A Method for Chlorophyll-a and Suspended Solids Prediction through Remote Sensing and Machine Learning

Lucas Silveira Kupssinskü, Tainá Thomassim Guimarães, Eniuce Menezes De Souza, Daniel C. Zanotta, Maurício Roberto Veronez, Luiz Gonzaga Jr., Frederico Fábio MauadSensors 2020 · DOI · OpenReview

TL;DR

This paper introduces a robust methodology for predicting chlorophyll-a and Total Suspended Solids (TSS) in water bodies using remote sensing data and machine learning. By combining satellite and drone imagery with advanced ML techniques, the approach offers a cost-effective and scalable solution for water quality monitoring, outperforming traditional methods.

What Problem Does This Paper Address?

Monitoring water quality in lakes and reservoirs is crucial for environmental management, but traditional methods involving laboratory analysis of collected samples are expensive, time-consuming, and provide only point-in-time data, failing to capture the dynamic spatio-temporal variations. Eutrophication, a process leading to water deterioration, is often indicated by elevated levels of chlorophyll-a, turbidity, and suspended solids. These parameters are optically active, meaning they interact with light in a way that can be detected by remote sensing. However, accurately estimating these parameters from remote sensing data, especially in complex aquatic environments with non-linear relationships, presents a significant challenge. Existing methods often struggle with varying spatial resolutions and the need for extensive ground-truthing.

Methods

The proposed methodology integrates several key steps:

  • Water Quality Assessment: This involves strategic water sample collection, laboratory analysis for chlorophyll-a and TSS (using gravimetric and spectrophotometric methods, respectively), and recording precise geospatial coordinates.
  • Remote Sensing: Data is acquired using either Sentinel-2 satellite imagery or Unmanned Aerial Vehicles (UAVs). Sentinel-2 was chosen for its spectral bands, including the red edge, and higher spatial resolution compared to Landsat. UAVs offer even higher spatial resolution, crucial for smaller water bodies. Atmospheric correction is applied to satellite data, while for low-altitude UAV flights, it's deemed less critical. Key spectral bands (Blue, Green, NIR) are selected for their relevance to chlorophyll-a and TSS.
  • Geographic Information System (GIS): To increase the dataset size and spatial coverage, Inverse Distance Weighting (IDW) interpolation is used to create continuous maps from discrete sample points. This is followed by data integration and extraction into a raster format, ensuring consistent spatial resolution and projection. A grid of points is generated, and an erosion step is applied to discard pixels potentially affected by margins, shadows, or shallow bottoms, ensuring focus on pure water pixels.
  • Machine Learning (ML): A supervised learning approach is employed. Data is preprocessed through standardization and one-hot encoding for categorical features. A train-test split (90% train, 10% test) is performed to prevent bias. K-fold cross-validation (specifically 10-fold) is used for robust model training and parameter tuning. Several ML algorithms are evaluated, including Linear Regression, LASSO, Support Vector Regression (SVR), K Nearest Neighbors (KNN), Random Forest (RF), and Artificial Neural Networks (ANN). Model performance is assessed using R² and Mean Squared Error (MSE) metrics.

Datasets

The study utilizes two distinct real-world datasets:

  • Unisinos Lake: A small artificial lake (0.025 km²) in São Leopoldo, Brazil. Data was collected using UAV imagery (SenseFly, Swinglet CAM with Canon ELPH 110HS camera) and field samples over two years (2016 and 2017). This dataset was expanded from 42 initial points to 35,028 data points through interpolation.
  • Broa Reservoir: A larger dam (6.8 km²) in Itirapina, Brazil. Data was collected using Sentinel-2 satellite imagery and field samples across three occasions in 2018. This dataset was expanded from 30 initial points to 3,933 data points.

The paper does not introduce a new benchmark dataset but rather applies the methodology to existing, real-world water bodies with collected field and remote sensing data.

Key Insights and Ideas

  • IDW Interpolation as a Data Augmentation Tool: The study highlights how IDW interpolation, when applied carefully, can significantly increase the number of data points for ML training, making studies feasible in scenarios where collecting a vast number of physical samples would be impractical.
  • Importance of Feature Space Definition: The paper demonstrates that carefully selecting input features (spectral bands, pH, and spatial location encoded as quadrants) is crucial to avoid overfitting and achieve better model performance. Directly using raw x-y coordinates led to overfitting.
  • Non-linear Models Excel: The results strongly suggest that non-linear ML models, particularly Random Forest (RF) and Artificial Neural Networks (ANN), outperform linear models for predicting chlorophyll-a and TSS, indicating complex, non-linear relationships between spectral reflectance and water quality parameters.
  • Robustness Across Different Water Bodies: The methodology proved effective in two distinct water bodies with different sizes, usage patterns, and environmental conditions, suggesting its potential for broader application.

Key Results

  • High Predictive Accuracy: The Random Forest (RF) algorithm consistently achieved the highest R² values, often above 0.8, and the lowest MSE values across both study areas for both chlorophyll-a and TSS prediction. For Broa, RF outperformed other algorithms, while for Unisinos, RF, KNN, and ANN performed comparably well.
  • Spatial Distribution Consistency: Visual comparison of maps generated by the RF algorithm and IDW interpolation showed similar spatial distributions, with the RF maps potentially capturing finer details and smoother transitions due to their consideration of spectral reflectance and interpolated pH.
  • Overcoming Data Limitations: For Unisinos Lake, where laboratory data had a limited range for chlorophyll-a, the interpolation process and ML models helped to estimate values beyond the directly measured range.
  • R² values close to one: Scatter plots comparing measured and estimated values for the best models (RF) showed R² values close to 1, indicating a strong correlation between predicted and actual measurements.

Limitations and Future Work

  • Seasonality Not Explicitly Modeled: The paper acknowledges that seasonality was not explicitly modeled as an input variable. While reflectance data implicitly captured seasonal variations due to sampling dates, a dedicated seasonality component could further improve models.
  • Potential for Other ML Techniques: The study explored six ML learners but suggests that other advanced techniques, such as Convolutional Neural Networks (CNNs), could benefit from incorporating texture features and learning through backpropagation, potentially improving performance.
  • Adaptability to Local Features: While the method is flexible regarding sampling points, imagery resolution, and algorithm selection, the authors recommend conducting new experiments to fine-tune models for specific local environmental characteristics to achieve optimal performance in different scenarios.
  • Further Exploration of Spatial Features: While using quadrants for spatial information helped avoid overfitting, further research could explore more sophisticated ways to incorporate spatial context.

Summary updated

Abstract

Total Suspended Solids (TSS) and chlorophyll-a concentration are two critical parameters to monitor water quality. Since directly collecting samples for laboratory analysis can be expensive, this paper presents a methodology to estimate this information through remote sensing and Machine Learning (ML) techniques. TSS and chlorophyll-a are optically active components, therefore enabling measurement by remote sensing. Two study cases in distinct water bodies are performed, and those cases use different spatial resolution data from Sentinel-2 spectral images and unmanned aerial vehicles together with laboratory analysis data. In consonance with the methodology, supervised ML algorithms are trained to predict the concentration of TSS and chlorophyll-a. The predictions are evaluated separately in both study areas, where both TSS and chlorophyll-a models achieved R-squared values above 0.8.
BibTeX
@article{kupssinsku2020method,
  author = {Lucas Silveira Kupssinskü and Tainá Thomassim Guimarães and Eniuce Menezes De Souza and Daniel C. Zanotta and Maurício Roberto Veronez and Luiz Gonzaga Jr. and Frederico Fábio Mauad},
  title = {{A Method for Chlorophyll-a and Suspended Solids Prediction through Remote Sensing and Machine Learning}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20072125},
}
Figure