Evaluation of Vegetation Index-Based Curve Fitting Models for Accurate Classification of Salt Marsh Vegetation Using Sentinel-2 Time-Series

Chao Sun, Jialin Li, Luodan Cao, Yongchao Liu, Song Jin, Bingxue Zhao2020 · DOI

TL;DR

This paper evaluates different vegetation indices (VIs) and curve-fitting methods for classifying salt marsh vegetation using Sentinel-2 time-series data. The study finds that the SAVI and NDVI VIs, combined with Double Logistic (DL) or Two-term Fourier (TF) fitting methods, offer the most robust classification, with specific recommendations based on regional climate.

What Problem Does This Paper Address?

Salt marshes are vital ecosystems that provide numerous ecological services and act as indicators of coastal health. Accurate and large-scale mapping of salt marsh vegetation is crucial for monitoring their status, especially given threats from climate change and human activities. Traditional methods using single high-spatial-resolution images are limited by cost and availability, while coarser resolution time-series data (like MODIS) lack the detail needed for heterogeneous salt marsh environments. Sentinel-2 offers a solution with its frequent revisit times and moderate spatial resolution, enabling the construction of dense time-series. However, the effectiveness of these time-series analyses hinges on selecting appropriate vegetation indices (VIs) and curve-fitting methods to handle the complexities of coastal environments, such as tidal flooding and cloud cover, which introduce noise and data gaps. This paper aims to critically compare various VI and curve-fitting combinations to identify the optimal approach for salt marsh classification.

Methods

The study employed a pixel-differential time-series approach using Sentinel-2 MSI Level-1C imagery. Key steps included:

  • Atmospheric Correction: Converting Top-Of-Atmosphere (TOA) reflectance to Bottom-Of-Atmosphere (BOA) reflectance using the Sen2Cor processor.
  • Cloud and Shadow Masking: Utilizing the Scene Classification (SC) map to identify and remove pixels affected by clouds, cloud shadows, and snow.
  • Vegetation Index (VI) Calculation: Computing five different VIs: NDVI, SAVI, EVI, GNDVI, and WAVI. These were chosen for their sensitivity to vegetation characteristics and their suitability for salt marsh environments, considering factors like soil brightness, vegetation density, and water influence.
  • Curve Fitting: Applying three curve-fitting methods to the VI time series to smooth noise and extract phenological information: Asymmetric Gaussian (AG), Double Logistic (DL), and Two-term Fourier (TF). These methods were selected for their ability to model vegetation phenology.
  • Phenological Metric Extraction: Deriving six key phenological metrics (Base Value, Maximum Value, Start of Season, End of Season, Rate of Increase, and Rate of Decrease) from the fitted curves.
  • Separability Analysis: Quantifying the separability between different salt marsh vegetation types using the Jeffries-Matusita distance (JMD).
  • Classification Mapping: Generating classification maps using the Random Forest algorithm, with accuracy assessed through Overall Accuracy (OA), Quantity Disagreement (QD), and Allocation Disagreement (AD).

Datasets

The study utilized Sentinel-2 MSI Level-1C images acquired in 2018 for three coastal sites in China: the Yellow River Delta (YRD), Jiangsu Middle Coast (JMC), and Yangtze River Estuary (YRE). A total of 1502 salt marsh vegetation samples were collected through field surveys, Unmanned Aerial Vehicle (UAV) investigations, and Google Earth imagery interpretation. These samples represented six dominant salt marsh vegetation types: Spartina alterniflora (SA), Phragmites australis (PA), Suaeda salsa (SS), Tamarix chinensis (TC), Imperata cylindrica (IC), and Scirpus mariqueter (SM).

Key Insights and Ideas

  • SAVI and NDVI Shine for Stability: The SAVI and NDVI VIs demonstrated superior time-series stability, showing less noise from tidal effects and cloud cover compared to other indices like EVI. This stability is crucial for reliable phenological analysis.
  • EVI's Sensitivity to Noise: The EVI index, while often achieving high fitting accuracy, proved highly sensitive to errors from omitted clouds and shadows, leading to conspicuous outliers that negatively impacted classification.
  • DL and TF Methods Excel: The Double Logistic (DL) and Two-term Fourier (TF) curve-fitting methods generally outperformed the Asymmetric Gaussian (AG) method, especially when combined with stable VIs. DL was particularly effective for capturing phenological trends.
  • Climate Matters for Model Selection: The optimal VI-fitting model combination is influenced by regional climate. For high-latitude regions with continental climates, the SAVI/NDVI-based DL model is recommended. For mid- to low-latitude regions with monsoonal climates, the SAVI/NDVI-based TF model appears more suitable.
  • Separability is Key to Accuracy: High salt marsh vegetation separability, as measured by JMD, directly correlated with accurate classification maps. Models with high separability showed fewer omission and commission errors.

Key Results

  • VI Stability: SAVI exhibited the best time-series stability, while EVI showed poor stability with significant outliers.
  • Fitting Accuracy: EVI, SAVI, and NDVI generally yielded higher fitting accuracy than WAVI and GNDVI. The DL method typically resulted in the highest fitting accuracy, especially when paired with EVI.
  • Vegetation Separability: SAVI and NDVI time series, combined with DL and TF methods, were the most effective for distinguishing salt marsh vegetation types, consistently achieving high OJMD scores across all study sites.
  • Classification Accuracy: The SAVI/NDVI-based DL/TF models achieved the highest classification accuracy. For example, the NDVI+TF model achieved an OA of 0.898 in the YRE. Models with high separability consistently outperformed those with low separability, with a notable difference in overall accuracy.

Limitations and Future Work

  • Cloud and Shadow Impact: Despite cloud masking efforts, omitted clouds and shadows still introduced errors, particularly affecting the blue and green bands and leading to VI time-series instability, especially for EVI.
  • Tidal Flooding Influence: Periodic tidal flooding significantly impacted VI time series, causing abrupt declines and fluctuations, particularly for species with low plant height and density. NDVI and GNDVI were found to be most sensitive to this.
  • Data Distribution Effects: The study highlighted that the suitability of curve-fitting methods is influenced by the distribution of available data, which varies with VI and site characteristics. For instance, DL performed well with data concentrated around the peak season, while TF was more flexible for capturing lagged phenology.
  • Future Work: The paper implicitly suggests that further research could explore more advanced cloud detection algorithms, investigate the impact of different tidal regimes on specific VI indices more deeply, and potentially develop hybrid models that combine the strengths of different VIs and fitting methods to further enhance salt marsh classification accuracy. The authors also emphasize the need to consider data distribution characteristics when selecting models for future phenology-based applications.

Summary updated

BibTeX
@misc{sun2020evaluation,
  author = {Chao Sun and Jialin Li and Luodan Cao and Yongchao Liu and Song Jin and Bingxue Zhao},
  title = {{Evaluation of Vegetation Index-Based Curve Fitting Models for Accurate Classification of Salt Marsh Vegetation Using Sentinel-2 Time-Series}},
  year = {2020},
  doi = {10.3390/s20195551},
}
Figure