Introduction
The rapid and often unregulated expansion of rubber plantations (Hevea brasiliensis) across tropical regions, particularly on Hainan Island, China, has raised significant ecological concerns. Replacing natural rainforests with monoculture rubber stands leads to biodiversity loss and altered hydrological cycles. Accurate, large-scale monitoring is essential for sustainable land management. However, distinguishing rubber plantations from natural evergreen forests via satellite imagery is notoriously difficult because they often share nearly identical spectral signatures—a challenge known as the "different objects, same spectrum" phenomenon.
Methodology: Leveraging the "Winter Sleep"
This research addresses the limitations of traditional pixel-based classification by combining Object-Oriented Analysis (OOA) with Multi-temporal Phenology.
The Phenological Advantage
The key innovation lies in identifying the rubber tree's unique "winter window." Unlike the surrounding tropical evergreen forests, rubber trees undergo a distinct dormant period between January and February, during which they shed their leaves. By analyzing MODIS NDVI time-series data, researchers identified this period as the optimal time for differentiation.
Object-Oriented Decision Tree
Instead of classifying individual pixels, which often results in "salt-and-pepper" noise, the study uses OOA to group pixels into meaningful "objects" based on spectral, spatial, and textural similarities. A five-level decision tree was then constructed using several key indices:
- NDVI (Normalized Difference Vegetation Index): To measure vegetation greenness.
- LSWI (Land Surface Water Index): To detect changes in leaf water content and soil moisture, which drop significantly during the rubber tree's dormant phase.
- MNDWI (Modified Normalized Difference Water Index): To mask out water bodies.
- Slope Analysis: Derived from a 30m Digital Elevation Model (DEM) to filter out topographic shadows in mountainous terrain.
Results and Performance
The integration of multi-temporal data (comparing February dormant-season images with December growing-season images) significantly improved classification performance.
- Overall Accuracy: The model achieved a high overall accuracy of 94.5%.
- Rubber-Specific Precision: The producer accuracy for rubber plantations reached 96.2% when using two-temporal images.
- Temporal Comparison: Using a single image from December (growing season) yielded only 78.1% accuracy. A single February image (dormant season) improved this to 86.1%. The jump to 96.2% confirms that capturing the phenological transition is critical for high-fidelity mapping.
Discussion and Future Outlook
The study highlights that the Land Surface Water Index (LSWI) is as vital as NDVI; as rubber trees lose leaves, the satellite detects increased soil moisture and decreased canopy water, providing a robust secondary signature.
Despite these successes, tropical cloud cover remains a significant barrier to obtaining clear Landsat imagery during the narrow "winter window." Future research should focus on integrating data from satellites with higher revisit frequencies, such as China’s HJ-1A/1B constellations, to ensure the capture of these brief phenological stages. Additionally, accounting for human management factors like irrigation will be necessary to further refine these models for global application.
Summary updated