TL;DR
This paper introduces a Convolutional Neural Network (CNN) approach to accurately estimate chlorophyll-a (Chl-a) concentrations in the optically complex coastal waters of Hong Kong, outperforming traditional machine learning models. The study also investigates the impact of input data characteristics, such as patch size and spectral bands, on the CNN's performance, providing valuable insights for future water quality monitoring.
What Problem Does This Paper Address?
Accurately measuring chlorophyll-a (Chl-a) concentration is crucial for understanding and monitoring the health of aquatic ecosystems, particularly for assessing eutrophication levels. However, traditional methods like laboratory analysis are time-consuming and labor-intensive. While remote sensing offers a more scalable solution, traditional algorithms struggle in optically complex coastal waters due to variations in water constituents. Existing pixel-based machine learning models also fail to capture the spatial context, which is vital for accurate inversion. This paper aims to bridge this gap by developing a more robust Chl-a inversion model that leverages the spatial information processing capabilities of CNNs in the challenging Hong Kong coastal environment.
Methods
The researchers employed a Convolutional Neural Network (CNN) as their primary model for Chl-a inversion. Unlike traditional pixel-based machine learning methods (Support Vector Regression, Random Forest, XGBoost, and BP Neural Network), CNNs process data in image "patches," allowing them to learn spatial relationships and contextual information. This approach is particularly beneficial for remote sensing data where neighboring pixels often share similar characteristics.
Key methodological aspects include:
- Sentinel-2 MSI Data: Utilized Sentinel-2 satellite imagery, a high-resolution multispectral sensor, for its rich spectral information.
- Data Preprocessing: Standardized procedures like atmospheric correction (using Sen2cor) and resampling to a uniform 10m spatial resolution were applied to ensure data consistency.
- Patch-Based Analysis: The CNN model was trained and evaluated using image patches of varying sizes (from 3x3 to 15x15 pixels) to determine the optimal spatial context for inversion.
- Model Evaluation: Performance was assessed using standard metrics: R-squared (R²), Root Mean Square Error (RMSE), and Mean Relative Error (MRE).
- Interpretability: Shapley Additive Explanations (SHAP) was used to understand which spectral bands contributed most to the CNN model's predictions, offering insights into the "black box" nature of deep learning models.
Datasets
The study utilized:
- Sentinel-2 MSI Satellite Imagery: Acquired for the Hong Kong coastal area in April and October 2022.
- In-situ Chl-a Data: Empirical data collected from 76 fixed monitoring stations by the Hong Kong Environmental Protection Department. This data, collected in November-December 2021 and January 2022, was used for training and validation. A total of 213 sampling points were used after excluding cloud-obscured data.
Key Insights and Ideas
- CNNs Excel in Complex Coastal Waters: The study strongly supports the hypothesis that CNNs, by considering spatial context, are superior to traditional pixel-based machine learning models for Chl-a inversion in optically complex coastal environments.
- Patch Size Matters: The size of the input image patch significantly impacts CNN performance. Too small a patch might miss crucial spatial features, while too large a patch can introduce too much variability, hindering accurate learning. The 7x7 patch size emerged as the optimal choice in this study.
- Spectral Band Importance is Revealed: SHAP analysis provided a clear understanding of which Sentinel-2 bands are most influential for Chl-a estimation. This is crucial for optimizing future model development and potentially reducing data processing requirements by focusing on key bands.
- Spatial Distribution Patterns: The generated Chl-a maps revealed distinct spatial patterns, with higher concentrations in semi-enclosed northeastern areas (like Tolo Harbour) and lower concentrations in open southeastern waters, aligning with geographical and environmental factors.
Key Results
- CNN Superiority: The CNN model achieved the best performance among the five tested models, with an R² of 0.810, an RMSE of 1.165 µg/L, and an MRE of 35.578%.
- Optimal Patch Size: A 7x7 pixel patch size yielded the highest R² (0.810) and the lowest RMSE (1.165 µg/L) and MRE (35.578%) for the CNN model.
- Most Important Bands: SHAP analysis identified bands B6, B12, and B8 as the most important for the CNN model's Chl-a inversion. B6, B12, and B2 positively influenced the model's output, while B8, B10, B11, and B4 had a mitigating effect.
- Spatial Distribution Maps: The study successfully generated thematic maps showing Chl-a distribution for April and October 2022, highlighting higher concentrations in northeastern and southwestern areas and lower concentrations in southeastern regions.
Limitations and Future Work
The authors acknowledge several limitations and suggest future research directions:
- Patch Size Sensitivity: While 7x7 was optimal, the study notes that patch size selection is critical and may need to be tailored to specific environments.
- Temporal Resolution: The reliance on specific time points (April and October) might not fully capture seasonal variations. Future work should involve extended time series analysis using more satellite imagery to improve adaptability to seasonal and environmental changes.
- Inclusion of Other Parameters: The model currently focuses solely on Chl-a. Expanding the model to include other water quality indicators like total nitrogen (TN) and total phosphorus (TP) would provide a more comprehensive assessment.
- Real-time Monitoring: Developing a framework for real-time monitoring that integrates in-situ measurements with the CNN model could account for dynamic weather and climate effects on water quality.
- Model Interpretability: While SHAP provides insights, the inherent complexity of CNNs still presents challenges in fully understanding all aspects of their decision-making process. Further exploration into model interpretability is warranted.
Summary updated