1. Introduction
Inland waters play an important role in human lives, providing water for drinking, irrigation, and industrial use. With satellite remote sensing, the state of inland waters can be monitored through synoptic observations collected at frequent intervals. Generally, remote sensing of inland waters is much more difficult than that of open oceans. This is because the optical properties of inland waters are not determined only by phytoplankton (like open oceans), but are also strongly influenced by other constituents (
i.e., non-algal particles [NAP] and colored dissolved organic matter [CDOM]). As a result, inland waters have more complex optical properties. Both NAP and CDOM have larger absorptions at the blue and green spectral regions, and thus make these bands inappropriate for retrieving chlorophyll-
a concentration (Chl-
a) in many inland waters [
1–
6].
To address this problem, several indices have been proposed to remove or minimize the effects of NAP and CDOM based on use of remote sensing reflectance at the red and near infrared (NIR) spectral regions (NIR-red algorithms, [
1,
2,
7–
9]). However, two challenges remain. First, although the design of these indices was based on theoretical studies of the inherent optical properties (IOP) of inland and coastal waters, Chl-
a estimation models were still empirically calibrated through regression analyses of the proposed indices and measured Chl-
a. Therefore, the developed Chl-
a estimation models strongly depend on the calibration dataset used. Ideally, the calibration dataset used does not have a sampling bias. Practically, however, it is difficult to collect enough water samples to represent all the water conditions in the world, and thus the applicability of the models will be limited. Second, the proposed indices were developed based on several specific assumptions, some of which may not be valid in highly turbid lakes such as Lake Taihu and Lake Dianchi in China and Lake Kasumigaura in Japan [
10–
12].
Recently, Yang
et al. proposed a semi-analytical model-optimizing and look-up-table (SAMO-LUT) method [
11], which can potentially estimate Chl-
a in wide range of inland waters. Since the SAMO-LUT was proposed based on three wavelengths—665 nm, 708 nm and 753 nm—for Chl-
a estimation, it is a NIR-red algorithm. In the SAMO-LUT algorithm, a comprehensive synthetic dataset of reflectance spectra related to various combinations of water constituents with a wide dynamic range was used to calibrate the Chl-
a estimation model, instead of an
in situ dataset. It thus improved the applicability of the model. In addition, a different assumption—
i.e., concentrations of NAP and CDOM are constants—was adopted to further minimize the effects due to the previous assumptions. The new assumption may not be valid in real waters, but can be valid in simulated cases. Chl-
a estimation models were then prepared in advance for various combinations of NAP and CDOM, which were increased in small increments, and saved in a look-up-table (LUT). An iterative search strategy was used to obtain the most appropriate Chl-
a estimation model for a given pixel. A more detailed description of the SAMO-LUT algorithm can be found in Section 3 and in the previous study [
11].
Although the SAMO-LUT algorithm has been validated in Lake Dianchi, China, and Lake Kasumigaura, Japan, further validation is necessary to determine both the advantages and potential limitations of the algorithm. The present study had three objectives: (1) to evaluate the performance of the SAMO-LUT algorithm using a more extensive dataset collected from five Asian lakes (three in Japan and two in China); (2) to determine whether the fixed Specific Inherent Optical Properties (SIOPs, collected from Lake Dianchi) affected the accuracy of the SAMO-LUT algorithm in Chl-a estimation for other lakes; and (3) to provide recommendations for the operational application of the SAMO-LUT algorithm for remote monitoring of Chl-a in inland waters.
2. Study Areas
The data used in this study were collected from five Asian lakes (
Figure 1), which cover trophic categories from oligotrophic to hypertrophic. The first is Lake Biwa, which is located in the western part of Japan (35.33°N, 136.17°E, Shiga Prefecture). It is the largest freshwater lake in Japan with a surface area of 670 km
2, a maximum depth of 104 m, and an average depth of 41 m. Lake Biwa serves as reservoir for the cities of Kyoto and Ōtsu, and provides drinking water for about 15 million people in the Kansai region. It is also a valuable water resource for many kinds of nearby textile industries. Water quality in the lake is currently good, and it belongs to the oligotrophic category.
The second is Lake Suwa, which is located in the central part of Japan (36.05°N, 138.08°E, Nagano Prefecture). It has a surface area of 13.3 km2, an average depth of 4.7 m, and a maximum depth of 7.2 m. In the 1960s, the lake underwent a very rapid hypertrophication. This was caused by the spectacular growth of industrial activity around the lake, and is indicated by heavy blooms of blue-green algae. Water quality has been remarkably improved since the end of 1990s, thanks to effective management. Therefore, the current trophic status of the lake is close to the boundary between mesotrophic and eutrophic.
The third lake is Lake Kasumigarua, situated in the eastern part of Japan’s Kanto plain (36.03°N, 140.40°E, Ibaraki Prefecture). It is Japan’s second largest lake, with a surface area of 171 km
2, an average depth of 4 m, and a maximum depth of 7.3 m (only for Nishiura). This lake is considered hypertrophic because of its high loads of nutrients and shallow depth [
13]. Although average Chl-
a has decreased from 87 to 61 mg·m
−3 during the past three decades, the mean total phosphorus concentration increased from 116 to 138 mg·m
−3. Secchi disk depth decreased from 70 to 52 cm in the last twenty years [
14]. Total suspended sediment (TSS) concentrations increased from 14.1 to 26.4 g·m
−3 during the last decade, due mainly to the resuspension of bottom sediments [
15]. Diatoms (e.g.,
Cyclotella sp. or
Synedra sp.) are generally observed during winter, spring, and autumn, while harmful blooms (blue green algae, e.g.,
Microcystis sp. or
Anabaena sp.) are sometime observed during summer. The concentration of dissolved organic carbon (DOC), which is often correlated with CDOM concentration, is always low (2.9–4.2 g·m
−3; [
13,
16]) compared to lakes such as Lake Taihu (8.8–20.2 g·m
−3 in [
17]) and Finnish lakes (6.0–12.3 g·m
−3 in [
18]). The absorption coefficients of CDOM at 420 nm ranged from 0.5 to 0.6 m
−1 when DOC concentrations ranged from 1.9 to 2.7 g·m
−3 [
14], which was lower than the absorption coefficients of CDOM at 420 nm in Finnish lakes (1.7–7.7 m
−1; [
18]).
The fourth lake in our database is Lake Dianchi, located in a plateau area of the southwestern part of China (24.83°N, 102.72°E). It has a surface area of 300 km
2 and is the largest lake in Yunnan Province, with an average depth of 4.3 m, and a maximum depth of 11.3 m. Eutrophication has become more and more serious in the lake over the past 20 years due to the large quantities of industrial wastewater and municipal sewage discharged into it. Algal blooms occur frequently from April to November each year [
19]. The trophic status of Lake Dianchi also belongs to the hypertrophic category.
The fifth and final lake is Lake Erhai (25.82°N, 100.18°E), which is the second largest lake in the Yunnan Province of China. It has a surface area of 249 km
2, an average depth of 11 m, and a maximum depth of 21 m. Lake Erhai is an important drinking water resource for the local people, supplying drinking water of 8.3 × 10
4 m
3 per day. It is also utilized for local industries, irrigation, and domestic water in the coastal area. During the past 30 years of rapid economic development and increasing population, Lake Erhai has faced a serious threat of intensive eutrophication due to anthropogenic inputs and overuse [
20]. The water quality status of Lake Erhai is already at the initial stage of eutrophication, with organic matter and phytoplankton biomass increasing rapidly, and cyanobacteria blooms breaking out in the embayment and some parts of the lake.
5. Discussion
One advantage of the SAMO-LUT algorithm is that the model calibration process requires a minimum of field data [
11]. In the SAMO-LUT, only the SIOPs of water constituents must be collected from the target water. After that, a comprehensive synthetic dataset is generated based on the obtained SIOPs and a bio-optical model; and the LUTs are constructed to save the coefficients of water constituent estimation models (e.g., Chl-
a estimation models). Yang
et al. suggested that the LUTs should be reconstructed if the SAMO-LUT algorithm was applied to a water body with different SIOPs [
11]. In practice, the SIOPs do change according to the dominant phytoplankton species in water bodies. For example, the dominant phytoplankton species changed seasonally in Lake Kasumigaura [
14]. Generally,
Cyanophyceae is the dominant species during the summer season, while
Bacillariphyceae is dominant from autumn to spring. Therefore, the SIOPs changed accordingly [
30]. The dominant phytoplankton species in Lakes Erhai and Dianchi were
Cyanophyceae during the field surveys [
31].
In this study, the SIOPs collected from Lake Dianchi in July 2009 were used to construct LUTs, which were then applied to other periods of Lake Dianchi and other lakes without any adjustment. The results in Section 4.2 showed that the SAMO-LUT algorithm was able to achieve acceptable accuracy for all tested cases, with NMAE less than 20%, except for Lake Biwa. These findings indicate that not only did the variation of SIOPs not affect the performance of the SAMO-LUT dramatically, but also that the use of a comprehensive synthetic dataset for model calibration can prevent the uncertainty in estimation models caused by sampling bias. Our results thus present a strong case for the use of the SAMO-LUT algorithm as an operational method for estimating Chl-a from remote sensing data. Moreover, the best performance of the SAMO-LUT algorithm in Lake Dianchi and its slightly degraded performance in Lake Kasumigaura in February 2006 suggest that the algorithm has the potential to provide more accurate Chl-a estimation if the SIOPs of the target water were available.
We do not believe, however, that the failure of the SAMO-LUT algorithm in the cases of Lakes Biwa and Suwa were caused by the differences in SIOPs between them and Lake Dianchi. In the other words, the performance of the SAMO-LUT algorithm would not have been satisfactorily improved had SIOPs been collected from these two lakes. From
Figure 2 it can be seen that the
Rrs in the red-NIR spectral region of Lakes Biwa and Suwa are much lower than those of the other three lakes. Weak signals from water bodies with too low TSS will be hidden by environmental noise, and thus cannot provide enough information for estimating water constituents. Therefore, the SNR at selected red-NIR bands used in the SAMO-LUT algorithm are not applicable to relatively clear waters. Alternatively, for clear waters with negligible effects of NAP and CDOM, a blue-green algorithm (e.g., OC4E) can be used to replace the NIR-red algorithms (e.g., Lakes Biwa and Suwa) because the signals in blue-green spectral regions remain strong even for this kind of water. A classification method for selecting an appropriate algorithm in the case of different types of water is needed.
On the other hand, the SAMO-LUT algorithm was developed based on the bandwidths of Mediun Resolution Imaging Spectrometer (MERIS). The most important prerequisite is the band around 708 nm, as used in the three-band index. Therefore, it can be directly applied to water-color sensors with configurations similar to MERIS, such as the Ocean and Land Color Instrument (OLCI, [
32]), the Hyperspectral Infrared Imager (HyspIRI, [
33]) or the Airborne Imaging Spectrometer for Applications (AISA, [
5]). Extension of the SAMO-LUT method to more satellite sensors without the 708 nm band, such as the Moderate Resolution Imaging Spectroradiometer (MODIS), and the Sea-viewing Wide Field-of-view Sensor (SeaWiFS), will be carried out in future studies.