time-saving and labour-saving. However, the accuracy of the algorithm in this product is strongly related to the number of training samples and the distribution of the training samples, and the algorithm used to create this product will need to consider the impact of clouds on the results in the future.

Author Contributions

Pan, X. and Yang, Y.B designed the overall algorithms of dataset. Yang, Z. and Yuan, J. contributed to the data processing and analysis. Song, H., Xu,K. and Wu,Y.H. verify the data. Yang, Z. wrote the data paper.

Conflicts of Interest

The authors declare no conflicts of interest.

References

[1]      Qin, B. Q., Gao, G., Zhu, G. W., et al. Lake eutrophication and its ecosystem response [J]. Chinese Science Bulletin, 2013, 58(10): 855?C864.

[2]      Zhu, L. Y. Remote sensing monitoring and assessment of water quality for lakes [D]. Beijing: Institude of Remote Sensing Applications, Chinese Academy of Sciences, 2006

[3]      Yang, Y., Hang, W. L., Xie, H. B., et al. A study on water information extraction method of cyanobacteria lake based on Landsat8 [J]. Remote Sensing for Land and Resources, 2020, 32(4) : 130?C136.

[4]     Wang, M., Zheng, W., Liu, C., Application of Himawari-8 data with high-frequency observation for Cyanobacteria bloom dynamically monitoring in Taihu Lake [J]. Journal of Lake Sciences, 2017, 29(5): 1043?C1053.

[5]      Shi, H., Li, X. W., Niu, Z. C., et al. Remote sensing information extraction of aquatic vegetation in Taihu Lake based on Random Forest Model [J]. Journal of Lake Sciences, 2016, 28(3): 635?C644.

[6]      Xia, X. R., Wei, Y. C., Xu, N., et al. Decision tree model of extracting blue-green algal blooms information based on Landsat TM/ETM+ imagery in Taihu Lake [J]. Journal of Lake Sciences, 2014, 26(6): 907?C915.

[7]      Huang, J. Z., Zhao, R., Satellite remote sensing monitoring of cyanobacteria outbreak in Taihu Lake [J]. Remote Sensing Information, 1999(4): 43?C44.

[8]      Li, X. W., Shi, H., Zhang, Y., et al. Cyanobacteria blooms monitoring in Taihu Lake based on the Sentinel-2A satellite of European Space Agency [J]. Environmental Monitoring in China, 2018, 34(4): 169?C176.

[9]      Miao, S., Wang, R., Li, J. C., et al. Retrieval algorithm of phycocyanin concentration in inland lakes from Sentinel 3A-OLCI images [J]. Journal of Infrared and Millimeter Waves, 2018, 37(5): 621?C630.

[10]    Li, X. Z., Lv, H., Li, Y. M., et al. Spatial scale difference analysis of cyanobacteria bloom extraction based on MODIS and GOCI data [C]. Jiangsu Society of Oceanology and Limnology, 2013.

[11]    Li, Y. C., Sun, J. L., Xie, Z. Q., et al. Extraction methods of cyanobacteria bloom in Lake Tai based on MODIS vegetation index [J]. Journal of the Meteorological Sciences, 2011, 31(6): 737?C741.

[12]    Yang, Z., Pan, X., You, C. S., et al. Spatio-temporal variation of fractional vegetation coverage in the Aydingkol Lake Basin [J]. Journal of Applied Remote Sensing, 2022, 16(1): 1?C23.

[13]   Yang, Z., Pan, X., Yuan, J., et al. Cyanobacteria dataset of random forest algorithm for satellite monitoring in Taihu Lake (2019) [J/DB/OL]. Digital Journal of Global Change Data Repository, 2023. https://doi.org/10.3974/geodb.2023.12.01.V1. https://cstr.escience.org.cn/CSTR:20146.11.2023.12.01.V1.

[14]    GCdataPR Editorial Office. GCdataPR data sharing policy [OL]. https://doi.org/10.3974/dp.policy.2014.05 (Updated 2017).

[15]    Breiman, L., Cutler, R. A. Random forests machine learning [J]. Journal of Clinical Microbiology, 2001, 45(1): 5?C32.

[16]    Pan, X., Yang, Z., Yang, Y. B., et al. Comparison and applicability analysis of methods for extracting cyanobacteria from Taihu Lake based on GF-6 data [J]. Journal of Lake Sciences, 2022, 34(6): 1866?C1876.

[17]    Yin, J., Zhu, Y. F., Comparative study of water extraction methods in different regions of OLI images [J]. Jiangxi Science, 2020, 38(5): 743?C747.