The full name of WOS is Web of Science, which is Clarivate’s product (formerly Thomson Reuters intellectual property and Technology). WOS includes three famous Citation Index Databases (SCI, SSCI, and A & HCI) and collects authoritative and influential journals in various disciplines. WOS, as a document retrieval tool due to its strict selection criteria and citation index mechanism, has also become one of the essential basic evaluation tools of bibliometrics and scientific metrology. WOS can retrieve relevant papers in various ways, such as keywords, authors, topics, DOI, etc. And it can enable authors or researchers to find the articles and data efficiently. EI is the citation abbreviation of the Engineering Index, a largescale comprehensive retrieval tool with the most extended history, and was founded by the American Federation of Engineers in 1884 (Palmer et al., 2015; Li et al., 2020). EI enjoys a high reputation worldwide in the academic, engineering, and information circles and is a critical retrieval tool commonly recognized by scientific and technological circles (Aasen et al., 2015). The content includes the research results of all engineering disciplines and engineering activities.

According to the statement of “TS = (hyperspectral AND water quality AND monitor)”, Retrieve relevant papers in the WOS and EI databases. The data source is set to “search in web of science core collection” in the WOS platform (Hilton et al., 2012). Then, select the citation index as “Editions: All”, select the document field as “Topic” and set the retrieval time as “Publication: All years (1994–2022)”. Six hundred thirtytwo records were retrieved from WOS. Select the retrieval time from “all years” to “till now”, and 404 records are extracted in the EI platform. Two hundred fiftyone publications are retrieved by the two databases simultaneously, and 221 publications are irrelevant. Finally, 564 practical items were obtained after selection.

The bibliometric analysis aims to summarize a series of results by describing publications objectively, systematically, and quantitatively. It is generally divided into four steps: sampling, recording entries, cataloging, and metrics. The first three steps are to record the titles, author, keywords, publication time, main methods, quoted data, etc., of the publication through reasonable rules to form standardized data. Then, the metrics are realized to extract and analyze the hidden information in all publications, which can provide a direction for future research. The above work is more standardized and efficient with the support of software represented by VOSviewer and CiteSpace (Govender et al., 2007). VOSviewer is a free software based on Java developed by the Centre for Science and Technology Studies (CWTS) of Leiden University in the Netherlands in 2009 (Ben-Dor et al., 2002). It is mainly adapted to the mock examination undirected network analysis and focuses on the visualization of scientific knowledge. CiteSpace, known as “citation space”, is a citation visual analysis software that focuses on analyzing the potential knowledge contained in scientific analysis and is gradually developed under the background of scient metrics and data visualization (O’Grady et al., 2021). This paper analyzes the scientific research status and trend in the field of water quality hyperspectral monitoring from four aspects with the support of the above software. The research scheme includes word counting, concept classification, spatial analysis, and semantic intensity analysis.

It can be concluded that the water quality monitoring work based on hyperspectral technology is generally divided into three research stages according to the statistics of the number of publications and citation frequency in the last 30 years (Figure 1).

Statistics on the number of books published by all countries can reflect the significant international cooperation countries in water quality hyperspectral research. The larger the circle, the greater the number of publications (Figure 2). Discharge the number of publications in a clockwise direction. The number of publications in 18 countries exceeded 8. The country with the largest international cooperation papers is China, with 157 publications. The following four countries are the United States (133), Germany (56), Italy (38), and the United Kingdom (33). According to the 2021 China Water Resources Bulletin, the total amount of water resources in China in 2021 was 2963.82 billion cubic meters. As stated by the US Environmental Protection Agency, the United States has approximately 300,000 freshwater lakes, 5,200 rivers, and thousands of miles of coastline. Germany has abundant groundwater and surface water resources, most of which are used for water supply and agricultural irrigation. Italy has abundant freshwater resources, including about 61,000 km of rivers, lakes, and groundwater reserves. According to the UK Environment Agency, the UK has some large reservoirs and water supply systems to meet the country’s water demand, as well as many small rivers and lakes. China and the United States have coauthored publications with most countries, representing the highest level in water quality hyperspectral research. One notable feature is that these two countries are the first to conduct relevant research. The results of other countries are delayed for about 2 years compared with China and the United States. The support for new technologies and the spirit of exploring new fields are the main reasons for the two countries’ outstanding development in this field.

Drawing the frequency diagram of keywords and calculating their occurrence times after counting the keywords of all publications (Figure 3). It is worth noting that this is a “cloud words picture”, which is a visual highlight of the key words that appear frequently in the paper. The more they appear, the larger the font is displayed, the more prominent the keyword is. The color here is only used to distinguish from adjacent words, and has no representational meaning. The distribution of words is random and has no direct relationship with the geographical location of the response. The main research points and technologies of hyperspectral water quality monitoring can be obtained from the keywords (Table 1). The application scenario of hyperspectral technology for water quality monitoring, the organization format of hyperspectral data, the target of water quality monitoring, and the main algorithm ideas can be learned according to the top 20 words that appear most frequently. Firstly, the keywords “Water”, “Quality”, “Waters”, “Lake”, “Coastal” and “Inland” indicate that the application scenario of hyperspectral technology has covered the main water types on the Earth’s surface, including lakes, oceans, and rivers. The scholars designed the technical process of each field and drew valuable conclusions according to the needs of the application. Secondly, the keywords such as “Reflection”, “Hyperspectral”, “Remote”, “Imagery”, “Optical”, “Spectral”, “Data” indicate that as passive remote sensing technology, the image data format is mainly used in water quality monitoring application. Image data is acquired primarily through the spectrometer mounted on three platforms, such as satellite, airborne or UAV. In addition, the spectral range is visible to nearinfrared, which is the response band of the main substances in the water (Zhang D. et al., 2021). Thirdly, the keywords “Chlorophyll”, “Matter”, “Content”, “Blooms”, and “Color” indicate the material information that has been proved can be accurately extracted by hyperspectral technology repeatedly. Chlorophyll content has been widely studied as the most common substance in water. The study of water bloom and color, closely related to chlorophyll concentration, has also indirectly become a research hotspot (Song K. et al., 2010). The former mainly focuses on the distribution law in natural waters, while the latter focuses on urban malodorous black river water research. In addition, the research on total nitrogen, total phosphorus, ammonia nitrogen, and other substances has also yielded many results that can be popularized. Fourthly, the keywords “model” and “retrieval” indicate that the algorithm of hyperspectral technology is still based on inversion technology. Largearea information extraction combines limited in-situ data and spectral data to establish a correlation model has become a mainstream method (Wang et al., 2005). The machine learning algorithm has become the core of the new modeling algorithm and has received a lot of new research in this process.

Draw a pennant plot with “water quality” as the goal, and infer the research trend according to the keywords (Figure 4). The idf value is an algorithm calculation result of tag words after clustering, which is used to infer mainstream research methods. The main contents of the literature are generated into cluster labels, and the research intensity of each research field is obtained after the weighted calculation. Usually, the idf is plotted after calculating the logarithm (Jianguang et al., 2005; Yan et al., 2005; Yan et al., 2006). Green, orange, and yellow in the figure, respectively, represent the most important keywords, the key keywords under study, and the trend distribution of follow-up studies taking idf value as the evaluation index. Keywords with higher idf values depend more on keywords with lower idf values. It can be obtained from the figure that spectral reflectance is one of the fundamental technologies for water quality monitoring, and subsequent research is developed based on the optical mechanism of water. It can be concluded from the orange area that the current research mainly focuses on specific application scenarios, including lakes, water colors, internal roads, and water bodies. The essential research difficulties are leaf green extraction, atmospheric correction, and inversion algorithm. The follow-up research trend shows an explosive state, with many keywords in the yellow area in the figure. In addition, some more complex issues have been paid attention to, including the research and development of optical instruments, the extraction of micro materials, the application of reservoirs, the application of UAV remote sensing, the development of fluorescence technology, and the research of numerical simulation technology (Lee et al., 2007). On the whole, the pennant plot shows the divergent research rules of water quality remote sensing monitoring from the spectroscopy principle to the application of different scenarios and then to the study of complex problems (Bresciani et al., 2017).

Burst chart can show the mutation law of keywords in the field of water quality hyperspectral monitoring (Table 2). The mutation represents the whole process that the research frequency of scholars began to increase over some time, and gradually tended to be stable and cool with time. The red line in the figure represents the rise and continuation of research. The results of calculating the top 24 keywords of the last 15 years indicate that.

(1) Data: Airborne sensors represented by CASI were the primary data acquisition form from 2009 to 2015 on the platform of hyperspectral sensors. Airborne sensors can obtain hyperspectral data of thousands of square kilometers within a few hours, making remarkable achievements in inland water and reservoir water quality monitoring. However, the high data acquisition costs limit this technology’s largescale promotion (Fenocchi et al., 2015). A large number of micro spectrometers have been studied from 2016 to 2019. Micro spectrometers have played a good role in monitoring cities’ odorous water bodies and coastal zones. Sensors mounted on UAVs can obtain hyperspectral data flexibly and at a low cost, becoming the most important data source with the promotion of UAV technology from 2019 to 2022.

The main application scenarios can be classified as urban water, lake, natural river and ocean after the hyperspectral remote sensing technology is introduced into the water quality monitoring demand (Figure 5). The monitoring needs of urban water are concentrated on the assessment of reservoir water quality (Song K. et al., 2010), black-odor water management (Wei et al., 2019), sewage discharge monitoring (Pascucci et al., 2012), and flood warning (Lin et al., 2019). Hyperspectral data are acquired mainly by UAV and ground sensors, which meet the requirements of information extraction at the microscale. Lake’s monitoring needs are significantly different, and the study core is the chlorophyll evaluation (Katlane et al., 2020). In particular, a large number of studies have been carried out on the rapid identification of cyanobacteria outbreaks because the lake water body is close to the territory of human production and living, and the water flow is relatively poor, which is prone to water bloom (Pyo et al., 2021). The research on natural rivers is more complicated than in the previous two scenarios. The critical point of the study is how to make hyperspectral technology partially or entirely replace the traditional in-situ assay method (Hong et al., 2021). Although more than ten indicators are considered adequate, the leading indicators reported in academic publications are total phosphorus, total nitrogen, ammonia nitrogen, and dissolved organic matter. The most complex application scenario is the ocean. The water quality monitoring accuracy is insufficient only through hyperspectral technology due to the interference of salinity and sea waves. Lidar, fluorescence, and ultraviolet sensors are also introduced to realize the joint monitoring of multiple sensors (Gorkavyi et al., 2021). Good results have been achieved in monitoring seawater conductivity, water depth, and algae. Representative typical cases are listed in Table 3.

The hyperspectral sensors mentioned in the publication can be divided into three platforms: satellite, airborne, and UAV, according to statistics. Most of the higher-solution spectrometers are not specially designed for monitoring the water environment except for a few sensors. Spectral resolution refers to the ability of the sensor to distinguish different wavelengths, usually expressed in microns (μm). The main spectrum range is 0.401 μm and 0.402 μm, considering the application needs of vegetation, rock mining, and cities. All sensors covering a spectrum range of 0.400 μm can play a role in water quality monitoring since it is the main spectral response band of water quality parameters (Mielke et al., 2014). UAV hyperspectral sensors can usually reach the submeter level, and airborne sensors can get the meter level in spatial resolution. While the highest answer of hyperspectral satellite data is the OHS sensor of China Orbita company, which reaches 10 m. And the mainstream spatial resolution of other satellite sensors is 30 m, which is the result of balancing data transmission efficiency and monitoring effect. Various countries are actively launching new hyperspectral sensors, and there is a lot of research and development of sensors on multiple platforms. Therefore, the number of sensor classes is much more than those listed in Table 4.

Hyperspectral water quality monitoring parameters can be obtained by making statistics on the research objects of all publications and drawing thermodynamic maps. The more the color in the figure tends to be red, the higher the research concentration. Four clusters can be formed like “islands”, corresponding to the monitoring focus of different fields (Figure 6). Firstly, research about geometrical optics of water bodies. The research target generally has a significant characteristic spectrum based on the analysis of the optical mechanism of water, such as chlorophyll-a, phytoplankton, dissolved organic matter, CDOM, turbidity, segments, and water color. These indicators can significantly show different characteristics in the visible band with additional water content, so they have been studied extensively. Mature research technology, sufficient theoretical basis, and reliable conclusions are the characteristics of these studies. Secondly, research about water bloom outbreak monitoring. Focusing on monitoring cyanobacteria, phycocyanin, algal and suspended sectors, and the water quality changes in different seasons are focused. Multiperiod remote sensing data is generally used to identify the dynamic changes of water bloom in the water area. The rivers and lakes near the urban are the main areas of study. The transparency and water depth of the water area are evaluated. Thirdly, research about chemical parameters of water quality. It belongs to the quantitative analysis of water quality parameters. The indicators studied include carbon, fluorescence, organic matter, particulate matter, total phosphorus, biomass, vegetation indicators, pH, and temperature. The water-holding capacity of the water body is evaluated to guide the development of water quality based on the calculated content value. Fourthly, research about water quality evaluation and management. The water pollution investigation technology is studied under the premise of the first three explorations, including organic carbon, acid mine drainage, microplastics, and contamination. These pollutants have attracted worldwide attention, and hyperspectral technology is becoming one of the effective technologies (Table 5).

The algorithms used in all publications are counted to obtain the most commonly used algorithms (Figure 7). The research generally requires a lot of preprocessing and pre-analysis of spectral data, “classification”, “atmospheric correction”, “spectral indexes” and “radial transfer” have appeared in abundant publications to illustrate this point. Cause accurate spectral identification and atmospheric correction can ensure the correctness of spectral data and provide data sets for water quality parameter modeling. After spectral preprocessing, the algorithms for calculating water quality parameters can be divided into regression and machine learning methods. The most commonly used regression methods are semiempirical models based on limited in-situ data. “Partial least squares”, “multiple linear regression”, “principal component analysis”, and “random forest” are all commonly used methods. The keywords that appear most frequently in machine learning methods include “neural network”, “artificial intelligence”, and “deep learning”. In recent years, machine learning is the most studied method. In addition, the terms of “calibration”, “visualization”, “prediction”, and “verification” also show that scholars attach importance to the accuracy of the algorithm (Table 6).