Combining Multiple-source Data for Continuous Monitoring Across Large Areas with BULC (Bayesian Updating of Land Cover)

Abstract:

The proliferation of remote-sensing platforms and data has created a pressing need for a general method for fusing large amounts of data to create multi-category time series tailored to particular needs, in a region of choice over a specified time. To create coherent time series built from arbitrary sets of sensors, we developed the general-purpose Bayesian Updating of Land Cover (BULC) algorithm. BULC ingests classified images and computes, for each pixel, a constantly updating vector of probabilities of each potential class, modifying them in light of new evidence from a new provisional classification from any data source. Land use/land cover applications are shown here, but BULC is suitable for any landscape partitioned into any number or type of category. Here we detail one BULC application: tracking forest loss for the establishment of agriculture over the entire satellite record in a 12,000-km² area in Mato Grosso, Brazil. In Google Earth Engine, we roughly classified 140 images from 14 sensors of different spatial resolution and spectral characteristics—all Landsats, Sentinels, ASTER, CBERS, etc. BULC fused the classifications into a coherent five-decade time series, revealing the timing and complex spatial patterns as ~10% of its area was converted for Agriculture across five decades. The BULC time series was more reliable than single-day classifications, while estimating land use across the entire area at all time steps even when some individual classifications covered only a small portion. Beyond these results, we survey other projects to illustrate how BULC may be used. These include sharpening a coarse global classification to finer Landsat scale, and tracing fire growth across British Columbia with three sensors. In an era of almost unlimited free satellite data, BULC can be a useful, general approach to merging interpretations from multiple platforms that leverages each platform’s strengths to produce time series with user-specified categories.