Can single empirical algorithms accurately predict inland shallow water quality status from high resolution, multi-sensor, multi-temporal satellite data?

Ioanna Theologou, M. Patelaki, Κωνσταντίνος Καράντζαλος2015 · DOI

Abstract

Assessing and monitoring water quality status through timely, cost-effective, and accurate means is of fundamental importance for environmental management and policy-making. There is a current need for validated methodologies that can effectively exploit the vast amount of Earth observation imaging datasets from various high-resolution satellite multispectral sensors in an unsupervised way. This study investigates the applicability of single empirical models across multi-temporal, multi-sensor datasets for predicting key water quality indicators in shallow inland systems, specifically Lake Karla in Greece. By fusing Landsat 7 and Landsat 8 data with in-situ ground truth and hyperspectral measurements over a four-year period, we found that while single linear regression models fail to establish concrete relations across diverse observations, sensor-specific and depth-stratified models yield promising results. High correlation rates were achieved for chlorophyll-a ($R^2=89.80\%$), dissolved oxygen ($R^2=88.53\%$), and conductivity ($R^2=88.18\%$), while nutrients like total phosphorus and nitrates showed lower correlation rates.

Introduction

Monitoring the health of inland water bodies is critical for environmental management, yet traditional in-situ sampling is often logistically challenging, expensive, and limited to discrete point data. Remote sensing via satellites offers a potential solution for continuous spatial monitoring. However, inland waters—particularly shallow systems like Lake Karla—are "optically complex." Unlike the open ocean, these waters contain varying concentrations of chlorophyll, inorganic suspended solids, and dissolved organic matter that create overlapping spectral signatures.

Figure 1: Geographic location of the study area, Lake Karla, in Greece.

This research addresses a fundamental question: Can a single, robust empirical model accurately predict water quality across different satellite sensors (Landsat 7 vs. Landsat 8) and over several years, or does the inherent complexity of shallow water require a more tailored, stratified approach?

Methodology

The study utilized a comprehensive dataset covering a four-year period at Lake Karla. The workflow integrated satellite imagery with ground truth data through several key stages:

  1. Pre-processing: Raw satellite digital numbers were converted into top-of-atmosphere reflectance using ATCOR2 and MODTRAN4 to correct for atmospheric interference.
  2. Gap-Filling: A specific algorithm was implemented to resolve the "striping" data gaps in Landsat 7 imagery caused by the failure of its Scan Line Corrector (SLC-off).
  3. Empirical Modeling: Hundreds of band combinations, ratios, and indices (such as Red/NIR ratios) were tested to identify the strongest correlations with water quality parameters.
  4. Stratified Analysis: Modeling scenarios were developed for relatively deep sections, very shallow sections, and the lake as a whole to account for bottom reflectance and aquatic vegetation.
  5. Validation: Satellite-derived results were cross-validated against in-situ chemical analyses and high-resolution field hyperspectral data collected using a GER 1500 spectroradiometer.
Figure 2: Comprehensive overview of the data processing workflow, including in-situ sampling locations, raw vs. processed satellite imagery, and the final value-added water quality map.

Results and Discussion

The analysis revealed that a "universal" linear regression model cannot accurately bridge different sensors and time periods. Variations in sensor design and the shifting optical properties of the lake necessitate "per-sensor" and "per-depth" models to maintain high accuracy.

Depth emerged as a decisive factor; models trained specifically on deeper sections performed significantly better than those attempting to model the entire lake simultaneously. This is likely due to the interference of bottom reflectance and submerged vegetation in shallower areas. Spectral analysis indicated that Chlorophyll-a correlated best with Red and Near-Infrared (NIR) bands, while nutrients like Ammonium showed stronger relationships with Green and Blue bands.

When using optimized, sensor-specific models, the study achieved remarkably high correlation rates for several key indicators:

  • Chlorophyll-a (chl-a): $R^2 = 89.80\%$
  • Dissolved Oxygen (DO): $R^2 = 88.53\%$
  • Conductivity: $R^2 = 88.18\%$
  • Ammonium ($NH_4$): $R^2 = 87.2\%$
  • pH: $R^2 = 86.35\%$

In contrast, Total Phosphorus ($R^2 = 70.55\%$) and Nitrates ($R^2 = 55.50\%$) proved more difficult to predict, suggesting these parameters may require higher spectral resolution or more advanced analytical modeling techniques.

Conclusion

While a single empirical algorithm remains elusive for multi-sensor, multi-temporal monitoring of optically complex shallow waters, this study demonstrates that consistent per-sensor and per-depth prediction models can be successfully established. These models provide a cost-effective means of generating high-resolution spatial maps of water quality. Future work should focus on inter-calibration between sensors and leveraging newer platforms like Sentinel-2 to move toward standardized, operational global water-quality monitoring services.

Summary updated

Abstract

Assessing and monitoring water quality status through timely, cost effective and accurate manner is of fundamental importance for numerous environmental management and policy making purposes. Therefore, there is a current need for validated methodologies which can effectively exploit, in an unsupervised way, the enormous amount of earth observation imaging datasets from various high-resolution satellite multispectral sensors. To this end, many research efforts are based on building concrete relationships and empirical algorithms from concurrent satellite and in-situ data collection campaigns. We have experimented with Landsat 7 and Landsat 8 multi-temporal satellite data, coupled with hyperspectral data from a field spectroradiometer and in-situ ground truth data with several physico-chemical and other key monitoring indicators. All available datasets, covering a 4 years period, in our case study Lake Karla in Greece, were processed and fused under a quantitative evaluation framework. The performed comprehensive analysis posed certain questions regarding the applicability of single empirical models across multi-temporal, multi-sensor datasets towards the accurate prediction of key water quality indicators for shallow inland systems. Single linear regression models didn’t establish concrete relations across multi-temporal, multi-sensor observations. Moreover, the shallower parts of the inland system followed, in accordance with the literature, different regression patterns. Landsat 7 and 8 resulted in quite promising results indicating that from the recreation of the lake and onward consistent per-sensor, per-depth prediction models can be successfully established. The highest rates were for chl-a (r2=89.80%), dissolved oxygen (r2=88.53%), conductivity (r2=88.18%), ammonium (r2=87.2%) and pH (r2=86.35%), while the total phosphorus (r2=70.55%) and nitrates (r2=55.50%) resulted in lower correlation rates.
BibTeX
@inproceedings{theologou2015can,
  author = {Ioanna Theologou and M. Patelaki and Κωνσταντίνος Καράντζαλος},
  title = {{Can single empirical algorithms accurately predict inland shallow water quality status from high resolution, multi-sensor, multi-temporal satellite data?}},
  booktitle = {˜The},
  year = {2015},
  doi = {10.5194/isprsarchives-xl-7-w3-1511-2015},
}
Figure