Introduction
Generating a high-resolution, seamless global map from satellite imagery is a monumental task. For the Global Land Survey 2005 (GLS-2005), researchers from NASA and the USGS faced the challenge of selecting the best possible images for approximately 9,500 locations across the Earth's landmass. With a pool of over 300,000 candidate images from Landsat 5 and Landsat 7, the selection process is akin to solving a massive jigsaw puzzle where each piece must meet strict quality and consistency criteria.
The Complexity of Image Selection
The selection process is not merely about finding the clearest picture for a single spot. It involves balancing several conflicting objectives:
- Image Quality: Minimizing cloud cover is paramount, but researchers also aim to maximize vegetation health, measured by the Normalized Difference Vegetation Index (NDVI).
- Temporal Consistency: To avoid a jarring "patchwork" appearance, neighboring images should ideally be captured during the same season.
- Sensor Compensation: A 2003 malfunction in the Landsat 7 Enhanced Thematic Mapper Plus (ETM+) sensor resulted in data gaps (Scan Line Corrector-off). To produce a complete image, two separate scenes of the same location must be "composited." This requires finding a pair of images that are "out of phase" so their data gaps do not overlap.
Historically, this selection was done manually, a labor-intensive process that often led to sub-optimal results due to the sheer volume of data.
Methodology: The GMG-COP Framework
To automate this, the authors formulated the task as a Global Map Generation Constraint Optimization Problem (GMG-COP). They defined an objective function based on a weighted sum of 12 criteria, including cloud cover, NDVI, seasonality, and gap-phase statistics for Landsat 7 composites.
Because the complexity of this problem makes exact solvers computationally infeasible, the team employed a multi-start local search algorithm. The search begins with an initial "seed" selection. In each iteration, the algorithm randomly selects a scene location, evaluates all possible candidates for that spot, and chooses the one that most improves the global score.
Interestingly, the researchers found that starting with a random selection of images often yielded better results than a "greedy" start (picking the best individual image for each spot first), as the latter tended to trap the algorithm in local optima. To prevent cycling, they implemented a "taboo" list to ensure every scene is examined before any is revisited in a single iteration.
Human-in-the-Loop: The LASSI Interface
While automation provides efficiency, human expertise remains vital. Automated metadata for cloud cover can sometimes be inaccurate, missing small "popcorn" clouds that degrade image quality. To address this, the solver was integrated into the Landsat Scene Selector Interface (LASSI).
LASSI allows experts to visualize the solver's choices and manually override them if necessary. The interface displays candidate images, their metadata, and how they contribute to the overall mosaic. This is particularly useful for specific land types, such as agricultural areas, where experts might prefer Landsat 5 imagery over Landsat 7 composites to avoid artifacts that could interfere with crop analysis.
Key Results and Impact
The automated approach demonstrated significant improvements over previous manual efforts. When benchmarked against the GeoCover-2000 dataset, the GMG solver produced solutions with a 23% higher objective function score.
Beyond quality, the efficiency gains were dramatic. The solver can converge on a global solution for all 9,500 locations in approximately one minute. This system was successfully deployed to complete the North American portion of the GLS-2005 and was subsequently used for the rest of the globe.
Conclusion
The integration of local search optimization into the Landsat selection process has transformed global map generation. By treating map assembly as a constraint optimization problem, researchers can produce higher-quality datasets in a fraction of the time required for manual selection. Future work involves applying this framework to the GLS-2010 survey and potentially extending it to planetary mapping, such as generating lunar maps from Clementine mission data.
Summary updated