Proposal of a Method to Determine the Correlation between Total Suspended Solids and Dissolved Organic Matter in Water Bodies from Spectral Imaging and Artificial Neural Networks

Maurício Roberto Veronez, Lucas S. Kupssinskü, Tainá T. Guimarães, Emilie C. Koste, Juarez Machado da Silva, Lais V. de Souza, William F. M. Oliverio, Rogélio S. Jardim, Ismael É. Koch, Jonas G. de Souza, Luiz Gonzaga Jr., Frederico Fábio Mauad, Leonardo Campos Inocencio, Fabiane BordinSensors 2018 · DOI · OpenReview

TL;DR

This paper proposes a cost-effective method to monitor water quality by using a low-cost RGB camera on a drone to predict multispectral bands typically captured by expensive sensors. By training an Artificial Neural Network (ANN), they successfully correlated these predicted bands with Total Suspended Solids (TSS) and Dissolved Organic Matter (DOM), offering a promising alternative for water quality assessment.

What Problem Does This Paper Address?

Monitoring water quality in lakes and reservoirs is crucial for environmental health and resource management. Traditional methods involve manual water sampling and laboratory analysis, which are time-consuming, costly, and often impractical for large or inaccessible water bodies. Remote sensing offers a more efficient solution, but high-resolution multispectral sensors, essential for accurately assessing parameters like Total Suspended Solids (TSS) and Dissolved Organic Matter (DOM), are prohibitively expensive. This creates a barrier to widespread adoption, especially for smaller organizations or localized monitoring efforts. The paper aims to bridge this gap by developing a method that leverages affordable RGB cameras on Unmanned Aerial Vehicles (UAVs) to achieve similar water quality monitoring capabilities.

Methods

The core of the proposed method lies in an Artificial Neural Network (ANN) trained to predict specific multispectral bands (LANDSAT 8 OLI bands b4, b5, and b9) using the Red, Green, and Blue (RGB) bands captured by a low-cost camera mounted on a UAV.

Here's a breakdown of the workflow:

  1. UAV Overflight and Image Acquisition: A hexacopter UAV equipped with a Canon RGB sensor captured images of the study lake. Simultaneously, water samples were collected from 21 points.
  2. Georeferencing and Orthophoto Generation: UAV images were georeferenced using ground control points (GCPs) and processed using Structure from Motion (SfM) techniques to create a high-resolution orthophoto.
  3. Multispectral Band Prediction with ANN:
    • Input: RGB bands from the low-cost UAV sensor.
  4. Output: Predicted multispectral bands (b4, b5, b9 from LANDSAT 8 OLI).
  5. Training: 80% of the data was used for training the ANN, with Adaptive Moment Estimation (Adam) as the optimization algorithm. The network topology was optimized through experimentation.
  6. Validation: The remaining 20% of the data was used for validation.
  7. Index Calculation: Using the ANN-predicted multispectral bands, Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI) were calculated. Two versions of these indices were explored: one using LANDSAT bands b4 and b5, and another using b4 and b9.
  8. Correlation Analysis: Polynomial regression was used to correlate the calculated NDVI and NDWI values with the laboratory-measured TSS and DOM concentrations.

The novelty lies in using an ANN to bridge the gap between low-cost RGB imagery and the spectral information required for water quality parameters, effectively "synthesizing" multispectral data from cheaper sensors.

Datasets

  • UAV RGB Imagery: Acquired by a Canon RGB sensor mounted on a hexacopter UAV over a lake at Unisinos University, Brazil.
  • LANDSAT 8 (OLI) Imagery: Used as a reference for the target multispectral bands (b4, b5, b9). These images were atmospherically corrected.
  • Water Samples: 21 water samples were collected from the study lake.
  • Laboratory Analysis Data: TSS and DOM concentrations were determined from the collected water samples using gravimetric methods.

No new benchmark datasets were introduced; the study utilized existing satellite data and field-collected samples.

Key Insights and Ideas

  • Democratizing Multispectral Data: The most significant insight is the demonstration that an ANN can effectively "predict" multispectral bands from low-cost RGB imagery. This opens doors for more accessible and affordable remote sensing applications in water quality monitoring.
  • ANN as a "Spectral Synthesizer": The ANN acts as a powerful tool to bridge the spectral resolution gap, allowing us to leverage the high spatial resolution of UAV RGB imagery with the spectral information typically found in more expensive multispectral sensors.
  • NDWI's Superiority for Water Quality: The study consistently shows that the NDWI, particularly when derived from specific spectral bands (b5 or b9), provides stronger correlations with TSS and DOM compared to NDVI. This highlights the importance of selecting appropriate spectral bands for specific water quality parameters.
  • Band Selection Matters: The exploration of using band b9 instead of b5 in the NDVI and NDWI calculations showed a slight improvement in correlation for TSS, suggesting that the longer infrared wavelength captured by b9 might be more sensitive to certain water constituents in this specific environment.

Key Results

  • ANN Prediction Accuracy: The ANN successfully predicted the LANDSAT 8 OLI bands (b4, b5, b9) from the low-cost RGB sensor with an average error of 19%. The linear regression between known and predicted pixel intensities yielded R² values above 0.81 for all three bands.
  • Correlation with TSS and DOM:
    • TSS: Correlations with NDVI and NDWI showed R² values greater than 0.60. The NDWI, especially when derived using band b5 or b9, demonstrated stronger correlations (R² up to 0.77) with TSS, with RMSE values around 1.3-1.4 mg/L.
  • DOM: Correlations with NDWI also showed promising results (R² of 0.60), with RMSE values around 5.3 mg/L. While correlations with NDVI were lower (R² of 0.52), the study noted that direct comparisons with literature for DOM were scarce.
  • Computational Efficiency: The ANN prediction process for generating multispectral bands and subsequently NDVI/NDWI images was computationally efficient, with a cost of around 30.5 seconds for generating the indices.

Limitations and Future Work

  • Water Sampling Conditions: The method's effectiveness is dependent on good light conditions for UAV image acquisition and cloud-free satellite imagery.
  • Orthophoto Generation for Large Water Bodies: Creating accurate orthophotos for very large and homogeneous water bodies can be challenging due to difficulties in identifying homologous points for processing.
  • Seasonal Variation: The current approach does not account for seasonal variations in TSS and DOM.
  • Future Work: The authors suggest incorporating climate variables (wind velocity, rainfall, temperature) into the ANN input to improve the model's robustness and account for temporal variations in water conditions. They also propose applying this method to other water bodies.

Summary updated

Abstract

Water quality monitoring through remote sensing with UAVs is best conducted using multispectral sensors; however, these sensors are expensive. We aimed to predict multispectral bands from a low-cost sensor (R, G, B bands) using artificial neural networks (ANN). We studied a lake located on the campus of Unisinos University, Brazil, using a low-cost sensor mounted on a UAV. Simultaneously, we collected water samples during the UAV flight to determine total suspended solids (TSS) and dissolved organic matter (DOM). We correlated the three bands predicted with TSS and DOM. The results show that the ANN validation process predicted the three bands of the multispectral sensor using the three bands of the low-cost sensor with a low average error of 19%. The correlations with TSS and DOM resulted in R2 values of greater than 0.60, consistent with literature values.
BibTeX
@article{veronez2018proposal,
  author = {Maurício Roberto Veronez and Lucas S. Kupssinskü and Tainá T. Guimarães and Emilie C. Koste and Juarez Machado da Silva and Lais V. de Souza and William F. M. Oliverio and Rogélio S. Jardim and Ismael É. Koch and Jonas G. de Souza and Luiz Gonzaga Jr. and Frederico Fábio Mauad and Leonardo Campos Inocencio and Fabiane Bordin},
  title = {{Proposal of a Method to Determine the Correlation between Total Suspended Solids and Dissolved Organic Matter in Water Bodies from Spectral Imaging and Artificial Neural Networks}},
  journal = {Sensors},
  year = {2018},
  doi = {10.3390/s18010159},
}
Figure