Inversion Method for Chlorophyll-a Concentration in High-Salinity Water Based on Hyperspectral Remote Sensing Data

Nan Wang, Zhiguo Wang, Pingping Huang, Yongguang Zhai, Xiangli Yang, Jianyu Su2024 · DOI

TL;DR

This paper introduces a novel machine learning approach to accurately estimate chlorophyll-a concentration in hypersaline lakes, a challenging environment for remote sensing. By incorporating salinity as a key predictive factor into a random forest model, the researchers achieved a significant improvement in accuracy, demonstrating the critical role of salinity in understanding water quality in these unique aquatic ecosystems.

What Problem Does This Paper Address?

Monitoring water quality in lakes is crucial for understanding ecosystem health and managing resources. Chlorophyll-a concentration is a key indicator of eutrophication and algal biomass. However, traditional methods are inefficient and resource-intensive. Remote sensing offers a scalable solution, but accurately estimating chlorophyll-a in hypersaline lakes presents a significant challenge. High salinity can alter the optical properties of water, affecting the spectral signatures that remote sensing relies on. Existing models often struggle to account for this, leading to inaccurate estimations. This paper aims to bridge this gap by developing a robust method for chlorophyll-a inversion in hypersaline environments, specifically focusing on Lake Daihai.

Methods

The study employed a multi-stage approach combining remote sensing data preprocessing, feature selection, and machine learning model development.

  • Remote Sensing Data Preprocessing: Hyperspectral imagery from the Zhuhai-1 satellite underwent standard preprocessing steps, including radiometric calibration and atmospheric correction, to obtain accurate surface reflectance values. Orthorectification was also performed to align the imagery with ground truth locations.
  • Feature Selection: Principal Component Analysis (PCA) was utilized to reduce the dimensionality of the hyperspectral data and identify the most informative spectral bands for chlorophyll-a inversion. This step is crucial for handling the high-dimensional nature of hyperspectral data.
  • Machine Learning Models: Three machine learning models were initially compared:
  • Random Forest: An ensemble learning method known for its robustness and ability to handle high-dimensional data.
  • Model Stacking (Ensemble Learning): A technique that combines multiple base models to improve prediction accuracy.
  • Ridge Regression: An enhanced linear regression model that addresses multicollinearity.
  • Salinity Integration: The core innovation of this study was the incorporation of salinity data as a predictive variable. This was achieved by re-evaluating feature selection and retraining the chosen model with the added salinity factor.
  • Model Evaluation: The performance of the models was evaluated using metrics such as Root Mean Square Error (RMSE) and the coefficient of determination (R²). Cross-validation was also employed to ensure the robustness and generalization ability of the final model.

Datasets

  • Zhuhai-1 Hyperspectral Satellite Data: Used for remote sensing imagery. The data was acquired on September 6, 2023, and consists of 32 bands with a spatial resolution of 10m and a spectral range of 400-1000 nm.
  • Field Data: 30 valid samples of chlorophyll-a concentration and corresponding salinity measurements were collected from Lake Daihai on September 6, 2023. These samples were used for model training and validation.

Key Insights and Ideas

  • Salinity is a critical factor: The study unequivocally demonstrates that salinity significantly impacts chlorophyll-a concentration in hypersaline lakes. Ignoring salinity leads to substantial underestimation or overestimation of chlorophyll-a.
  • Random Forest excels: Among the tested machine learning models, Random Forest consistently showed superior performance in both initial comparisons and after incorporating salinity. Its ability to handle complex relationships and high-dimensional data makes it well-suited for this task.
  • PCA for efficient feature selection: PCA proved effective in reducing the dimensionality of hyperspectral data, identifying key spectral bands that are most relevant for chlorophyll-a inversion, thereby simplifying the model and improving efficiency.
  • Hyperspectral data's potential: The study highlights the value of hyperspectral remote sensing for detailed water quality monitoring, especially when combined with appropriate analytical techniques.

Key Results

  • Model Performance without Salinity:
    • Model Stacking: RMSE = 0.109, R² = 0.52
    • Ridge Regression: RMSE = 0.084, R² = 0.61
    • Random Forest: RMSE = 0.056, R² = 0.64
  • Model Performance with Salinity:
    • Random Forest (with salinity): RMSE = 0.047, R² = 0.92
  • The introduction of the salinity factor led to a significant improvement in the Random Forest model's accuracy, with R² increasing from 0.64 to 0.92 and RMSE decreasing from 0.056 to 0.047.
  • The spatial distribution maps of chlorophyll-a concentration showed distinct patterns, with higher concentrations often observed in coastal areas and lower concentrations in the lake's center, influenced by factors like thermal discharge from a power plant and riverine inflow.

Limitations and Future Work

  • Data Availability: The study mentions that the data is available upon request from the corresponding author, with confidentiality requirements from a government department. This suggests that public access to the dataset might be limited.
  • Generalizability: While the model performed well for Lake Daihai, its generalizability to other hypersaline lakes with different hydrological and environmental characteristics would require further investigation.
  • Future Work: The authors suggest that continuous introduction of key variables and model optimization can further improve prediction accuracy. Future research could explore other environmental parameters that might influence chlorophyll-a concentration in hypersaline environments. The study also implies the need for further in-depth study and optimization to improve inversion accuracy, particularly in bridging the gap between retrieved and actual chlorophyll-a values.

Summary updated

BibTeX
@misc{wang2024inversion,
  author = {Nan Wang and Zhiguo Wang and Pingping Huang and Yongguang Zhai and Xiangli Yang and Jianyu Su},
  title = {{Inversion Method for Chlorophyll-a Concentration in High-Salinity Water Based on Hyperspectral Remote Sensing Data}},
  year = {2024},
  doi = {10.3390/s24134181},
}
Figure