Using hyperspectral data collected in January and June 2022 from the Sha River, the concentrations of total nitrogen (TN) and total phosphorus (TP) were estimated using the differential method. The results indicate that the optimal bands for estimation vary monthly due to temperature fluctuations. In the TN model, the power function model at 586 nm in January exhibited the strongest fit, yielding a fit coefficient (R^{2}) of 0.95 and F-value of 164.57 at a significance level (^{2} = 0.93 and F = 80.95 at ^{2} = 0.78 and F = 20.61. However, the overall fit in June outperformed that in January. Specifically, the quadratic and linear model fits of the differential values at 824 and 863 nm with TP achieved R^{2} = 0.96 and F-values of 34.42 and 203.34, respectively.

Total Nitrogen (TN) refers to the nitrogen content in water, encompassing ammonia nitrogen, nitrate nitrogen, nitrite nitrogen, and organic nitrogen [

Currently, in quantitative research on water quality indicators, three main categories prevail: empirical, semi-empirical, and analytical methods [

The Sha River, an important water system in Chengdu, China, was chosen as the research subject in this paper. The author collected water samples at various locations in the Sha River during January and June 2022 and analyzed the water’s spectrum using an ASD ground objective spectrometer. The collected water samples were sent to the laboratory to determine the concentrations of total nitrogen and total phosphorus. This study investigates the correlation between total total nitrogen, total phosphorus, and spectral data, with the intention of identifying the optimal band for quantitatively estimating total nitrogen and total nitrogen concentration. An estimation model is subsequently developed to establish a scientific foundation. The use of spectral sensors offers a novel approach for estimating water quality parameters on a large scale in the Sha River, Chengdu.

To validate the model’s applicability, Landsat9 remote sensing data was chosen for analysis. The Landsat9 satellite is considered superior to many other satellites because of its high spatial resolution and data availability. It offers a relatively comprehensive range of bands for water quality inversion, as indicated in

Bands | Wavelength range (nm) | Resolution (m) |
---|---|---|

B1 | 433~453 | 30 |

B2 | 450~515 | 30 |

B3 | 525~600 | 30 |

B4 | 630~680 | 30 |

B5 | 845~885 | 30 |

B6 | 1560~1660 | 30 |

B7 | 2100~2300 | 30 |

B8 | 500~680 | 15 |

B9 | 1360~1390 | 30 |

B10 | 1060~1119 | 100 |

B11 | 1150~1251 | 100 |

This Landsat data was downloaded from the United States Geological Survey (USGS,

The Sha River is one of the three major urban rivers in Chengdu. It is part of the Minjiang River system (

The water samples used for this study were collected on January 05 and June 23, 2022.

Hyperspectral data were measured by Analytical Spectroscopy Device (ASD) FieldSpec4 produced by ASD Corporation (USA). The instrument parameters are as follows: the measurement spectrum range is 350–2500 nm, the wavelength accuracy is ±1 nm, the sampling time is 10 seconds/time, and the front field of view angle is 25°. For each measurement, the upward radiance (

Using the spectral data processing software ViewSpecPro to calculate the average value of the spectral curve of the sample. The spectral data is converted to ASCII format and then stored for further processing.

The determination of water quality parameters primarily involves direct measurements at the sampling site and subsequent physical and chemical analysis conducted in the laboratory. The field data collected for this study primarily includes measurements of water transparency as well as the latitude and longitude coordinates of the sampling points. The laboratory data primarily consists of the samples collected and sent to physical and chemical laboratories for analysis after being refrigerated. According to the specified indicators, the sampler should collect water samples from a depth of 50 cm below the water surface and distribute them equally into eight 500 ml brown glass bottles. At the same time, add H_{2}SO_{4} to the brown bottles to keep the pH below 2. It was sent to the Sichuan Xiye Testing Laboratory for determination. The detection environment conditions were set at a temperature of 22°C and relative humidity of 55%. Determination of TN and TP followed the “Surface Water Environmental Quality Standard” (GB 11893-89), using alkaline potassium persulfate digestion ultraviolet spectrophotometry (GB 11894-89) and potassium persulfate digestion ammonium molybdate spectrophotometry (GB 11893-89) as the standard analysis methods.

Because of variations in measurement time and location, the spectrum exhibits variability. One approach to mitigate the impact of temporal and spatial variations on determining the absolute reflectance value is through the use of a spectrum normalization method. This method effectively reduces the influence of time and location differences on the determination of the absolute reflectance value. Normalized reflectance is calculated by dividing the reflectance at each wavelength by the average reflectance value across the entire wavelength range [

In the formula, the normalized water body remote sensing reflectance is the original calculated water body remote sensing reflectance, and n is the number of bands between 420 and 900 nm.

Spectral differential technology allows for identifying shape changes in spectral data [

First-order differential processing has the capability to remove linear or nearly linear backgrounds, mitigate the impact of noise spectrum on the target spectrum (which must be nonlinear), and minimize the influence of water on the reflection spectrum of water bodies [

Half of the data are kept for this paper’s accuracy verification. Accuracy was assessed using the Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Mean Absolute Error (MAE) [

In the formula,

This paper aims to analyze the seasonal fluctuations of the spectral reflection curve of the water body in the Sha River by processing the spectral data curve using ENVI. Additionally, we describe the shape characteristics of the absorption bands in the spectrum.

From

From a spectral perspective, the primary distinction between the winter and summer curves occurs at approximately 800 nm. During summer, a prominent reflection peak is observed at 800 nm, whereas the winter curve exhibits a more gradual change. The lower winter temperature hinders the formation of algae in water, leading to a decrease in chlorophyll and carotene concentration.

During summer, the reflectance of water in the 400–600 nm region of the spectral curve is reduced due to the light energy absorption by chlorophyll-a and yellow substances. Additionally, chlorophyll and carotene have a weak light energy absorption capability, further compounded by the scattering effect of cells. A prominent reflection peak is observed at 800 nm, serving as a quantitative benchmark for chlorophyll concentration. The absorption of phycocyanin in the 620–630 nm range induces a dip in the reflection spectrum. In addition, the strong absorption of chlorophyll-a around 670 nm results in additional valleys in the reflection spectrum [

The objective of this study is to establish the relationship between the measured TN-TP concentration and the spectrum in order to analyze the seasonal impact on the water quality of the Sha River. The construction of the water quality inversion model typically involves analyzing the spectral characteristics of the water body to identify bands with strong responses. These bands are then used in correlation analysis with the concentration of the water quality parameter to establish the inversion model based on the principle of the largest correlation coefficient. This study examines the correlation between TN-TP, normalized remote sensing reflectance data, and multi-order differential algorithm spectrum in different seasons using Pearson correlation analysis [

Correlation analysis was conducted between the TN-TP concentration and the normalized reflectance data of each point. The maximum positive and negative correlation coefficients in January and June were obtained through correlation analysis (

Based on the correlation analysis conducted above, the differential single-band with a relatively high correlation was chosen as the inversion factor for TN-TP. The fitting was performed using SPSS software with the quadratic polynomial, exponential, power function, and linear function. After evaluating the fitting effect of the TN model using data collected in January, it was determined that the power function at 586 nm exhibited the highest performance, with R^{2} = 0.95 and F = 164.57. In contrast, during June, the exponential model at 477 nm yielded the most satisfactory fitting effect, with a fitting coefficient R^{2} = 0.93 and F = 80.95 at a significance level of

During the fitting model in January, the differential value of 851 nm with a 1 nm interval demonstrated the best fitting effect for TP, with an R^{2} value of 0.78. However, the fitting effect in June surpassed that of January, reaching an R^{2} value of 0.96 for TP.

Utilize the remaining four sample points to test, the RMSE, MAPE, and MAE were calculated (^{2} and F values of the regression model and RMSE, MAPE, MAE between the predicted and measured values, the best model is determined.

Month | Bands | Model type | Model | R^{2} |
F | MAPE | RMSE | MAE | |
---|---|---|---|---|---|---|---|---|---|

TN | January | 586 nm | Power function | y = 0.0976 * (1 + x)^(9.52E + 05) | 0.95 | 164.57 | 0.06 | 0.008 | 0.007 |

922 nm | Power function | y = 0.20857 * (1 + x)^(1.25E + 06) | 0.94 | 163.12 | 0.17 | 0.023 | 0.020 | ||

857 nm | Index | y = 0.03791 * exp(x/(2.83E-07)) + 0.12931 | 0.94 | 96.62 | 0.27 | 0.036 | 0.034 | ||

June | 653 nm | Linear | y = −2818.01469x + 0.25271 | 0.84 | 16.31 | 0.44 | 0.061 | 0.053 | |

477 nm | Index | y = 0.01574 * exp(−x/(3.69E-05)) + 0.14411 | 0.93 | 80.95 | 0.24 | 0.03 | 0.028 | ||

620 nm | Linear | y = 0.21149–4040.46562 * x | 0.78 | 11.62 | 0.79 | 0.08 | 0.074 | ||

TP | January | 418 nm | Unary quadratic | y = 1.09E-01 + (5.74E-03) * x^1 + (1.14E-04) * x^2 | 0.72 | 1.27 | 0.63 | 0.02 | 0.02 |

359 nm | Power function | y = 0.03245 * (1 + x)^(−1.50E + 06) | 0.76 | 45.67 | 0.65 | 0.01 | 0.01 | ||

851 nm | Index | y = 0.02566 * exp(−x/(7.46E-08)) + 0.00473 | 0.78 | 25.61 | 0.51 | 0.02 | 0.03 | ||

June | 785 nm | Linear | y = 0.05601 + 399.52885 * x | 0.89 | 24.06 | 0.11 | 0.006 | 0.01 | |

824 nm | Unary quadratic | y = 0.05357 + 1286.41757 * x^1 + (−1.09E+08) * x^2 | 0.96 | 34.42 | 0.16 | 0.010 | 0.01 | ||

863 nm | Power function | y = 0.07687 * (1 + x)^ −4863.75412 | 0.96 | 203.34 | 0.14 | 0.012 | 0.05 |

The best models are applied to Landsat9 remote sensing satellite images, and the Spatial Model module in ArcMAP was used to build a spatial model of TN-TP inversion, and the spatial distribution of TN-TP concentration in part of the Sha River basin was obtained after image by image calculation (

The best model was used to invert the spatial distribution of TN and TP by combining 2/3 of the measured data with the reflectance data of Landsat images, in order to verify the model results.

At present, the inversion mechanism of total nitrogen and total phosphorus based on hyperspectral data remains unclear, hence the reliance on empirical models for inversion. The inversion effect of total phosphorus is superior to that of total nitrogen. This is primarily due to the susceptibility of the 530 nm nitrogen-sensitive band to interference from non-pigmented substances and xanthin in water, and the influence of pure water signals on the 693 nm band. In contrast, the sensitive band of total phosphorus is relatively less affected compared to total nitrogen. Moreover, Researchers have observed covariation during their study [

(1) The Sha River in Chengdu is very different. Winter is the dry season of the Sha River, with little precipitation, small river runoff, and shallow river water; the water level of the Sha River rises in summer and the precipitation increases. In terms of total nitrogen and total phosphorus content, the difference between the two is not obvious, but the corresponding spectral curve difference is more obvious. The difference between the two is most obvious at 800 nm, and after correlation analysis it was concluded that the inversion model based on this range is better.

(2) The differential method is highly sensitive to the concentration of total nitrogen and total phosphorus, enabling effective extraction of weak signals. The inversion model, generated using the differential method and multiple linear regression, exhibits relatively high prediction accuracy. For the total nitrogen model, the power function model at 586 nm demonstrates the best fitting effect, with a fitting coefficient of R^{2} = 0.95 and F = 164.57 at a significance level of ^{2} = 0.93 and F = 80.95 at the significance level of ^{2} value of 0.78 and F = 20.61 at the significance level of ^{2} values of 0.96, with respective F-values of 34.42 and 203.34. Furthermore, the relative error between the predicted and sample points was within 20%, indicating the model’s excellent prediction accuracy and stability.

The hyperspectral inversion of total nitrogen and total phosphorus is still in the exploratory stage and requires further investigation into molecular spectroscopy and optical mechanisms to improve the inversion model. This research is in its early stages, and the current collection of experimental data is insufficient. In the future, the scope of study will be expanded both temporally and spatially, and the data set will be enlarged for more comprehensive exploration.

Interpretation 1

Interpretation 2

Porosity

Skin factor

We are very grateful to the USGS EarthExplore for providing the Landsat9. In addition, we thank the anonymous reviewers whose comments improved the manu-script.

This study is funded by the Gansu Provincial Science and Technology Plan Project, Innovation Base and Talent Project (Project Number: 21JR1RA284).

The authors confirm contribution to the paper as follows: study conception and design: Na Guo; data collection: Yaping Luo, Xinchen Wang, and Jie Wu; analysis and interpretation of results: Yaping Luo, Na Guo; draft manuscript preparation: Shuming Peng, Liu Dong. All authors reviewed the results and approved the final version of the manuscript.

This Landsat data was downloaded from the United States Geological Survey (USGS,

The authors declare that they have no conflicts of interest to report regarding the present study.