PS 72-124
BULC: Bayesian Updating of Land-Cover classifications in a data-rich environment

Thursday, August 13, 2015
Exhibit Hall, Baltimore Convention Center
Jeffrey Cardille, McGill University, Ste. Anne de Bellevue, QC, Canada
Jacky Lee, Natural Resource Sciences, McGill University, Ste. Anne de Bellevue, QC, Canada
Julie Fortin, Natural Resource Sciences, McGill University, Ste. Anne de Bellevue, QC, Canada
Seonah Han, Natural Resource Sciences, McGill University, Ste. Anne de Bellevue, QC, Canada
Jaaved Singh, Natural Resource Sciences, McGill University, Ste. Anne de Bellevue, QC, Canada

What was the land cover of Las Vegas, Nevada on July 10, 2014? Given observations from every available sensor, where has the Brazilian cerrado been converted for agriculture since 1984?

For ecologists interested in using land-cover classifications, questions like these may be enticing, but they are remarkably difficult to answer given the current state of the art for interpreting satellite imagery. For all but highly specialized remote-sensing experts, users must either commit to months or years of diligent work gathering, fusing, and interpreting raw images, or wait for the occasional marshaling of precious resources by others to produce/update broad-purpose land-cover classifications. Yet for many would-be users of land-cover pattern information, proper maps are only one step of a multi-part analysis. Without a straightforward way to roll land-cover classifications forward or backward in time using a study-specific classification legend, we are often left with highly precise but wrongly dated land-cover data, with categories that don’t quite fit.

To address these limitations, we have developed the BULC algorithm (Bayesian Updating of Land-Cover), designed for the continuous updating of land-cover classifications through time. BULC ingests remote-sensing data from multiple satellites and sensors—including cloudy imagery— and is designed to allow non-expert users to adjust a given classification in time to any desired date. We tested this algorithm in central Quebec with Landsat 8 for summer 2013, when a series of large fires erupted and burned through forests. 


In this very dynamic setting, many of the summer’s images were cloudy and contaminated by ordinary standards, yet parts of the region could be seen through the clouds, smoke, and haze. Across 10 images of widely varying quality, BULC tracked the fires, ignored contaminated image information, and mapped fire borders consistently throughout the summer. Building on images averaging 81% accuracy, the BULC algorithm created 10 changing maps each having more than 90% overall accuracy. Despite variable-quality inputs, individual categories were mapped between 90% and near 100% accuracy. We are now porting this algorithm to Google Earth Engine, where we intend to build a general tool for non-experts that can be run without downloading any imagery. As we leave remote sensing’s data-poor era and enter a period with many views from multiple sensors of varied resolutions over a short period of time, the BULC algorithm may help to sift through images of fluctuating quality to map Earth’s surface on demand for a specific study’s needs.