Abstract
Accurate high-resolution land cover mapping is essential for environmental monitoring but is often hindered by the trade-off between spatial and temporal resolution in satellite imagery. This paper proposes a Directed Probabilistic Graphical Model (PGM) framework to intelligently fuse 30m Landsat data with coarse-resolution MODIS time-series data. Unlike traditional "blind" data-stacking, the PGM approach employs selective weighting based on data quality, allowing the model to prioritize reliable observations. Evaluation in two diverse regions of China demonstrates significant accuracy improvements, particularly in cloud-affected scenarios, increasing overall accuracy from 74.0% to 81.8% in complex landscapes.
Introduction
Creating precise 30-meter land cover maps involves navigating the limitations of current satellite constellations. Landsat provides the necessary spatial resolution (30m) but suffers from a 16-day revisit cycle, which often results in data gaps due to cloud cover. Conversely, MODIS offers daily revisits but at a much coarser spatial resolution (250m–500m).
Traditional data fusion methods typically stack these sources into a single feature vector for classification. However, this "blind stacking" often fails to account for temporal inconsistencies or localized noise, such as cloud shadows in one image date that are absent in another. There is a critical need for a "quality-aware" fusion mechanism that can selectively trust data sources based on their reliability at any given pixel and time.
Methodology
The proposed framework utilizes a Directed Probabilistic Graphical Model to represent the dependencies between observed satellite data and the underlying land cover classes.
PGM Architecture and Selective Fusion
The model decomposes the classification task into independent estimations followed by a probabilistic fusion step. First, Support Vector Machines (SVM) are used for single-date Landsat images, and Random Forests (RF) are applied to MODIS time-series data. These classifiers generate probability distributions for each pixel.
The PGM structure represents these estimations as nodes. The core innovation lies in the Conditional Probability Distributions (CPDs), which are dynamically adjusted based on data quality metrics like cloud and shadow masks. If a Landsat pixel is flagged as cloudy, its influence on the final result is automatically diminished, shifting the decision weight toward the MODIS time-series or alternative Landsat dates.
Inference and Efficiency
To determine the final land cover class, the model performs inference using a variable elimination algorithm to find the Marginal Maximum A Posteriori (MAP) estimate. By simplifying the CPDs—where many values are zero—the authors reduced the computational complexity to $O(CN)$, where $C$ is the number of classes and $N$ is the number of pixels. This efficiency allows the model to scale to large datasets without the overhead typical of complex ensemble methods.
Experimental Setup
The model was validated in two regions in China:
- Beijing (Area-1): A complex landscape featuring urban impervious surfaces, cropland, forest, and shrubland.
- Coastal Shandong (Area-2): A region dominated by cropland, forest, and waterbodies.
Data sources included multiple Landsat 8 scenes from 2013 and MODIS (MOD13Q1) 250m time-series data. Validation was performed against over 800 points interpreted from high-resolution Google Earth imagery and field surveys.
Results and Discussion
The PGM approach consistently outperformed both single-source mapping and traditional fusion techniques:
- Accuracy Improvements: In the Beijing study area, the PGM achieved an overall accuracy of 81.8%, a significant jump from the 74.0% achieved using a single Landsat image.
- Resilience to Noise: The model demonstrated exceptional performance in cloud-affected regions. In a test of 40 validation points covered by clouds or shadows, the PGM correctly classified 33, whereas the cloudy Landsat-only approach correctly identified only 19.
- The Value of Coarse Data: Even when high-resolution data was limited, the inclusion of MODIS data via the PGM improved accuracy from 74.0% to 78.4%. This confirms that coarse-resolution temporal data can effectively "fill the gaps" and provide a safety net for high-resolution snapshots.
Conclusion and Future Work
The PGM framework offers a robust, quality-aware solution for multi-source land cover mapping. By explicitly modeling the reliability of different data sources, it overcomes the pitfalls of traditional data fusion. Future research will focus on incorporating expert ecological rules directly into the PGM graph and integrating spatial context to move beyond pixel-wise analysis.
Summary updated