The influence of natural and anthropogenic drivers on vegetation dynamics was explored using a Generalized Additive Model (GAM). This model evaluated non-linear relationships between vegetation changes and key factors:
Spartina alterniflora was primarily influenced by marine environmental variables such as salinity and wave height.
Phragmites australis and Suaeda salsa were affected by precipitation, anthropogenic pressures (e.g., aquaculture), and interspecies competition.
Understanding these drivers supports adaptive ecosystem management and invasive species control.
GBF Alignment: Supports GBF Targets 6 and 8. Contribution: Predictive models improve on reactive conservation, offering measurable driver insights.
GAM effectively captured complex, non-linear interactions between drivers and vegetation changes.
Integration of environmental and human activity datasets enhanced the robustness of driver attribution.
Continuous data collection and model refinement are critical for long-term predictive accuracy.
Mechanistic understanding of ecological drivers underpins the development of forward-looking conservation strategies.
A comprehensive geospatial database was developed, integrating vegetation cover data derived from remote sensing with key environmental, climatic, and anthropogenic variables. Included metrics encompassed soil salinity, sea surface temperature, seawater salinity, and locations of aquaculture ponds, providing a robust analytical foundation.
GBF Alignment: Supports GBF Target 21. Contribution: Integrates diverse data layers for holistic analysis, adding value to fragmented conservation datasets.
Field validation confirmed the accuracy of remote sensing interpretations (see Figures 1 and 2).
The database facilitated the integration of spatial and environmental data, supporting multi-variable analyses and ecological modeling.
Accurate ground-truth data are vital for validating remote sensing outputs and ensuring database reliability.
A well-structured, multi-source database improves analysis efficiency and enables more sophisticated correlation and causality studies.
Vegetation index time series were smoothed using Gaussian fitting to reduce noise and extract key phenological features. A random forest deep learning algorithm was applied to classify wetland vegetation into three dominant types: Spartina alterniflora, Phragmites australis, and Suaeda salsa. Classification accuracy from 1990 to 2022 was validated through field surveys.
GBF Alignment: Contributes to GBF Target 6. Contribution: Reduces invasive species impact by accurately identifying Spartina alterniflora for targeted control, addressing a key biodiversity threat.
Gaussian curve fitting effectively minimized noise in raw vegetation index curves, enhancing classification accuracy.
The random forest algorithm leveraged spectral differences between species, enabling robust feature extraction and reliable identification.
Spectral features related to vegetation moisture and structural attributes significantly improved interspecies separability.
Preprocessing steps such as curve fitting and denoising were essential for improving the reliability of long-term classification.
Using the Google Earth Engine (GEE) platform, Landsat series remote sensing data from 1990 to 2022 were systematically acquired, encompassing TM5, ETM+ (Landsat 7), OLI (Landsat 8), and OLI (Landsat 9) sensors. To ensure data quality for subsequent analyses, key spectral bands—near-infrared (NIR), red, and green—were selected and fused.
GBF Alignment: Supports GBF Target 21. Contribution: Enhances decision-making with real-time, validated datasets, adding value to existing conservation efforts through technological innovation.
Only remote sensing images with cloud cover ≤10% were selected, followed by radiometric and atmospheric correction via batch processing.
Vegetation-specific information was extracted using optimized band combinations, particularly leveraging the high reflectance of vegetation in the NIR range.
Limitations in spatial, temporal, and spectral resolutions introduced potential uncertainties, highlighting the importance of robust radiometric and geometric correction methods.
Data fusion across different Landsat sensors was essential for achieving consistent long-term time series, though it required substantial additional processing to harmonize spatial and temporal resolutions.
