Remote Sensing and Machine Learning for Eutrophication Detection: Assessing the Trophic State in Reservoirs Using Multispectral Indices and Deep Learning
Eutrophication is defined as the excessive enrichment of water bodies with nutrients, leading to planktonic algal blooms. This phenomenon causes significant ecological disturbances, such as extreme fluctuations in oxygen levels, reduced water transparency, and increased oxygen consumption in deeper...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11053172/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Eutrophication is defined as the excessive enrichment of water bodies with nutrients, leading to planktonic algal blooms. This phenomenon causes significant ecological disturbances, such as extreme fluctuations in oxygen levels, reduced water transparency, and increased oxygen consumption in deeper water layers. The main challenge in assessing the trophic status of water bodies lies in the need to collect data in situ, which complicates the analysis of multiple reservoirs, especially in large-scale studies. To address this limitation, remote sensing techniques are proposed as an alternative method to assess trophic states. In this study, satellite imagery is utilized to monitor 20 reservoirs from 2017 to 2024, resulting in a dataset of 496 images paired with corresponding in situ measurements. From these images, 198 multispectral indices are computed for analysis, which are combined with 12 primary spectral bands. These indices and bands undergo an automated feature selection process, which identifies modified chlorophyll absorption ratio index, near-infrared reflectance of vegetation, difference vegetation index, and weighted difference vegetation index as the most relevant features for distinguishing whether the reservoir is in a eutrophic state. Using these indices, false-color images are generated by combining red, green, blue, alpha channels. When integrated with a deep learning model, these images enable the remote classification of the reservoirs’ trophic state. Balanced accuracy scores above 0.89 and accuracy levels above 0.92 are achieved across the entire reservoir dataset. These results demonstrate the model’s strong ability to distinguish between eutrophic and noneutrophic conditions, regardless of the specific reservoir. These results underline the potential of remote sensing systems to effectively detect variations in water quality, characterize the trophic status of reservoirs, and monitor changes in their environmental conditions over time. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |