Next Article in Journal
Warehouse Drone: Indoor Positioning and Product Counter with Virtual Fiducial Markers
Next Article in Special Issue
Yield Prediction of Four Bean (Phaseolus vulgaris) Cultivars Using Vegetation Indices Based on Multispectral Images from UAV in an Arid Zone of Peru
Previous Article in Journal
Exploring the Feasibility of Mitigating Flood Hazards by an Existing Pond System in Taoyuan, Taiwan
Previous Article in Special Issue
Assessing the Temporal and Spatial Variability of Coffee Plantation Using RPA-Based RGB Imaging
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Drones
Volume 7
Issue 1
10.3390/drones7010002
drones-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Water Chlorophyll a Estimation Using UAV-Based Multispectral Data and Machine Learning

by
Xiyong Zhao
1,2,3,
Yanzhou Li
1,
Yongli Chen
2,
Xi Qiao
1,3,* and
Wanqiang Qian
3,*
1
College of Mechanical Engineering, Guangxi University, Nanning 530004, China
2
Guangxi Bossco Environmental Protection Technology Co., Ltd., Nanning 530007, China
3
Shenzhen Branch, Guangdong Laboratory for Lingnan Modern Agriculture, Genome Analysis Laboratory of the Ministry of Agriculture and Rural Affairs, Agricultural Genomics Institute at Shenzhen, Chinese Academy of Agricultural Sciences, Shenzhen 518120, China
*
Authors to whom correspondence should be addressed.
Drones 2023, 7(1), 2; https://doi.org/10.3390/drones7010002
Submission received: 26 November 2022 / Revised: 14 December 2022 / Accepted: 16 December 2022 / Published: 21 December 2022
(This article belongs to the Special Issue Yield Prediction Using Data from Unmanned Aerial Vehicles)
Browse Figures
Review Reports Versions Notes

Abstract

:
Chlorophyll a (chl-a) concentration is an important parameter for evaluating the degree of water eutrophication. Monitoring it accurately through remote sensing is thus of great significance for early warnings of water eutrophication, and the inversion of water quality from UAV images has attracted more and more attention. In this study, a regression method to estimate chl-a was proposed; it used a small multispectral UAV to collect data and took the vegetation indices as intermediate variables. For this purpose, ten monitoring points were selected in Erhai Lake, China, and two months of monitoring and data collection were conducted during a cyanobacterial bloom period. Finally, 155 sets of valid data were obtained. The imaging data were obtained using a multispectral UAV, water samples were collected from the lake, and the chl-a concentration was obtained in the laboratory. Then, the images were preprocessed to extract the information from different wavebands. The univariate regression of each vegetation index and the regression using band information were used for comparative analysis. Four machine learning algorithms were used to build the model: support vector machine (SVM), random forest (RF), extreme learning machine (ELM), and convolutional neural network (CNN). The results showed that the effect of estimating the chl-a concentration via multiple regression using vegetation indices was generally better than that via regression with a single vegetation index and original band information. The CNN model obtained the best results (R2 = 0.7917, RMSE = 8.7660, and MRE = 0.2461). This study showed the reliability of using multiple regression based on vegetation indices to estimate the chl-a of surface water.

1. Introduction

Since eutrophication began to appear in inland lakes in the 1930s, approximately 40–50% of lakes and reservoirs in the world have been affected by eutrophication to varying degrees, and lake eutrophication has become one of the most intractable water environmental problems [1]. Chlorophyll a (chl-a) concentration is an important indicator to measure the abundance of phytoplankton in lake water and an important parameter to determine water quality, biological status, and eutrophication degree [2,3]. Chl-a, which is a photosynthetic pigment found in algae species [4], has been considered the largest weighted factor to calculate the Carlson trophic state indices [5]. Assessments of the timing and extent of algal biomass and eutrophication states in aquatic ecosystems may frequently be performed using extensive long-term chl-a concentration measurements [6]. These eutrophication insights can help us to understand primary production, biogeochemical cycling, and overall inland water quality, which can result in environmental changes and useful mitigation tactics [7,8].
Using remote sensing technology to measure a series of parameters indicating water quality (temperature, transparency, turbidity, photosynthetic pigment concentration, colored dissolved organic matter, etc.), it was shown that this variability in aquatic ecosystems can be effectively captured [9,10]. Data sources are mainly multispectral sensors carried by satellites, such as Moderate Resolution Imaging Spectroradiometer (MODIS) Aqua and Terra, Medium Resolution Imaging Spectrometer (MERIS), Operational Land Imager (OLI), and Sentinel 2-A/B [11,12,13,14,15,16,17,18,19]. However, low spatial and temporal resolution, atmospheric correction challenges, cloud cover, slow data–product turn-around times, and the cost of some satellite products can limit the application of satellite remote sensors for monitoring water quality in inland systems [20]. In addition, the concentration of chl-a at the water surface varies at different times of the same day. It increases with the temperature and is usually highest at noon. This also leads to space-time errors between water sample data and satellite data during a study [21]. Unmanned aerial systems (UASs) (remotely piloted remote sensing platforms, i.e., UAVs), show great potential for bridging the gap between in situ water sampling and satellite remote sensing. UASs collect high-resolution aerial data with minimal atmospheric disturbance from cloud coverage, allow flexible flight planning with rapid turn-around times, and are available in a variety of wavelength combinations [22]. There are already many commercial UAVs equipped with multiband and multispectral sensors, their performance is stable, and they can meet daily monitoring needs.
In traditional chl-a remote sensing monitoring research, according to the different reflectivities of chl-a in different spectral bands, some band algorithms for chl-a were proposed, such as the single-band threshold, band interpolation algorithms, and band ratio algorithms [23,24], which are simple and suitable for general estimation. The three-band method [25,26] and four-band method [27] weaken the influence of turbid water on chl-a and improve the accuracy of the model. However, the band algorithm requires specific band data; therefore, it requires special multispectral or hyperspectral sensors. It is widely used in satellite remote sensing monitoring, and large UAVs are required to carry hyperspectral equipment in UAV monitoring.
With the development of artificial intelligence, machine learning has become an important technology in remote sensing image processing [28,29,30], and has also been widely used in water quality monitoring [31]. In many chl-a remote sensing monitoring studies, Peterson et al. [32,33] used an extreme learning machine (ELM) to estimate chl-a and obtained good results. An ELM also showed good performance in organic carbon estimation [34]. Philipp M et al. used a convolutional neural network (CNN) to invert chl-a and obtained promising results, which were better than those from the Bayesian regularization (BR) algorithm model [35,36]. Many scholars have used support vector machines (SVMs) and random forests (RFs) to estimate chl-a, and all of them achieved good results [37,38,39,40]. The effect is better than that of the models established using artificial neural networks (ANNs) [41,42], dynamic programming (DP) [43], generalized linear models (GLMs) [44], and other algorithms. Therefore, we chose SVM, RF, ELM, and CNN to be evaluated.
At present, the estimation model of chl-a is more established by using regression of the band reflectivity [45,46,47]. In the monitoring of chl-a using a small multispectral UAV, multiband multispectral sensors mainly include the red (R), green (G), blue (B), red-edge (RE), and near-infrared (NIR) bands, which can be used to calculate the vegetation indices to represent the concentration of chl-a. Based on G. Edna’s correlation analysis on different vegetation indices and chl-a [48], this study proposed a method to estimate chl-a concentration by using vegetation indices as an intermediate variable, which provided a new idea for small multispectral UAV monitoring of chl-a.
In many studies of chl-a remote sensing inversion, many of them analyzed image data through existing algorithms, but only used a small amount of data as validation, and lacked a lot of actual data to support their claims [27,49]. Most of the regression algorithms used to predict chl-a adopt the method of multiple sampling points in one experiment to collect data, which has a small amount of data and is in the same environment [42,50]. The lack of long-term monitoring and data collection makes the performance of the model worse in complex environments.
In light of this context, this study aimed to monitor the chl-a concentration and multispectral data of Erhai Lake (Yunnan, China) for two months. The main objectives of this study were as follows: (1) explore the feasibility of using vegetation indices as intermediate variables to estimate the concentration of chl-a; (2) the effect of using vegetation indices as an intermediate variable to estimate chl-a concentration was analyzed compared with that of using single vegetation index and original variable; (3) analyze the applicability of SVM, RF, ELM, and CNN machine learning algorithms in the estimation of chl-a concentration.

2. Materials and Methods

2.1. Study Area

The data collection area was located in the northwestern part of Erhai Lake, as shown in Figure 1. Erhai Lake is located in Dali city, Yunnan Province, China (25°36′–25°58′ N, 100°06′–100°18′ E, 1972 m above sea level). It has an average depth of 10.5 m and a surface area of about 251 km2. Since the 1990s, with the development of the social economy, the disturbance from human activities has been increasingly aggravating Erhai Lake, its water quality has deteriorated, a large number of aquatic plants have disappeared, and the eutrophication of water has become gradually obvious, which has gradually degraded the ecological environment of Erhai Lake [51]. The water quality of Erhai Lake is mainly classified as class II in the dry season and class III in the rainy season, and large blooms mainly occur around October.
The concentration of chl-a in the northern coastal area of Erhai Lake was high in previous years. However, the shore water is shallow, the grass is more abundant, and the bottom color has a greater impact on measurements. Therefore, 100 m away from the shore, one sampling point every 120 m and a total of 10 sampling and monitoring points were selected.

2.2. Data Acquisition and Preprocessing

Multispectral image acquisition and lake water samples were collected from 10 monitoring sites over a total of 60 days between September and November 2021. Experiments were conducted every three days. Due to rain and other weather-related reasons, a field experiment was not possible, and images were affected by surface ships and wind; therefore, the data were not sufficiently accurate. Therefore, a total of 18 experiments were conducted and 155 groups of data were collected.
The Phantom 4 Multispectral (P4M) UAV produced by Dajiang Innovation Technology Co., Ltd. (DJI), Shenzhen, China, was used in this study. DJI P4M and its multispectral camera were used for image collection in this experiment. The multispectral sensor has a six-camera array, including a high-definition color sensor and five monochrome sensors: R, G, B, RE, and NIR. It has a light intensity sensor at the top that captures solar irradiance and records it in image files. When the data is post-processed, the solar irradiance data can be used for illumination compensation of the image to eliminate the interference of ambient light on data acquisition; no whiteboard calibration was required. The specifications of DJI P4M are shown in Table 1.
The flight path was delimited on DJI GS Pro software to pass through each monitoring point. The flight altitude was 10 m, the camera angle was −90°, and the exposure was automatic. The resolution of the photos obtained was 1600 × 1300, with an accuracy of approximately 0.53 cm/pixel.
Each experiment began at 1 p.m. when temperatures were the highest and the chl-a concentrations were the highest throughout the day. First, the UAV flew according to the predetermined route to obtain multispectral images. Then, water sample collection was carried out. Water samples were collected at a depth of 20 cm from the surface of the water using a water sampler. Finally, the water quality parameters (chl-a concentration, turbidity, and water temperature) of the water samples were tested in the laboratory.
For the water samples, we used the chl-a concentration module of the HX-200 multi-parameter controller produced by Beijing Hongxinhengce Technology Co., Ltd. (Beijing, China) to analyze the chl-a concentration. It makes use of the characteristics of chl-a with absorption peaks and reflection peaks in the spectrum. The monochromatic light at the absorption peak band of the chl-a spectrum was emitted into the water. The chl-a in the water absorbs the energy of the light and reflects monochromatic light with another wavelength. The light intensity reflected by chl-a is directly proportional to the content of chl-a in the water, which is used to calculate the concentration of chl-a.
To collect higher-resolution images, the UAV was flown at a lower altitude and the surface of the lake would ripple under the influence of the wind, causing the image to be affected by the reflection of light. Therefore, the main processing of the image was the extraction and removal of reflective areas. Finally, the data from non-reflective areas were determined to reduce the influence of reflection on the results.
The fuzzy C-means (FCM) algorithm is a common clustering algorithm based on the fuzzy theory introduced by J. C. Birdek [52]. It maximizes the similarity of data objects within classes and minimizes the similarity of data objects between classes. It is an improvement on the traditional C-means clustering. It divides the samples according to the degree of the image pixels belonging to the clustering center and has been widely used in image segmentation. The reflective area extraction effect of the FCM algorithm is shown in Figure 2.
After the extraction of the reflective region in the visible image, the non-reflective region was covered on the multispectral band image using a mask, and the non-reflective region of the multispectral band image was extracted. Then, the gray value of each multispectral band image was extracted and all gray values of each image were averaged. Finally, the reflectance of different bands was calculated according to the gray value.

2.3. Vegetation Index

According to the vegetation indices (band ratio algorithm) containing the five bands R, G, B, RE, and NIR in relevant studies, 28 common vegetation indices with formulas were finally selected for this research, among which 14 contained only the RGB visible band and 14 contained the RE or NIR band. The vegetation indices and their formulas are shown in Table 2.

2.4. Modeling Techniques

The algorithms used for image preprocessing and regression in this study were all completed in MATLAB software, version 2019a. The parameters in the following models were the results after the gradient test.

2.4.1. Support Vector Machine (SVM)

An SVM is a kind of generalized linear classifier that classifies bivariate data according to supervised learning. Its decision boundary is a hyperplane with a maximum margin to solve the learning samples, and the problem can be transformed into a convex quadratic programming problem [72]. Compared with logistic regression and neural networks, support vector machines provide a clearer and more powerful way to learn complex nonlinear equations. The optimization problem of an SVM considers both empirical risk and structural risk minimization; therefore, it is stable. An SVM is mainly used for classification and regression analysis. The main parameters of the SVM model in this study were as follows: SVM type(-s), 3 (epsilon-SVR); kernel type(-t), 2 (radial basis function); cost(-c), 2.2; gamma(-g), 2.8; epsilon(-p), 0.01. The main parameters of the vegetation index univariate regression, original variable multiple regression, and vegetation indices multiple regression models remained the same.

2.4.2. Random Forest (RF)

In 2001, Leo Breiman combined classification trees into an RF, that is, the use of variables and data were randomized to obtain a certain number of classification trees. Then, the results of the classification trees were summarized, and the random forest algorithm was proposed [73]. An RF is a type of ensemble learning. Ensemble learning is a very popular machine learning strategy, and almost all problems can be improved by using its ideas, and it is more stable when analyzing small sample data. The basic starting point is to gather algorithms and various strategies together. Ensemble learning can be used for both classification problems and regression problems, and it is often seen in the field of machine learning. The main parameters of the RF model in this study were as follows: leaf, 5; tree, 500. The main parameters of the vegetation index univariate regression, original variable multiple regression, and vegetation indices multiple regression models remained the same.

2.4.3. Extreme Learning Machine (ELM)

An ELM is a kind of machine learning system or method based on a feedforward neural network (FNN), which is suitable for both supervised and unsupervised learning problems [74]. The most significant feature of an ELM is that the input weights and the bias of hidden nodes are randomly generated within a given range, which was shown to have high learning efficiency and strong generalization ability. The main purpose of training is to solve the weights of the input layer. An ELM is widely used in classification, regression, clustering, feature learning, and other problems [75]. The main parameters of the ELM model in this study were as follows: number of hidden neurons (N), 30; transfer function (TF), sig (sigmoidal function); type, 0 (regression). The main parameters of the vegetation index univariate regression, original variable multiple regression, and vegetation indices multiple regression models remained the same.

2.4.4. Convolutional Neural Networks (CNN)

A CNN is a class of feedforward neural networks with a deep structure and convolutional computation, which is one of the representative algorithms of deep learning [76]. A CNN imitates biological visual perception mechanism construction, which can carry out supervised learning and unsupervised learning. The convolution kernel parameter shares in the hidden layers and the sparsity of the connections between layers allow the CNN to use grid-like topological features with less computation. The main parameters of the CNN model in this study were as follows. In the univariate regression of the vegetation index, only the final regression layer was used. The multiple regression model included one convolution layer, one pooling layer, and two fully connected layers. Convolution layer: convolution kernel size, 3 × 1; step length, 16; podding, same. Pooling layer: pool kernel size, 2 × 1; step length, 2. The number of neurons in the two fully connected layers: both were 9 in the original multivariate regression model; in the multivariate regression model of vegetation indices, both were 52. Training parameters: epoch, 50; batch size, 4; learn rate, 0.005. The training parameters of all CNN models were the same.

2.5. Modeling Strategy and Validation Metrics

In this study, the performance of the model was evaluated using the coefficient of determination (R2) [56], root-mean-square error (RMSE) [57], and mean relative error (MRE) [51], where the related formulas are as follows:
R 2 = 1 − ∑ i = 1 n y ^ i − y i 2 ∑ i = 1 n y i − y ¯ 2
R M S E = 1 n ∑ i = 1 n y ^ i − y i 2
M R E = ∑ i = 1 n y ^ i − y i y i n
where y i is the measured value, y ^ i is the predicted value, y ¯ is the average of the measured values, and n denotes the number of samples.
The larger the R2 and the smaller the RMSE and MRE, the better the model performance.

2.6. Data Analysis

The maximum, minimum, median, mean, and standard deviation of the observed chl-a concentration data were analyzed statistically. Furthermore, the significance analysis of the estimated concentration of chl-a was carried out. The software used was IBM SPSS Statistics (v. 25.0).

3. Results

3.1. Descriptive Statistics

To build and test the model, the total dataset was randomly divided into a training set and a test set. In this study, the Rand perm function was used for random classification. The original ratio was 3:1, but the number cannot contain decimals, and considering the convenience of calculation and statistics, the number of training sets was 120, and the rest were test sets. The statistical values of the training and test datasets for the chl-a concentration estimation are shown in Table 3. The datasets were analyzed using the maximum, minimum, median, mean, and standard deviation. According to the data in Table 3, the distribution of the whole dataset and training and testing datasets was similar. Therefore, the training set and test set could be used to establish the model and test the model accuracy.

3.2. Univariate Regression Results of Vegetation Indices

Table 4 shows the test results of the univariate regression model established using the four algorithms on 28 vegetation indices. According to the analysis of the best results of each vegetation index among the four algorithms, there were 10 vegetation indices with R2 values greater than 0.5 (9 of which contained NIR bands), 6 of which were between 0.3 and 0.5, and 12 of which were less than 0.3. The DVI achieved the best results using the ELM model (R2=0.6923). In the univariate regression of the vegetation index, the ELM achieved the best results in most vegetation index regressions, especially when the R2 values were greater than 0.5. The second was the RF, which performed better than the CNN and SVM in most cases.

3.3. Original Band Reflectance Regression Results

We used data from the five original bands as independent variables for regression as a reference, and the results are shown in Table 5. The results showed that the ELM model achieved the best results (R2 = 0.7202, RMSE = 8.3023, and MRE = 0.2728), and in terms of the R2, RMSE, and MRE values, the results were significantly better than the models established by other algorithms.

3.4. Multiple Regression Results of the Vegetation Indices

First, linear correlation analysis was conducted between the vegetation indices and the chl-a concentration (represented by R2), and the analysis results are shown in Figure 3. The figure shows that the variable with the highest correlation was the RNDVI, and its R2 value was 0.3435. The RNDVI was originally developed to differentiate between vegetation and soil to estimate crop growth without the need to measure near-infrared wavelengths, and it is not typically included in UAS aquatic ecosystem monitoring studies. There were 12 variables with correlations greater than 0.1, among which 11 vegetation index formulas required NIR band data. Moreover, the 10 vegetation indices with R2 values greater than 0.5 in the univariate regression of the vegetation indices were all in the above 12 vegetation indices.
To explore the difference between the effects of multiple regression with the vegetation indices as the independent variable and univariate regression with the vegetation index, 28 planting indices were used as the independent variables; the multiple regression results of the model that was established using the four regression algorithms are shown in Table 6. The results showed that the CNN model achieved the best results (R2 = 0.7917, RMSE = 8.7660, and MRE = 0.2461), followed by the ELM model, RF model, and SVM model. The results of the evaluation indices R2, RMSE, and MRE values, were consistent with the above conclusions.
The results of the multiple regression test data of the vegetation indices are shown in Figure 4. As seen from the figure, the dispersion degree increased with decreasing R2 values, the CNN model was the most stable, and the dispersion degree was the lowest. Because there were fewer data with high concentrations, the prediction results of the model with high concentrations were not as good as those with low concentrations. In particular, the SVM model had poor performance when predicting high concentrations.
Finally, we used the combination of vegetation indices and original band data as independent variables for regression, and the results are shown in Table 7. The best performance was from the CNN model (R2 = 0.7710, RMSE = 9.1317, and MRE = 0.2485), but its performance was inferior to that of the vegetation index multiple regression model.
In general, compared with the univariate regression of the vegetation index and the original band information regression, the effect of the multiple regression of the vegetation indices was significantly improved. Among the different algorithms, ELM displayed the best performance in the univariate regression of the vegetation index using band information regression, and the CNN displayed the best performance in the multiple vegetation indices variable regression. The CNN multivariate model based on the vegetation indices achieved the best results among all the models (R2 = 0.7917, RMSE = 8.7660, and MRE = 0.2461).

4. Discussion

4.1. Correlation between Vegetation Index and Chl-a

Many multiband multispectral sensors are specifically designed to record the peak reflectance and absorption characteristics of chl-a. The band of the multispectral sensor of the DJI P4M UAV used in this study has some similarities with the characteristic band of chl-a. Studies using UAVs and satellites to monitor water quality also found that vegetation indices including red and near-infrared bands, such as the RVI and NDVI, perform well when estimating chl-a [77,78]. In the univariate regression of the vegetation index in this study, the ELM model of the vegetation index DVI achieved the best result (R2 = 0.6923), and the regression results of RVI and NVDI were also greater than 0.5. The DVI is mainly used to detect vegetation growth status and vegetation coverage; the NDVI is the most widespread vegetation index and it can reflect the background influences of the plant canopy, such as soil and dead leaves [79]; and the RVI is very sensitive to vegetation coverage. Moreover, there were 10 vegetation indices with R2 values greater than 0.5, of which 9 contained the NIR band and 7 contained the NIR and R bands. This was also consistent with the results that water with a higher chl-a concentration is significantly different from ordinary water in the NIR band [80]. The importance of the NIR band in the results also suggested that UAVs equipped with multispectral sensors have advantages in chl-a estimation over UAVs equipped with ordinary HD cameras. In this study, the DVI achieved the best results, but there were slight differences in the best-realized vegetation index among different studies, which indicated that the vegetation indices could produce different results when using data under different circumstances; therefore, it was not rigorous enough to use a single vegetation index to represent the concentration of chl-a.

4.2. Improvement of the Chl-a Estimation Effect Using Multiple Regression of Vegetation Indices

In the multiple regression of the vegetation indices, the CNN model achieved the best result (R2 = 07917, RMSE = 8.7660, and MRE = 0.2461). Compared with the wave segment reflectance, the effect of multiple regression was significantly improved. In addition, among the models established by the four algorithms, only the accuracy of the SVM model decreased (R2 decreased by 0.0073), while the accuracy of the other algorithm models increased, among which the R2 values of the RF increased by 0.0660 and the R2 values of the ELM increased by 0.0608. In particular, the effect of the CNN model was significantly improved (R2 values increased by 0.1319). This indicated that using the vegetation indices as an intermediate variable could explore more information in the data and increase the number of features, which was more conducive to the regression model established by the machine learning algorithm, thus improving the chl-a regression effect. The linear correlation analysis results between the vegetation indices and chl-a concentration showed that the R2 value of the RNDVI was the largest (R2 = 0.3435) and the correlation was generally small; therefore, it is not reliable to use a single vegetation index to represent the chl-a concentration. By using vegetation index univariate regression, there was a nonlinear relationship between the chl-a concentration and R2 values that were significantly improved, but the effect of using vegetation indices multiple regression was better. This also showed that a single vegetation index had many features that could not be involved, and the multiple regression of multiple vegetation indices could enrich these features and make the regression estimation more accurate. Using the combination of vegetation indices and original band data as independent variables, the results showed that the performance was lower than that of the multivariate regression of vegetation indices. It showed that this combination produced some redundancy, resulting in poor model performance. The multivariate regression of vegetation indices was the most suitable for the estimation of chl-a concentration with multispectral data.
Compared with satellites and UAVs equipped with hyperspectral sensors, small multispectral UAVs are easy to operate and save labor and material costs. In the estimation of chl-a with hyperspectral data, most of them used the characteristic bands of chl-a for regression [22,38,78,81]. However, since the band of multispectral data is not completely the same as the characteristic band of chl-a, it becomes the best choice to estimate the concentration of chl-a via regression using vegetation indices as the intermediate variable.

4.3. Adaptability of the Algorithm

In the results of univariate regression using the vegetation indices and multiple regression using band reflectance, the ELM algorithm outperformed the other algorithms. The ELM has a single hidden layer and is considered better than other shallow learning systems, such as the SVM, in terms of the learning rate and generalization ability [74]. When using the vegetation indices for multiple regression, the CNN model achieved the best results. In univariate regression, since the single variable could not be convolved, only the regression layer was used in the CNN algorithm, and its effect was worse than that of other algorithms. When using the vegetation indices for multiple regression, 28 vegetation indices were convolved and, finally, regression was performed. It provided better results than the other algorithms. This shows that when the number of features was large, the CNN could better extract effective features and improve the regression results [82]. An SVM and RF have a certain degree of robustness but do not have a deep structure; in the face of complex situations, the effect will be worse, but is suitable for small-sample analysis [83,84]. A CNN has an automatic feature extraction capability, which is more suitable for analysis in complex situations [85].

4.4. Deficiency and Prospects

In this study, the vegetation indices were used as an intermediate variable to estimate the concentration of chl-a. After verification, this method was more suitable for small multi-spectral UAVs monitoring the concentration of chl-a. In this study, the concentration of chl-a was mainly below 50 μg/L, and there were few data above 50 μg/L. Therefore, the high-concentration data were slightly insufficient in the model training, which could lead to greater error in the prediction of high concentrations compared with low concentrations. In the future, more data will be added to increase the model performance. In this study, 28 vegetation indices were used for analysis. The use of hyperspectral sensors that have a higher spectral resolution can also be considered for use. In this way, more vegetation indices can be calculated [86] so that the chl-a estimation model displays better performance in complex environments.

5. Conclusions

To explore the method of chl-a estimation for small multispectral UAVs, this study proposed a multiple regression method using the vegetation indices as an intermediate variable. Then, the FCM clustering segmentation algorithm was used to extract and remove the reflective area in the image, and then the gray value of other areas of the image was extracted and the mean value was taken. The next step was to calculate the reflectance according to the extracted gray value to calculate 28 vegetation indices. Then, the calculated vegetation index was used as the independent variable and the detected chl-a was used as the dependent variable, and the data set was divided into the training set and the test set using random classification. Finally, the data sets were established and tested using SVM, RF, ELM, and CNN regression algorithms, and R2, RMSE, and MRE were used to evaluate the model performance to identify the model with the best performance. The univariate regression of the vegetation index and the multiple regression of the original band data were used as references. Using the above process, the following findings were obtained in this study:
(1)
It was feasible to establish a chl-a regression estimation model with the vegetation indices as an intermediate variable. Moreover, compared with the previous regression estimation using band data and using a single vegetation index to express chl-a concentration, the effect was significantly improved.
(2)
Previous studies found that, due to different data and other reasons, when a single vegetation index is used to represent chl-a, the best-performing vegetation index is not completely consistent. The best-performing vegetation index in this study was the DVI, which was different from many studies. Therefore, it is not reliable to use a single vegetation index to represent the chl-a concentration. Only by combining multiple vegetation indices can we find more features in the data and make the results more convincing.
(3)
The ELM model displayed the best performance in the univariate regression of the vegetation index and multiple regression of the original band data. In the multiple regression of the vegetation indices, the CNN model obtained the best results. Therefore, in the case of univariate regression and fewer variables, the ELM could analyze data better and faster. In the case of more variables, the CNN could better utilize the advantages of the convolutional layer and extract more features from the data.

Author Contributions

Conceptualization, Y.L.; Methodology, X.Q.; Software, X.Z., X.Q. and W.Q.; Formal analysis, X.Z.; Resources, Y.C.; Data curation, X.Z. and Y.L.; Writing – original draft, X.Z.; Writing – review & editing, X.Z., X.Q. and W.Q.; Project administration, Y.L.; Funding acquisition, Y.L. and Y.C. All authors have read and agreed to the published version of the manuscript.

Funding

The work in this study was supported by the National Key R&D Program of China (2021YFD1400100 and 2021YFD1400101) and the Guangxi Ba-Gui Scholars Program of China (2019A33).

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Wei, J.; Ji, X.; Hu, W. Characteristics of Phytoplankton Production in Wet and Dry Seasons in Hyper-Eutrophic Lake Taihu, China. Sustainability 2022, 14, 11216. [Google Scholar] [CrossRef]
  2. Le Coz, M.; Chambord, S.; Meire, P.; Maris, T.; Azémar, F.; Ovaert, J.; Buffan-Dubau, E.; Kromkamp, J.C.; Sossou, A.C.; Prygiel, J.; et al. Test of some ecological concepts on the longitudinal distribution of zooplankton along a lowland water course. Hydrobiologia 2017, 802, 175–198. [Google Scholar] [CrossRef]
  3. Pirasteh, S.; Mollaee, S.; Fatholahi, S.N.; Li, J. Estimation of Phytoplankton Chlorophyll-a Concentrations in the Western Basin of Lake Erie Using Sentinel-2 and Sentinel-3 Data. Can. J. Remote Sens. 2020, 46, 585–602. [Google Scholar] [CrossRef]
  4. Cao, Z.; Ma, R.; Duan, H.; Pahlevan, N.; Melack, J.; Shen, M.; Xue, K. A machine learning approach to estimate chlorophyll-a from Landsat-8 measurements in inland lakes. Remote Sens. Environ. 2020, 248, 111974. [Google Scholar] [CrossRef]
  5. Carlson, R.E. A trophic state index for lakes. Limnol. Oceanogr. 1977, 22, 361–369. [Google Scholar] [CrossRef] [Green Version]
  6. O’Reilly, J.E.; Maritorena, S.; Mitchell, B.G.; Siegel, D.A.; Carder, K.L.; Garver, S.A.; Kahru, M.; McClain, C. Ocean color chlorophyll algorithms for SeaWiFS. J. Geophys. Res. Ocean. 1998, 103, 24937–24953. [Google Scholar] [CrossRef] [Green Version]
  7. Mishra, S.; Mishra, D.R. Normalized difference chlorophyll index: A novel model for remote estimation of chlorophyll-a concentration in turbid productive waters. Remote Sens. Environ. 2012, 117, 394–406. [Google Scholar] [CrossRef]
  8. Smith, M.E.; Lain, L.R.; Bernard, S. An optimized Chlorophyll a switching algorithm for MERIS and OLCI in phytoplankton-dominated waters. Remote Sens. Environ. 2018, 215, 217–227. [Google Scholar] [CrossRef]
  9. Woolway, R.I.; Kraemer, B.M.; Lenters, J.D.; Merchant, C.J.; O’Reilly, C.M.; Sharma, S. Global lake responses to climate change. Nat. Rev. Earth Environ. 2020, 1, 388–403. [Google Scholar] [CrossRef]
  10. Ho, J.C.; Michalak, A.M.; Pahlevan, N. Widespread global increase in intense lake phytoplankton blooms since the 1980s. Nature 2019, 574, 667–670. [Google Scholar] [CrossRef]
  11. Brezonik, P.; Menken, K.D.; Bauer, M. Landsat-based Remote Sensing of Lake Water Quality Characteristics, Including Chlorophyll and Colored Dissolved Organic Matter (CDOM). Lake Reserv. Manag. 2005, 21, 373–382. [Google Scholar] [CrossRef]
  12. Kutser, T.; Metsamaa, L.; Strömbeck, N.; Vahtmäe, E. Monitoring cyanobacterial blooms by satellite remote sensing. Estuar. Coast. Shelf Sci. 2006, 67, 303–312. [Google Scholar] [CrossRef]
  13. Moses, W.J.; Gitelson, A.A.; Berdnikov, S.; Povazhnyy, V. Satellite Estimation of Chlorophyll-$a$ Concentration Using the Red and NIR Bands of MERIS—The Azov Sea Case Study. IEEE Geosci. Remote Sens. Lett. 2009, 6, 845–849. [Google Scholar] [CrossRef]
  14. Matthews, M.W. A current review of empirical procedures of remote sensing in inland and near-coastal transitional waters. Int. J. Remote Sens. 2011, 32, 6855–6899. [Google Scholar] [CrossRef]
  15. Matthews, M.W. Eutrophication and cyanobacterial blooms in South African inland waters: 10years of MERIS observations. Remote Sens. Environ. 2014, 155, 161–177. [Google Scholar] [CrossRef]
  16. Shi, K.; Zhang, Y.; Liu, X.; Wang, M.; Qin, B. Remote sensing of diffuse attenuation coefficient of photosynthetically active radiation in Lake Taihu using MERIS data. Remote Sens. Environ. 2014, 140, 365–377. [Google Scholar] [CrossRef]
  17. Giardino, C.; Bresciani, M.; Stroppiana, D.; Oggioni, A.; Morabito, G. Optical remote sensing of lakes: An overview on Lake Maggiore. J. Limnol. 2014, 73, 201–214. [Google Scholar] [CrossRef] [Green Version]
  18. Dörnhöfer, K.; Oppelt, N. Remote sensing for lake research and monitoring—Recent advances. Ecol. Indic. 2016, 64, 105–122. [Google Scholar] [CrossRef]
  19. Ansper, A.; Alikas, K. Retrieval of Chlorophyll a from Sentinel-2 MSI Data for the European Union Water Framework Directive Reporting Purposes. Remote Sens. 2018, 11, 64. [Google Scholar] [CrossRef] [Green Version]
  20. Lomax, A.S.; Corso, W.; Etro, J.F. Employing Unmanned Aerial Vehicles (UAVs) as an Element of the Integrated Ocean Observing System. In Proceedings of the OCEANS 2005 MTS/IEEE, Washington, DC, USA, 17–23 September 2005; 2005; pp. 184–190. [Google Scholar]
  21. Behrangi, A.; Wen, Y. On the Spatial and Temporal Sampling Errors of Remotely Sensed Precipitation Products. Remote Sens. 2017, 9, 1127. [Google Scholar] [CrossRef]
  22. Kislik, C.; Dronova, I.; Kelly, M. UAVs in Support of Algal Bloom Research: A Review of Current Applications and Future Opportunities. Drones 2018, 2, 35. [Google Scholar] [CrossRef] [Green Version]
  23. Allee, R.J.; Johnson, J.E. Use of satellite imagery to estimate surface chlorophyll a and Secchi disc depth of Bull Shoals Reservoir, Arkansas, USA. Int. J. Remote Sens. 1999, 20, 1057–1072. [Google Scholar] [CrossRef]
  24. Binding, C.E.; Greenberg, T.A.; Bukata, R.P. Time series analysis of algal blooms in Lake of the Woods using the MERIS maximum chlorophyll index. J. Plankton Res. 2011, 33, 1847–1852. [Google Scholar] [CrossRef] [Green Version]
  25. Zhang, Y.; Feng, L.; Li, J.; Luo, L.; Yin, Y.; Liu, M.; Li, Y. Seasonal-spatial variation and remote sensing of phytoplankton absorption in Lake Taihu, a large eutrophic and shallow lake in China. J. Plankton Res. 2010, 32, 1023–1037. [Google Scholar] [CrossRef] [Green Version]
  26. Duan, H.; Ma, R.; Hu, C. Evaluation of remote sensing algorithms for cyanobacterial pigment retrievals during spring bloom formation in several lakes of East China. Remote Sens. Environ. 2012, 126, 126–135. [Google Scholar] [CrossRef]
  27. Tao, M.; Duan, H.; Qi, L.; Zhang, Y.; Ma, R. An operational algorithm to estimate chlorophyll-a concentrations in Lake Chaohu from MODIS imagery. J. Lake Sci. 2015, 27, 1140–1150. [Google Scholar]
  28. Mountrakis, G.; Im, J.; Ogole, C. Support vector machines in remote sensing: A review. ISPRS J. Photogramm. Remote Sens. 2011, 66, 247–259. [Google Scholar] [CrossRef]
  29. Pahlevan, N.; Smith, B.; Schalles, J.; Binding, C.; Cao, Z.; Ma, R.; Alikas, K.; Kangro, K.; Gurlin, D.; Hà, N.; et al. Seamless retrievals of chlorophyll-a from Sentinel-2 (MSI) and Sentinel-3 (OLCI) in inland and coastal waters: A machine-learning approach. Remote Sens. Environ. 2020, 240, 111604. [Google Scholar] [CrossRef]
  30. Guo, Y.; Fu, Y.; Hao, F.; Zhang, X.; Wu, W.; Jin, X.; Bryant, C.R.; Senthilnath, J. Integrated phenology and climate in rice yields prediction using machine learning methods. Ecol. Indic. 2021, 120, 106935. [Google Scholar] [CrossRef]
  31. Hassan, N.; Woo, C.S. Machine Learning Application in Water Quality Using Satellite Data. IOP Conf. Ser. Earth Environ. Sci. 2021, 842, 12018. [Google Scholar] [CrossRef]
  32. Peterson, K.T.; Sagan, V.; Sloan, J.J. Deep learning-based water quality estimation and anomaly detection using Landsat-8/Sentinel-2 virtual constellation and cloud computing. GIScience Remote Sens. 2020, 57, 510–525. [Google Scholar] [CrossRef]
  33. Zhang, Y.; Gao, Q. Prediction method of Chlorophyll-a concentration in seawater based on extreme learning machine regression. Mar. Environ. Sci. 2015, 34, 107–112. [Google Scholar]
  34. Chang, N.B.; Imen, S. Improving the control of water treatment plant with remote sensing-based water quality forecasting model. In Proceedings of the 2015 IEEE 12th International Conference on Networking, Sensing and Control, Taipei, Taiwan, 9–11 April 2015; 2015; pp. 51–57. [Google Scholar]
  35. Maier, P.M.; Keller, S.; Hinz, S. Deep Learning with WASI Simulation Data for Estimating Chlorophyll a Concentration of Inland Water Bodies. Remote Sens. 2021, 13, 718. [Google Scholar] [CrossRef]
  36. Yu, B.; Xu, L.; Peng, J.; Hu, Z.; Wong, A. Global chlorophyll-a concentration estimation from moderate resolution imaging spectroradiometer using convolutional neural networks. J. Appl. Remote Sens. 2020, 14, 34520. [Google Scholar] [CrossRef]
  37. Martinez, E.; Gorgues, T.; Lengaigne, M.; Fontana, C.; Sauzède, R.; Menkes, C.; Uitz, J.; Di Lorenzo, E.; Fablet, R. Reconstructing Global Chlorophyll-a Variations Using a Non-linear Statistical Approach. Front. Mar. Sci. 2020, 7, 464. [Google Scholar] [CrossRef]
  38. Guan, Q.; Feng, L.; Hou, X.; Schurgers, G.; Zheng, Y.; Tang, J. Eutrophication changes in fifty large lakes on the Yangtze Plain of China derived from MERIS and OLCI observations. Remote Sens. Environ. 2020, 246, 111890. [Google Scholar] [CrossRef]
  39. Park, J.; Kim, H.-C.; Bae, D.; Jo, Y.-H. Data Reconstruction for Remotely Sensed Chlorophyll-a Concentration in the Ross Sea Using Ensemble-Based Machine Learning. Remote Sens. 2020, 12, 1898. [Google Scholar] [CrossRef]
  40. Hu, C.; Feng, L.; Guan, Q. A Machine Learning Approach to Estimate Surface Chlorophyll a Concentrations in Global Oceans From Satellite Measurements. IEEE Trans. Geosci. Remote Sens. 2021, 59, 4590–4607. [Google Scholar] [CrossRef]
  41. Tenjo, C.; Ruiz-Verdú, A.; Van Wittenberghe, S.; Delegido, J.; Moreno, J. A New Algorithm for the Retrieval of Sun Induced Chlorophyll Fluorescence of Water Bodies Exploiting the Detailed Spectral Shape of Water-Leaving Radiance. Remote Sens. 2021, 13, 329. [Google Scholar] [CrossRef]
  42. Kown, Y.; Baek, S.; Lim, Y.; Pyo, J.; Ligaray, M.; Park, Y.; Cho, K. Monitoring Coastal Chlorophyll-a Concentrations in Coastal Areas Using Machine Learning Models. Water 2018, 10, 1020. [Google Scholar] [CrossRef] [Green Version]
  43. Campbell, A.M.; Racault, M.F.; Goult, S.; Laurenson, A. Cholera Risk: A Machine Learning Approach Applied to Essential Climate Variables. Int. J. Env. Res. Public Health 2020, 17, 49378. [Google Scholar] [CrossRef] [PubMed]
  44. DeLuca, N.M.; Zaitchik, B.F.; Guikema, S.D.; Jacobs, J.M.; Davis, B.J.K.; Curriero, F.C. Evaluation of remotely sensed prediction and forecast models for Vibrio parahaemolyticus in the Chesapeake Bay. Remote Sens. Environ. 2020, 250, 112016. [Google Scholar] [CrossRef]
  45. Ouma, Y.O.; Noor, K.; Herbert, K. Modelling Reservoir Chlorophyll-a, TSS, and Turbidity Using Sentinel-2A MSI and Landsat-8 OLI Satellite Sensors with Empirical Multivariate Regression. J. Sens. 2020, 2020, 8858408. [Google Scholar] [CrossRef]
  46. Patra, P.P.; Dubey, S.K.; Trivedi, R.K.; Sahu, S.K.; Rout, S.K. Estimation of chlorophyll-a concentration and trophic states in Nalban Lake of East Kolkata Wetland, India from Landsat 8 OLI data. Spat. Inf. Res. 2017, 25, 75–87. [Google Scholar] [CrossRef]
  47. Hansen, C.H.; Williams, G.P.; Adjei, Z. Long-Term Application of Remote Sensing Chlorophyll Detection Models: Jordanelle Reservoir Case Study. Nat. Resour. 2015, 6, 123–129. [Google Scholar] [CrossRef] [Green Version]
  48. Fernandez-Figueroa, E.G.; Wilson, A.E.; Rogers, S.R. Commercially available unoccupied aerial systems for monitoring harmful algal blooms: A comparative study. Limnol. Oceanogr. Methods 2021, 20, 146–158. [Google Scholar] [CrossRef]
  49. Guimarães, T.; Veronez, M.; Koste, E.; Gonzaga, L.; Bordin, F.; Inocencio, L.; Larocca, A.P.; De Oliveira, M.Z.; Vitti, D.C.; Mauad, F.F. An Alternative Method of Spatial Autocorrelation for Chlorophyll Detection in Water Bodies Using Remote Sensing. Sustainability 2017, 9, 30416. [Google Scholar] [CrossRef] [Green Version]
  50. Xue, Y.; Zhu, L.; Zou, B.; Wen, Y.-M.; Long, Y.-H.; Zhou, S.-L. Research on Inversion Mechanism of Chlorophyll—A Concentration in Water Bodies Using a Convolutional Neural Network Model. Water 2021, 13, 664. [Google Scholar] [CrossRef]
  51. Zhao, X.; Li, Y.; Chen, Y.; Qiao, X.A. Method of Cyanobacterial Concentrations Prediction Using Multispectral Images. Sustainability 2022, 14, 12784. [Google Scholar] [CrossRef]
  52. Gao, T.; Li, A.; Meng, F. Research on Data Stream Clustering Based on FCM Algorithm. Procedia Comput. Sci. 2017, 122, 595–602. [Google Scholar] [CrossRef]
  53. Shen, G.R.; Sarris, A.; Huang, X.M. Remote sensing monitoring and analysis of expansion characteristics of Chongming Island in Shanghai, China. In Proceedings of the Second International Conference on Agro-Geoinformatics, Fairfax, VA, USA, 12–16 August 2013. [Google Scholar]
  54. Cui, W.; Li, J.; Si, Q.; Wang, R.; Luo, X.; Yang, F.; Li, Y. Research on Extraction Method of Desert Shrub Coverage Based on UAV Visible Light Data. Res. Soil Water Conserv. 2021, 28, 175. [Google Scholar]
  55. Gong, P.; Pu, R.L.; Biging, G.S.; Larrieu, M.R. Estimation of forest leaf area index using vegetation indices derived from Hyperion hyperspectral data. IEEE Trans. Geosci. Remote Sens. 2003, 41, 1355–1362. [Google Scholar] [CrossRef] [Green Version]
  56. Xu, B.; Liang, C.; Chai, D.; Shi, W.; Sun, G. Inversion of Natural Grassland Productivity from Remote Sensor Imagery in Zulihe River Basin. Arid Zone Res. 2014, 31, 1147–1152. [Google Scholar]
  57. Xiao, D.N.; Zhou, Z.F.; Yin, L.J.; Huang, D.H.; Zhang, Y.; Li, Q.X. Identification of Single Plant of Karst Mountain Pitaya by Fusion of Color Index and Spatial Structure. Laser Optoelectron. Prog. 2022, 59, 8151. [Google Scholar]
  58. Chea, C.; Saengprachatanarug, K.; Posom, J.; Wongphati, M.; Taira. Sugar Yield Parameters and Fiber Prediction in Sugarcane Fields Using a Multispectral Camera Mounted on a Small Unmanned Aerial System (UAS). Sugar Tech. 2020, 22, 605–621. [Google Scholar] [CrossRef]
  59. Gitelson, A.A.; Kaufman, Y.J.; Stark, R.; Rundquist, D. Novel algorithms for remote estimation of vegetation fraction. Remote Sens. Environ. 2002, 80, 76–87. [Google Scholar] [CrossRef] [Green Version]
  60. Hague, T.; Tillett, N.D.; Wheeler, H. Automated crop and weed monitoring in widely spaced cereals. Precis. Agric. 2006, 7, 21–32. [Google Scholar] [CrossRef]
  61. Li, P.; Guo, X.; Gu, Q.; Zhang, X.; Feng, C.; Guo, G. Vegetation coverage information extraction of mine dump slope in Wuhai City of Inner Mongolia based on visible vegetation index. J. Beijing For. Univ. 2020, 42, 102–112. [Google Scholar]
  62. Duan, B.; Fang, S.H.; Gong, Y.; Peng, Y.; Wu, X.T.; Zhu, R.S. Remote estimation of grain yield based on UAV data in different rice cultivars under contrasting climatic zone. Field Crops Res. 2021, 267, 108148. [Google Scholar] [CrossRef]
  63. Lukas, V.; Hunady, I.; Kintl, A.; Mezera, J.; Hammerschmiedt, T.; Sobotkova, J.; Brtnicky, M.; Elbl, J. Using UAV to Identify the Optimal Vegetation Index for Yield Prediction of Oil Seed Rape (Brassica napus L.) at the Flowering Stage. Remote Sens. 2022, 14, 4953. [Google Scholar] [CrossRef]
  64. Kataoka, T.; Kaneko, T.; Okamoto, H.; Hata, S. Crop growth estimation system using machine vision. In Proceedings of the IEEE/ASME International Conference on Advanced Intelligent Mechatronics, Kobe, Japan, 20–24 July 2003. [Google Scholar]
  65. Song, B.; Park, K. Detection of Aquatic Plants Using Multispectral UAV Imagery and Vegetation Index. Remote Sens. 2020, 12, 387. [Google Scholar] [CrossRef] [Green Version]
  66. Strong, C.J.; Burnside, N.G.; Llewellyn, D. The potential of small-Unmanned Aircraft Systems for the rapid detection of threatened unimproved grassland communities using an Enhanced Normalized Difference Vegetation Index. PLoS ONE. 2017, 12, 0186193. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  67. Wei, Q.; Zhang, B.; Wei, Z.; Han, X.; Duan, C. Estimation of Canopy Chlorophyll Content in Winter Wheat by UAV Multispectral Remote Sensing. J. Triticeae Crops 2020, 40, 365–372. [Google Scholar]
  68. Du, M.M.; Noguchi, N. Monitoring of Wheat Growth Status and Mapping of Wheat Yield’s within-Field Spatial Variations Using Color Images Acquired from UAV-camera System. Remote Sens. 2017, 9, 289. [Google Scholar] [CrossRef] [Green Version]
  69. Brivio, P.A.; Giardino, C.; Zilioli, E. Determination of chlorophyll concentration changes in Lake Garda using an image-based radiative transfer code for Landsat TM images. Int. J. Remote Sens. 2001, 22, 487–502. [Google Scholar] [CrossRef]
  70. Dong, T.F.; Liu, J.G.; Qian, B.D.; He, L.M.; Liu, J.; Wang, R.; Jing, Q.; Champagne, C.; McNairn, H.; Powers, J.; et al. Estimating crop biomass using leaf area index derived from Landsat 8 and Sentinel-2 data. ISPRS J. Photogramm. Remote Sens. 2020, 168, 236–250. [Google Scholar] [CrossRef]
  71. Adewopo, J.; Peter, H.; Mohammed, I.; Kamara, A.; Craufurd, P.; Vanlauwe, B. Can a Combination of UAV-Derived Vegetation Indices with Biophysical Variables Improve Yield Variability Assessment in Smallholder Farms? Agronomy 2020, 10, 1934. [Google Scholar] [CrossRef]
  72. Ding, S.; Zhang, N.; Zhang, X.; Wu, F. Twin support vector machine: Theory, algorithm and applications. Neural Comput. Appl. 2017, 28, 3119–3130. [Google Scholar] [CrossRef]
  73. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef] [Green Version]
  74. Huang, G.B.; Siew, C.K. Extreme learning machine with randomly assigned RBF kernels. Int. J. Inf. Technol. 2005, 11, 16–24. [Google Scholar]
  75. Huang, G.; Huang, G.-B.; Song, S.; You, K. Trends in extreme learning machines: A review. Neural Netw. 2015, 61, 32–48. [Google Scholar] [CrossRef]
  76. Gu, J.; Wang, Z.; Kuen, J.; Ma, L.; Shahroudy, A.; Shuai, B.; Liu, T.; Wang, X.; Wang, G.; Cai, J.; et al. Recent advances in convolutional neural networks. Pattern Recognit. 2018, 77, 354–377. [Google Scholar] [CrossRef] [Green Version]
  77. Beck, R.; Zhan, S.; Liu, H.; Tong, S.; Yang, B.; Xu, M.; Ye, Z.; Huang, Y.; Shu, S.; Wu, Q.; et al. Comparison of satellite reflectance algorithms for estimating chlorophyll-a in a temperate reservoir using coincident hyperspectral aircraft imagery and dense coincident surface observations. Remote Sens. Environ. 2016, 178, 15–30. [Google Scholar] [CrossRef] [Green Version]
  78. Cillero Castro, C.; Domínguez Gómez, J.A.; Delgado Martín, J.; Hinojo Sánchez, B.A.; Cereijo Arango, J.L.; Cheda Tuya, F.A.; Díaz-Varela, R. An UAV and Satellite Multispectral Data Approach to Monitor Water Quality in Small Reservoirs. Remote Sens. 2020, 12, 1514. [Google Scholar] [CrossRef]
  79. León-Tavares, J.; Roujean, J.-L.; Smets, B.; Wolters, E.; Toté, C.; Swinnen, E. Correction of Directional Effects in VEGETATION NDVI Time-Series. Remote Sens. 2021, 13, 1130. [Google Scholar] [CrossRef]
  80. Prangsma, G.J.; Roozekrans, J.N. Using NOAA AVHRR imagery in assessing water quality parameters. Int. J. Remote Sens. 2007, 10, 811–818. [Google Scholar] [CrossRef]
  81. Zhang, T.; Huang, M.; Wang, Z. Estimation of chlorophyll-a Concentration of lakes based on SVM algorithm and Landsat 8 OLI images. Environ. Sci. Pollut. Res. Int. 2020, 27, 14977–14990. [Google Scholar] [CrossRef]
  82. Yang, H.; Du, Y.; Zhao, H.; Chen, F. Water Quality Chl-a Inversion Based on Spatio-Temporal Fusion and Convolutional Neural Network. Remote Sens. 2022, 14, 1267. [Google Scholar] [CrossRef]
  83. Li, M.-W.; Geng, J.; Hong, W.-C.; Zhang, Y. Hybridizing Chaotic and Quantum Mechanisms and Fruit Fly Optimization Algorithm with Least Squares Support Vector Regression Model in Electric Load Forecasting. Energies 2018, 11, 2226. [Google Scholar] [CrossRef] [Green Version]
  84. Han, S.; Kim, H. On the Optimal Size of Candidate Feature Set in Random forest. Appl. Sci. 2019, 9, 898. [Google Scholar] [CrossRef] [Green Version]
  85. Bhatt, D.; Patel, C.; Talsania, H.; Patel, J.; Vaghela, R.; Pandya, S.; Modi, K.; Ghayvat, H. CNN Variants for Computer Vision: History, Architecture, Application, Challenges and Future Scope. Electronics 2021, 10, 2470. [Google Scholar] [CrossRef]
  86. Adão, T.; Hruška, J.; Pádua, L.; Bessa, J.; Peres, E.; Morais, R.; Sousa, J. Hyperspectral Imaging: A Review on UAV-Based Sensors, Data Processing and Applications for Agriculture and Forestry. Remote Sens. 2017, 9, 1110. [Google Scholar] [CrossRef]
Drones 07 00002 g001 550
Figure 1. (a) Study area overview and its location in Erhai Lake, with the used unmanned aerial vehicle in the inset, and (b) locations of the sampling points.
Figure 1. (a) Study area overview and its location in Erhai Lake, with the used unmanned aerial vehicle in the inset, and (b) locations of the sampling points.
Drones 07 00002 g001
Drones 07 00002 g002 550
Figure 2. Extraction of the reflective area of the image: (a) original visible light image and (b) reflection area extraction effect image.
Figure 2. Extraction of the reflective area of the image: (a) original visible light image and (b) reflection area extraction effect image.
Drones 07 00002 g002
Drones 07 00002 g003 550
Figure 3. Results of the linear correlation analysis (R2).
Figure 3. Results of the linear correlation analysis (R2).
Drones 07 00002 g003
Drones 07 00002 g004 550
Figure 4. Test results of multiple regression models of the vegetation indices: (a) SVM model, (b) RF model, (c) ELM model, and (d) CNN model.
Figure 4. Test results of multiple regression models of the vegetation indices: (a) SVM model, (b) RF model, (c) ELM model, and (d) CNN model.
Drones 07 00002 g004
Table
Table 1. DJI P4M UAV parameters.
Table 1. DJI P4M UAV parameters.
AircraftTypeQuadcopter
Position accuracyVertical: ± 0.5 m; horizontal: ± 1.5 m
Take-off weight1484 g
Time of flightAbout 27 min
Wind loading rating10 m/s
Image position compensationRelative positions are compensated in photo EXIF coordinates
Image sensorVisible light channelPixels: 2 million 80 thousand
Multispectral sensor: all sensors have 2 megapixelsG channelSpectral band: 450 ± 16 nm
B channelSpectral band: 560 ± 16 nm
R channelSpectral band: 650 ± 16 nm
RE channelSpectral band: 730 ± 16 nm
NIR channelSpectral band: 840 ± 26 nm
Data from DJI website; http://www.dji.com/cn/p4-multispectral/specs (accessed on 30 September 2022).
Table
Table 2. Vegetation indices and their formulas and value ranges.
Table 2. Vegetation indices and their formulas and value ranges.
Vegetation IndexFormulaValue Range
Normalized difference vegetation index [49] N D V I = N I R − R / N I R + R [−1,1]
Normalized difference vegetation structure index [53] N D V S I = N I R − R + G / 2 / N I R + R + G / 2 [−1,1]
Normalized green–blue difference index [54] N G B D I = G − B / G + B [−1,1]
Normalized green–red difference index [54] N G R D I = G − R / G + R [−1,1]
Ratio normalized difference vegetation index [55] R N D V I = N I R − R / N I R + R ∗ N I R / R (−∞, +∞)
Ratio vegetation index [56] R V I = N I R / R (0, +∞)
Red–green–blue vegetation index [57] R G B V I = G 2 − R ∗ B / G 2 + R ∗ B [−1,1]
Red-edge chlorophyll index [58] C I R E D E D G E = N I R / R E − 1 [−1, +∞)
Red–green ratio index [54] R G R I = R / G (0, +∞)
Visible atmospherically resistant index [59] V A R I C R E E N = G − R / G + R − B (−∞, +∞)
Visible band difference vegetation index [54] V D V I = 2 ∗ G − R − B / 2 ∗ G + R + B [−1,1]
Vegetative [60] V E G = G / R a ∗ B 1 − a a = 0.667 (−∞, +∞)
Blue–green ratio index [61] B G R I = B / G (−∞, +∞)
ENGBDI [61] E N G B D I = G 2 − B 2 / G 2 + B 2 [−1,1]
Two-band enhanced vegetation index [62] E V I 2 = 2.5 ∗ N I R − R / N I R + 2.4 ∗ R + 1 [−0.71,1.25]
Blue normalized vegetation index [63] B N D V I = N I R − B / N I R + B [−1,1]
Color index of vegetation extraction [64] C I V E = 0.441 ∗ R − 0.881 ∗ G + 0.385 ∗ B + 18.787 [17.906,19.586]
Difference vegetation index [65] D V I = N I R − R [−1,1]
Enhanced normalized difference vegetation index [66] E N D V I = N I R + G − 2 ∗ B / N I R + G + 2 ∗ B [−1,1]
Excess green index [54] E X G = 2 ∗ G − R − B [−2,2]
Excess green minus excess red [54] E X G R = E X G − 1.4 ∗ R − G [−4.4,2]
Green chlorophyll index [58] C I G R E E N = N I R / G − 1 [−1, +∞)
Green normalized difference vegetation index [67] G N D V I = N I R − G / N I R + G [−1,1]
Green–red ratio index [68] G R R I = G / R (0, +∞)
KIVU [69] K I V U = B − R / G (−∞, +∞)
Modified simple ratio red edge [70] M S R R E = N I R / R E − 1 / N I R / R E + 1 (−1,1)
Modified single ratio [67] M S R = N I R / R − 1 / N I R / R + 1 (−1,1)
Normalized difference red-edge index [71] N D R E = N I R − R E / N I R + R E [−1,1]
B, blue; G, green; NIR, near-infrared; R, red; RE, red edge. Data from the IDB website; https://www.indexdatabase.de/ (accessed on 25 August 2022).
Table
Table 3. Descriptive statistics of the whole dataset, training dataset, and test dataset.
Table 3. Descriptive statistics of the whole dataset, training dataset, and test dataset.
DatasetSample SizeMaximum (μg/L)Minimum (μg/L)Median
(μg/L)		Mean (μg/L)Standard Deviation
(μg/L)
Whole dataset15584.748.0319.2624.8915.12
Training dataset12081.328.0318.6723.6313.68
Test dataset3584.7411.7523.6229.2318.65
Table
Table 4. Coefficient of determination (R2) values of vegetation index univariate regression.
Table 4. Coefficient of determination (R2) values of vegetation index univariate regression.
Vegetation IndexSVRRFELMCNN
NDVI0.4649 **0.5760 **0.4783 **0.4246 **
NDVSI0.3799 *0.4865 **0.5660 **0.3701 *
NGBDI0.05710.3313 *0.2453 *0.0612
NGRDI0.00050.4952 **0.2711 *0.0002
RNDVI0.6063 **0.5762 **0.5742 **0.6111 **
RVI0.5863 **0.5746 **0.4573 **0.5080 **
RGBVI0.03540.03120.01230.0348
CIREDEDGE0.04680.00620.00880.0408
RGRI0.00380.4723 **0.18630.0013
VARIGREEN0.01420.15360.2286 *0.0008
VDVI0.03560.00730.00290.0355
VEG0.01720.03960.14500.0161
BGRI0.05770.3520 *0.4116 *0.0732
ENGBDI0.04900.3525 *0.3511 *0.0625
EVI20.4439 **0.5758 **0.5200 **0.4395 *
BNDVI0.2685 **0.2907 *0.2730 *0.2663 *
CVIE0.08420.00860.00030.0842
DVI0.4282 *0.6057 **0.6923 **0.4249 *
ENDVI0.12990.14610.10710.1297
EXG0.06820.00010.00010.0682
EXGR0.5798 **0.6030 **0.6794 **0.5800 **
CIGREEN0.3376 *0.3194 *0.6213 **0.3516 *
GNDVI0.2982 *0.3310 *0.6235 **0.3128 *
GRRI0.00300.4883 **0.18440.0001
KIVU0.04700.05060.2461 *0.0465
MSRRE0.04350.00580.00480.0418
MSR0.4649 **0.5832 **0.4835 **0.4246 **
NDRE0.04350.00520.00480.0418
Levels of significance: * p-value < 0.05; ** p < 0.01.
Table
Table 5. Original band reflectance regression results.
Table 5. Original band reflectance regression results.
AlgorithmR2RMSE (μg/L)MRE
SVM0.6753 **9.05350.2907
RF0.6426 **10.07720.3084
ELM0.7202 **8.30230.2728
CNN0.6598 **9.69630.3040
Level of significance: ** p < 0.01.
Table
Table 6. Results of the multiple regression of the vegetation indices.
Table 6. Results of the multiple regression of the vegetation indices.
AlgorithmR2RMSE (μg/L)MRE
SVM0.5800 **12.45700.2894
RF0.7086 **10.37130.2891
ELM0.7810 **9.50500.2576
CNN0.7917 **8.76600.2461
Level of significance: ** p < 0.01.
Table
Table 7. Multivariate regression results of the vegetation index and original band information.
Table 7. Multivariate regression results of the vegetation index and original band information.
AlgorithmR2RMSE (μg/L)MRE
SVM0.5489 **12.87150.2963
RF0.7065 **10.33940.2904
ELM0.7299 **9.91850.2751
CNN0.7710 **9.13170.2485
Levels of significance: ** p < 0.01.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Zhao, X.; Li, Y.; Chen, Y.; Qiao, X.; Qian, W. Water Chlorophyll a Estimation Using UAV-Based Multispectral Data and Machine Learning. Drones 2023, 7, 2. https://doi.org/10.3390/drones7010002

AMA Style

Zhao X, Li Y, Chen Y, Qiao X, Qian W. Water Chlorophyll a Estimation Using UAV-Based Multispectral Data and Machine Learning. Drones. 2023; 7(1):2. https://doi.org/10.3390/drones7010002

Chicago/Turabian Style

Zhao, Xiyong, Yanzhou Li, Yongli Chen, Xi Qiao, and Wanqiang Qian. 2023. "Water Chlorophyll a Estimation Using UAV-Based Multispectral Data and Machine Learning" Drones 7, no. 1: 2. https://doi.org/10.3390/drones7010002

Article Metrics

Back to TopTop