1. Introduction
The United States Geological Survey (USGS), in their National Water Quality Assessment Program (NAWQA), defines water quality monitoring as a continuous period of data collection (in lakes, streams, rivers, reservoirs, wetlands, or oceans), in order to evaluate the chemical, physical, and biological characteristics of the body of water with respect to its ecological conditions and designated water uses [
1]. Monitoring water quality typically involves a series of in-situ observations, measurements, and water sample collections that are analyzed for various parameters depending on the individual project goals, such as temperature, phosphorus (P), nitrogen (N), total solids, pH, fecal bacteria, conductivity, dissolved oxygen (DO), biochemical oxygen demand (BOD), hardness, alkalinity, suspended sediments, other nutrients, trace metals, and water clarity. Traditionally, water quality indicators are determined by the collection, field examination, and laboratory analyses of water samples, following consistent protocols and guidelines [
2].
Although properly collected and analyzed in-situ measurements are highly accurate, these measurements can be time-consuming, susceptible to errors (especially visual subjectivity), and can only be related to a specific point in time and space [
3,
4]. Due to these potential problems, water quality monitoring programs that rely solely on these types of measurements may fail to provide accurate spatial or temporal views of water quality.
Considering the above constraints, the use of remote sensing and satellite imagery in water quality monitoring and management has been widely implemented to estimate different water quality parameters [
5,
6,
7]. Images from different Earth observing satellites (e.g., Landsat 5, Landsat 7, Landsat 8, Terra, Aqua, SPOT, among others) with the capability of obtaining information in the visible (0.4–0.8 μm), near infrared (0.8–1 μm) (NIR), and thermal (10–12 μm) portions of the electromagnetic (EM) spectrum, have been used to estimate different water quality parameters, such as total suspended solids (TSS), chlorophyll-a (Chl-a), pH, colored dissolved organic matter (CDOM), and Secchi disk depth (SDD). In case 1 waters (i.e., oceans), determination of optical properties is affected by phytoplankton. In case 2 waters (i.e., inland waters), determination of optical properties is more complex, due to the of presence of dissolved mineral sediments and organic matter [
5,
6,
8,
9,
10,
11,
12,
13,
14,
15,
16,
17,
18,
19].
Satellite technology has proven not only to be able to obtain unbiased information on specific characteristics of lakes, but to also serve as a cost-effective complement to in-situ monitoring programs [
3]. The implementation of remote sensing and satellite imagery addresses two of the most important limiting factors when obtaining in-situ measurements: (1) The subjective error susceptibility associated with these types of measurements and (2) the limited, discrete sampling point coverage limitation [
12].
Despite the benefits of using this technology, a major challenge when using optical imagery in observing and determining water quality parameters is its excessive susceptibility to data limitations due to cloud coverage [
2,
6,
20,
21,
22]. At the same time, most applications have focused on detection, determination, and prediction of optical water quality parameters like Chl-a, SDD, and CDOM [
21,
23,
24,
25,
26].
However, in recent years, the collection of high-resolution images using small unmanned aerial systems (sUAS) has become more prevalent [
27,
28,
29,
30,
31]. Pairing sUAS with multispectral sensors may provide cloud free images with higher revisiting time (temporal resolution) and smaller spatial resolutions at relatively low costs [
32,
33]. Particularly, for water quality monitoring and modeling, different authors have developed site specific models using multispectral images collected with sUAS [
34,
35]. Su and Chou [
34] used a multispectral sensor onboard an sUAS in order to map the trophic state of Tain-Pu reservoir in Kinmen, Taiwan. As part of their findings, they determined that the ratio between multispectral bands has the ability to predict Chl-a, total phosphorous (TP), and SDD. At the same time, they corroborated that with the flexibility that sUAS offers not only in terms of temporal resolution; but also with respect to higher spatial resolution, stronger regression models can be obtained. Kageyama et al. [
35] used an sUAS in order to perform water quality analysis in the Miharu reservoir in Japan. Their findings indicate that the ratios between multispectral bands are helpful to determine chemical oxygen demand (COD), TSS and Chl-a concentrations.
The main purpose of this study was to develop algorithms capable of estimating optical (TSS, Chl-a, SDD) and non-optical (TP and total nitrogen (TN)) water quality parameters in an oligotrophic system and a eutrophic system using images collected with a multispectral sensor attached to an sUAS. Given the global economic and water security challenges posed by increasing nutrient over enrichment and resulting water quality degradation [
36,
37,
38,
39], systems representing end members of the biological productivity spectrum were selected to generate robust and widely applicable models. Additionally, this study evaluates whether using a well-accepted statistical interpolation method improves the algorithms between the different water quality parameters and the reflectance values.
4. Discussion
The main purpose of this study was to develop models capable of reliably estimating optical (TSS, Chl-a, and SDD) and non-optical (TP and TN) water quality parameters in two extremes of the aquatic biological productivity spectrum (oligotrophic and eutrophic systems), using in situ data and images collected with a multispectral sensor attached to an sUAS. In order to develop these algorithms, linear approaches using single and multiple variables were used. As a result, it was determined that linear models using multiple variables had stronger predictive capabilities for all water quality parameters. These algorithms have the capability of generating data that are not statistically different from the collected in-situ data for optical and non-optical water quality parameters.
In the paper “Comprehensive Review on Water Quality Parameters Estimation Using Remote Sensing Techniques”, Gholizabeh et al. [
52] references that different authors determined that the use of visible and near infrared bands of the EM spectrum from multispectral sensors can be used to obtain strong correlations between reflectance and optical water quality parameters. However, when exploring correlations for non-optical parameters, direct inference of these measurements had low predictive capabilities. Lim and Choi [
53] used Landsat 8 in order to correlate spectral bands with in-situ water quality measurements, in order to establish water quality models capable of estimating optical (TSS and Chl-a) and non-optical (TN and TP) parameters in the Nakdong River in Korea. As a result, they obtained algorithms that strongly estimated TSS and Chl-a (R
2 = 0.74 and 0.71, respectively), but were not as strong when estimating TP and TN (R
2 = 0.50 and 0.48, respectively). Due to this limitation, an indirect estimation approach has been taken by some authors in order to develop strong correlations that relate TP and TN to Chl-a concentrations and SDD [
54,
55].
When examining the multiple variable models determined by this study (
Table 3), it can be observed that the combination between the ratios Blue/Red, Green/Red, and Green/Blue provide the strongest correlation between reflectance and most of the optical and non-optical water quality parameters (except for Chl-a, for which the highest correlation was obtained with the Green and Red bands). These findings are in accordance with Gholizabeh et al. [
52]; however, it was determined by this study that these ratios not only have the capability of estimating optical parameters, but also non-optical values. Lui et al. [
56] determined that with the use of high-resolution imagery, linear (multiple linear regression) and non-linear (artificial neural network) models with strong predictive capabilities could be developed for TN and TP. The basis of these relationships is explained by the high spectral correlation that TN and TP have with SDD, TSS, and Chl-a [
57].
It is imperative to begin this discussion with this information because the study presented herein deviates from the traditional approach of using multispectral sensors attached to satellite platforms. Instead, this study uses a more compact multispectral sensor, attached to an sUAS. By doing this, not only can direct methods of estimating non-optical water quality parameters be derived, but the use of this tool enhances spatial and temporal resolutions while eliminating the cloud coverage issues.
Earth observation satellites are the most common platforms to monitor and collect information about the Earth [
58].
Table 6 presents some of the most common remote sensing satellites used for estimating water quality parameters, along with their respective spatial and temporal resolutions. From this table, it can be determined that the spatial resolution obtained by any of these platforms is much coarser when compared to the spatial resolution (6–8 cm) obtained with the sensor used in this study. To illustrate this concept,
Figure 14 presents a visual comparison between images taken from two commonly used remote sensing satellites (Landsat 8 and Sentinel-2A) versus the images captured by the sUAS in the eutrophic system used in this study. By looking at these aerial images, the pixel resolution significantly increases in the picture taken with the sUAS.
The use of satellite remote sensing tools helps to expand the limited discrete sampling point coverage of traditional monitoring plans [
3,
6,
26,
59,
60]. However, in addition to spatial resolution, two major drawbacks when using these tools are: (1) The longer revisiting time (temporal resolution) of these platforms and (2) cloud coverage limitations. Zhang and Kovacs [
61] point out that the longer temporal resolutions of some of these platforms presents a major difficulty when trying to monitor systems that are in a constant state of change. At the same time, other authors mention that the number of images that they are unable to use due to cloud coverage accounts in some cases for up to 97% of the available captured imagery for a particular region in a 25-year period [
3]. With the use of sUASs, these issues are no longer a concern. First, with an sUAS, the operator has the flexibility of deciding how often they want to capture multispectral imagery. Secondly, because sUAS fly below the clouds, all the imagery is 100% cloud coverage free.
In order to determine optical and non-optical water quality measurements from multispectral sensors, in-situ measurements are needed to develop and calibrate the different models [
26]. However, due to the above limitations for use of satellite imagery, selecting images for these types of correlations can become a non-trivial task. Hicks et al. [
62] suggest that ideal imagery for these types of studies should not be more than one day apart from the in-situ data collection. However, in most cases, this is not possible due to the temporal resolution of the platform or cloud coverage present in the imagery [
3,
6,
26,
60,
61]. Furthermore, Barrett and Frazier [
63] mention that water quality parameters can be directly influenced by rapidly changing environmental conditions in the study site, and as a result, the utility of predictive models developed from imagery that is generated days or weeks from the day of the in-situ sampling can be detrimentally impacted. As shown in
Table 1, with the use of an sUAS, the time window between water sample acquisition and multispectral imagery collection can be reduced to minutes to hours. In theory, and due to the flexibility that these portable platforms provide, decreasing the time window between water sample acquisition and multispectral imagery collection translates to stronger and more reliable water quality models.
A secondary objective of this study was to evaluate if using a statistical interpolation method improved the algorithms between the different optical and non-optical water quality parameters and the reflectance values. In order to do that, three scenarios were evaluated under the reflectance extraction procedures: (1) Point extraction, (2) buffer extraction, and (3) kriging extraction. Results indicated that models created from the first scenario (point extraction) presented stronger predictive capabilities. Mu et al. [
64] references that in spatial sampling, collected samples are not independent from each other and for that reason the number of samples that need to be taken in order to develop or validate remote sensing products can be decreased in order to improve accuracy. Considering the water quality results and the spatial distribution maps generated in this study, it makes sense that for fully mixed systems (such as the ponds used in this study), fewer sample stations led to more accurate models.
For all the points discussed, one can determine that the use of sUAS offers additional benefits than the traditional satellite remote sensing approach. However, it is important to point out that sUAS, just like any other remote sensing tool, have their limitations. The first and perhaps the most important limiting factor when using this technology is the weather. When planning missions with sUAS, the operator must be aware that these platforms are unable to fly under wet conditions (rain) and elevated wind speeds (higher than 5 m/s or as stipulated by the platform manufacturer). For the study presented above, these issues were not a concern. However, it is necessary to point this out, because even though sUAS offer more advantages when it comes to obtaining imagery capable of estimating optical and non-optical water quality parameters, there is a tradeoff that needs to be considered and evaluated by the user.
5. Conclusions
This study aimed to create different statistical water quality models for optical (TSS, SDD, and Chl-a) and non-optical (TP and TN) water quality parameters in oligotrophic and eutrophic aquatic systems using remote sensing images from an sUAS equipped with a multispectral sensor. From the results of this study, it can be concluded that: (1) When using a multiple linear regression approach, models capable of predicting optical and non-optical models (with strong prediction capability R2 > 0.80) can be created, (2) multiple variable linear regressions in the visible portion of the electromagnetic spectrum (blue, green, and red) best described the relationship between TSS (R2 = 0.99, p-value = <0.01), Chl-a (R2 = 0.85, p-value = <0.01), TP (R2 = 0.98, p-value = <0.01), TN (R2 = 0.98, p-value = <0.01), and SDD (R2 = 0.88, p-value = <0.01), (3) the use of statistical interpolation (ordinary kriging) does not improve the statistical relationship between the different water quality parameters and the reflectance values, (4) 100% cloud free imagery can be collected with the use of sUAS, (5) the use of sUAS for water quality monitoring allows the user more flexibility in terms of temporal and spatial resolution, and (6) future research should evaluate if the use of this technology improves the predictive capabilities of water quality models that rely on satellite imagery and if the models developed in this study have the capacity of determining water quality in reservoirs that fall in other portions of the biological productivity spectrum.