Prediction of lake chlorophyll concentration using the BP neural network and Sentinel-2 images based on time features

Lakes provide a variety of services, including aquaculture, species protection, transportation, agricultural irrigation and water resource storage, in addition to being vital freshwater resources (Wang et al. 2021a; Quan et al. 2022). Maintaining the health of lakes is critical for both productivity and survival (Longo et al. 2019). However, because of the lake's fast socio-economic growth, a huge volume of home and industrial effluent is dumped into it. The levels of nutrients such as nitrogen and phosphorus in the water bodies are well over the recommended levels (Li et al. 2021). This results in the eutrophication of lakes, which will lead to the abnormal proliferation of aquatic organisms such as algae and loss of water body functions, and then a large number of aquatic animals and plants will die (Zhang et al. 2019; Quan et al. 2020; Dong et al. 2021).

Lake eutrophication is a worldwide biological and environmental issue. Chl-a concentration in water bodies reflects phytoplankton dispersion and is an indicator of algal biomass (Guo et al. 2021a, 2021b) as well as an important indication of eutrophication status (Bramich et al. 2021; Chen et al. 2021). Physical sampling and chemical analysis are the most common approaches for monitoring Chl-a levels in lakes, although they do not offer a full picture of the water. Traditional methods are time-consuming and labor-intensive, with drawbacks such as discomfort, a long monitoring period and expensive costs, and they can only provide extremely restricted and discrete data. Due to the wide range and timeliness of satellite monitoring (Yang et al. 2022) and the low cost of image acquisition (Zhao et al. 2020, 2021), satellite remote sensing has become an efficient tool for monitoring Chl-a concentration in lakes (Mohebzadeh & Lee 2021). This is essential to prevent and control the eutrophication of water bodies. By analyzing the correlation between satellite band data and Chl-a concentration, we can find the bands or bands combination highly related to Chl-a concentration, and then a model of satellite images and Chl-a concentration can be created to predict Chl-a concentration.

Traditional Chl-a concentration prediction methods rely primarily on a multivariate statistical regression model between remote sensing images and measured Chl-a concentrations (Mishra & Mishra 2012; Zhang et al. 2015; Bi et al. 2018; Xu et al. 2020). Nowadays, machine learning is more used to build models. He et al. (2021) used the double hidden layer ANN (artificial neural network) model to intelligently predict the Chl-a concentration, which to some extent shows that deep learning has more advantages in extracting the main features of information than shallow learning. Pan et al. (2021) quantitatively inverted the Chl-a concentrations in Taihu Lake using a band ratio model and a three-band model. Yue et al. (2020) performed a quadratic polynomial regression of Chl-a concentration in summer with B5/B4 as the independent variable and obtained the best results, corresponding to the decision coefficient (R²) of 0.816. Martinez et al. (2020) used SVR (support vector regression) to reconstruct and study the annual variation of Chl-a in the Pacific and Indian Oceans. Shin et al. (2020) combined the RNN (Recurrent Neural Network) model with a rolling window learning method to obtain the best prediction of Chl-a concentration. Zhang et al. (2020) used SVM (support vector machine) to obtain the best combination of bands to invert Chl-a, corresponding to the R² = 0.774. Su et al. (2021) used the machine learning model LightGBM to obtain the best results for inverse Chl-a with the corresponding R² = 0.785. Park et al. (2020) used the RF model to reconstruct the missing values of Chl-a concentration data. Saberioon et al. (2020) used Sentinel-2A and machine learning algorithms to monitor Chl-a (R² = 0.85, RMSE (Root Mean Square Error) = 48.572) and TSS (R² = 0.80, RMSE = 19.55) in small inland waters. However, the reflection spectra of inland lake waters are usually complex (Mishra et al. 2009), and atmospheric variations in the Earth can interfere with the output values of the models (Feng 2021). This can lead to poor inversion accuracy of statistical and linear models for complex nonlinear problems (Zhang et al. 2009). The relationship between water quality parameters and the optical or physical properties of the water body is typically nonlinear, and neural networks are suitable for explaining this relationship (Hou et al. 2021).

In this study, based on the requirements of the project, Chl-a was measured monthly, and remote sensing images could also be obtained monthly. The 12 months of each year belong to four seasons. The impact of seasons on the lake environment is significant, so the monthly features of the sampled data should be very obvious. Additionally, the number of days between the image and the measurement time (day difference) could also be regarded as a feature. Therefore, the month features (Ms) and day difference features (Ds) would be used together with the remote sensing data as the input of the model to predict the Chl-a concentration. The purpose of this study was to verify the performance of time features in improving the prediction accuracy of the model, and establish a neural network model to predict the concentration of Chl-a in the Wuliangsuhai Lake.

The objectives of this study were (1) to compare and analyze the influence of different time features on the accuracy of neural network models, (2) to develop a neural network model with remote sensing images combined with time features as input and Chl-a concentration as output, (3) validate the accuracy of the neural network model and (4) use the established model to construct a distribution map of Chl-a concentration in the Wuliangsuhai Lake for a certain period. Through this work, the eutrophication status of the Wuliangsuhai Lake can be quickly obtained, which is of great significance for both water resources’ environmental monitoring and management.

The measured Chl-a data were from 19 sampling points in the Wuliangsuhai Lake (Figure 1(c)). The sample points were evenly distributed and located in the center of the lake. A handheld GPS device was used to record the coordinates. We used a collector to collect water within 0.5 m depth. The water samples were then taken back to the laboratory for Chl-a concentration extraction by Ultraviolet spectrophotometer. The measured Chl-a data in the study area were collected once a month from September 2015 to July 2018. During this period, 108 satellite images with measured data close to its acquisition time (within 2 days) were also selected as sample data to establish the model.

The remote sensing images are the Sentinel-2 satellite images released by the European Space Agency (ESA) (Tian et al. 2021a, 2021b). The images were selected from 2015 to 2018 when the weather was clear with low cloudiness and in a non-ice period. Sentinel-2 images are from two satellites, satellite A and satellite B (Yin et al. 2022). The satellite A was launched on June 23, 2015, and satellite B was launched on March 7, 2017. Sentinel-2 carries a multispectral imager (MSI), with an altitude of 786 km, can cover 13 spectral bands (Table 1), a width of 290 km and a ground resolution of up to 10 m. The revisit period of one satellite is 10 days, and the two satellites complement each other, and the revisit period is 5 days. From visible and near-infrared to short wave infrared, compared with other satellites, it has great advantages in revisit cycle and resolution (Sun et al. 2021; Mao et al. 2022). The Sentinel-2 images comes from the USGS official website (https://earthexplorer.usgs.gov) (Tian et al. 2019). In this study, sen2cor-02.08.00 is used to preprocess Sentinel-2 images, such as radiometric calibration and atmospheric correction.

Table 1

Sentinel-2 image bands introduction

Band .	Name .	Central wavelength (μm) .	Resolution (m) .
	Coastal aerosol	0.433	60
	Blue	0.490	10
	Green	0.560	10
	Red	0.665	10
	Vegetation Red Edge	0.705	20
	Vegetation Red Edge	0.740	20
	Vegetation Red Edge	0.783	20
	NIR	0.842	10
	Vegetation Red Edge	0.865	20
	Water vapor	0.945	60
	SWIR-Cirrus	1.375	60
	SWIR1	1.610	20
	SWIR2	2.190	20

Band .	Name .	Central wavelength (μm) .	Resolution (m) .
	Coastal aerosol	0.433	60
	Blue	0.490	10
	Green	0.560	10
	Red	0.665	10
	Vegetation Red Edge	0.705	20
	Vegetation Red Edge	0.740	20
	Vegetation Red Edge	0.783	20
	NIR	0.842	10
	Vegetation Red Edge	0.865	20
	Water vapor	0.945	60
	SWIR-Cirrus	1.375	60
	SWIR1	1.610	20
	SWIR2	2.190	20

View Large

When using the BP neural network to create a model, it is not necessary to specify the mathematical relationship between the model input and output (Ahmadi et al. 2020, 2021). By constantly adjusting (training) the connection weight of neurons in the network, the expected output value corresponding to the input can be obtained. The BP neural network model uses gradient descent method to adjust the connection weight, so as to minimize the mean square error between the model output value and the expected output value. The BP neural network model is mainly composed of input layer, hidden layers and output layer. Besides simple and relatively easy to achieve, its algorithm has strong self-learning, self-organizing and adaptive capabilities (Ahmadi et al. 2021).

Theoretically, it can simulate any complex nonlinear relationship. Therefore, the BP neural network model was chosen to predict the Chl-a concentration in the study area.

Sentinel-2 data originally had 13 bands. After preprocessing, disappeared, leaving only 12 bands. At present, there are no more scientific method to directly select the band combinations with high correlation with chlorophyll concentration. Therefore, this study selects all bands (aBs), bands with positive correlation (pBs) and bands with negative correlation (nBs) with Chl-a concentrations as the inputs of neural network, and the measured Chl-a concentration as the output for the construction and training of the BP neural network.

The correlation coefficient r_x between each band and Chl-a concentration is shown in Table 2. According to Table 2, the positive correlation bands are B05, B06, B07, B08, B8A and B09, and the negative correlation bands are B01, B02, B03, B04, B11 and B12.

The value of R² ranges from [0,1], and the closer to 1, the higher the prediction accuracy of the model. Root mean square error (RMSE) is used to illustrate the dispersion of the sample. For nonlinear fitting, the smaller the RMSE, the better.

It can be seen from Figure 2 that when the number of neurons reaches 8, even if the number of neurons continues to increase, the accuracy does not improve significantly, and R² was below 0.8. In fact, both satellite images and measured data have periodicity in sampling time, and the data itself has the characteristics of periodicity. We tried to add the month features on the basis of input bands, so that the bands and month features could be used as the input together, and explored whether the month features were helpful to improve the accuracy of the model. The month information here is represented by the month switch vector. For example, if the input sample comes from one of the six months, it is represented by a vector. The vector is composed of 0 and 1, with a total of six elements, representing 6 months in turn. If the current sample belongs to a month, the position of the corresponding element in the vector is set to 1 and the other positions are set to 0. For example, if the sample data come from May, June, July, August and September, and the current sample is the sample of August, the vector representing the month feature is (0 ,0, 0, 1, 0).

As mentioned earlier, there may be day differences between satellite images and measured values of chlorophyll, and a vector of day difference is also defined. If the day difference is within 0, 1 and 2 days, the elements in the vector can only be −2, −1, 0, 1 and 2. The reason why there are positive and negative numbers is that the day difference can be earlier or later. The definition rule of the day difference vector is: if the difference between the image and the measured data in the current sample is D days, the corresponding element position is set to D, and the other positions are set to 0. For example, if the current image is 1 day earlier than the actual measurement, the vector representing the day difference feature is (0, − 1,0,0,0). Next, combine the two time features with aBs, pBs and nBs, respectively, and use the combined features as the input of the neural network to test the impact of time features on the prediction accuracy.

Firstly, the prediction results of the model with all bands, positive correlation bands and negative correlation bands as input were compared. Then, the month features and the day differences were combined in each band combination separately to test the effects of these two features on the model accuracy. Finally, month features and day differences were added to each band combination at the same time to test the effect of both features on the model accuracy. The best form of model input could be obtained by comparing the effects of various inputs on model accuracy. At this point the optimal model was also achieved. After obtaining the optimal prediction model, the model was used to predict the Chl-a concentrations of the Wuliangsuhai Lake in a certain period. The prediction results were compared with the existing research conclusions to verify the practicability of the model.

Fortunately, when aBs, pBs and nBs were combined with Ms or Ds as input, the prediction accuracy could be greatly improved. Moreover, the accuracy of the joint Ms was improved more. This is because month, as a periodic time indicator, can have a great impact on the prediction accuracy. At this time, the model accuracy was aBs + Ms > nBs + Ms > pBs + Ms It can be found that adding the month feature does not change the previous accuracy ranking. Furthermore, after adding the month feature, the improvement rate of each previous band combination was also different, and the improvement rate of pBs was the largest. The accuracy of pBs + Ms was almost twice that of pBs. After aBs, pBs and nBs were respectively combined with the month feature as input, R² could all reach above 0.8. The R² of aBs + Ms as model input even exceeds 0.9, which was very good for practical application.

For aBs, pBs and nBs, if Ms and Ds were combined as model inputs at the same time, the model accuracy obtained was higher than that obtained by combining Ds only. However, it could not exceed the model accuracy obtained only by combining Ms, which shows that Ds has not played a very positive role in improving the model accuracy. Consequently, the above results showed that Ms, as an inherent time feature of remote sensing data, could fundamentally and directly improve the prediction accuracy of the model.

After the model predicted 29,695 coordinate points in the August 2017 image that formed the clear water area, it was found that the mean chlorophyll concentration of 8,191 coordinate points in the north area was 25.06 μg/L, the mean chlorophyll concentration of 7,660 coordinate points in the middle area was 21.59 μg/L, and the mean chlorophyll concentration of 13,844 coordinate points in the south area was 11.97 μg/L. There are external water sources in the north of the lake, mainly including Yellow River, domestic sewage and residual water after farmland irrigation. The water after farmland irrigation contains a large number of chemical fertilizers, pesticides and other pollutants, so that the concentration of nitrogen and phosphorus in the water body in this area is high and the degree of eutrophication is serious. The lake as a whole includes reed area, plant area and water area. The water quality in the southern part of the lake is gradually improved because only a small amount of external water is discharged. Therefore, most of the time, the chlorophyll concentration decreases from north to south.

In this study, all bands of Sentinel-2 satellite images, the bands with positive and negative correlation with the measured Chl-a concentration were used as inputs, and the neural network was constructed to predict the Chl-a concentration in the Wuliangsuhai Lake. After a lot of experiments, a neural network model with 12 bands of Sentinel-2 image combined with month feature as the input, one hidden layer, eight neurons and Chl-a concentration as outputs was determined. The R² could reach 0.929 when the test set accounts for 20%. The improvement of model accuracy by month feature was greater than that by day difference feature. The model test on the satellite images of the Wuliangsuhai Lake in some specific time periods demonstrated that the model developed in this work was suitable for prediction of Chl-a concentration, and had high accuracy and practicality. The model has considerable theoretical and practical relevance for directing lake ecosystem management and pollution prevention and control, and it has essential reference value for intelligent monitoring of lake water environment. At present, only three-band combinations are used as input in this research. Whether there are other band combinations that are more helpful to improve the prediction accuracy of the model remains to be explored with the help of scientific methods. Moreover, due to weather conditions, sampling time, satellite transit frequency and other reasons, the data available for model training are relatively limited, especially the day difference between images and samplings, which may lead to the inaccuracy of the model. This is bound to have a certain degree of negative impact on the generalizability of model application. In the future study, we are ready to try some statistical methods for bands selection, and look forward to obtaining higher model accuracy on larger data sets.

This research was supported by the Research Program of Science and Technology at the Universities of Inner Mongolia Autonomous Region (NJZZ23044), the National Natural Science Foundation of China (61962047), the National Key Research and Development Program of China (2019YFC0409205) and the Natural Science Foundation of Inner Mongolia Autonomous Region of China (2019MS06015, 2020MS06011, 2021MS06009).

All data are available from the corresponding author.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.