BULC: The Foundation of Multi-Sensor Data Fusion

Picture
Figure 5. Events and BULC classifications for 11 sequential Landsat 8 image dates between summer 2013 and summer 2014. Area shown is located in sector D1 of Fig. 2, Fig. 4. At each time step, each BULC classification represented an estimate of land cover given all the available imagery up to that point in time. These classifications showed a consistent picture of the study area, without noisy oscillation of the estimated class for the great majority of pixels. (Cardille, J.A., and Fortin, J.A., 2016)
We developed the foundational Bayesian Updating of Land Cover (BULC) algorithm as a general-purpose, sensor-independent data fusion tool. Unlike traditional methods that concentrate on a single sensor's time series, BULC uses Bayes' Theorem to synthesize "rough" classifications from any source—including optical satellites, radar, and even hand-drawn maps—into a single, high-quality, low-noise time series. Its probabilistic structure allows us to update the estimated status of each pixel as new imagery arrives without needing to store ever-lengthening archives of raw image values, keeping computational demands linear. This flexibility enables us to treat the entire global satellite constellation as a single machine, merging data from disparate platforms regardless of their varying spatial or spectral specifications.

Relevant Papers:
  • Cardille, J.A., and Fortin, J.A. (2016). Bayesian updating of land-cover estimates in a data-rich environment. Remote Sensing of Environment, 186, 234–249. (Link)
  • Crowley, M.A., Cardille, J.A., White, J.C., and Wulder, M.A. (2019). Multi-sensor, multi-scale, Bayesian data synthesis for mapping within-year wildfire progression. Remote Sensing Letters, 10(3), 302-311.
  • Crowley, M.A., Cardille, J.A., White, J.C., and Wulder, M.A. (2019). Generating intra-year metrics of wildfire progression using multiple open-access satellite data streams. Remote Sensing of Environment, 232, 111295.

BULC-D: High-Sensitivity Disturbance Tracking

Picture
Figure 4. For each area of interest, the burned pixels based on seasonality as identified by different methods, including a summer 2023 middle-infrared Sentinel-2 composite (RGB [B12, B8, B04]) of the area (4th row). (Pelletier, F., 2024).
To address the need for high-cadence hazard monitoring, we devised BULC-D, a high-sensitivity variant designed to detect stand-replacing disturbances like fire and harvest. This logic, which also powers our TIIC (Tracking Intra- and Inter-year Change) framework, employs binned Z-scores and harmonic curve fitting to identify significant departures from a pixel's expected seasonal baseline. By accumulating evidence over multiple intra-annual images, it can confirm change events within weeks, drastically reducing the latency typical of annual trend-based methods. We have successfully applied this to map millions of hectares of record-breaking fire seasons across Canada while simultaneously tracking the resulting aboveground biomass (AGB) dynamics.

Relevant Papers:
  • Cardille, J.A., Perez, E., Crowley, M.A., Wulder, M.A., White, J.C., and Hermosilla, T. (2022). Multi-sensor change detection for within-year capture and labelling of forest disturbance. Remote Sensing of Environment, 268, 112741.
  • Pelletier, F., Cardille, J.A., Wulder, M.A., White, J.C., and Hermosilla, T. (2024). Inter- and intra-year forest change detection and monitoring of aboveground biomass dynamics using Sentinel-2 and Landsat. Remote Sensing of Environment, 301, 113931.
  • Pelletier, F., Cardille, J.A., Wulder, M.A., White, J.C., and Hermosilla, T. (2024). Revisiting the 2023 wildfire season in Canada. Science of Remote Sensing, 10, 100145. (Link)

BULC-U : Unsupervised

Picture
Graphical abstract from Xiong, S et al., 2022.
Picture
Figure 5. An example of the workflow output for four time points (2015, 2010, 2005, and 2000). The fourth row depicts resulting cropland probabilities from the BULC-U algorithm for each time point, alongside Google image, Landsat composites and unsupervised classification with segmentation. The white locator box highlights high probability cropland that appears in 2010 and disappears in 2015, indicating a probable loss of cropland area. (Xiong, S et al., 2022).
We created the BULC-U (Unsupervised) flavor to overcome the chronic shortage of labeled training data required for large-scale mapping. This variant accepts unlabeled classifications from an unsupervised classifier and evaluates their most likely meaning by comparing them to a known, potentially coarse ''seed'' basemap, such as the 300-m GlobCover product. By iteratively updating probabilities backwards in time, BULC-U can sharpen spatial resolutions (refining 300-m data to 30-m Landsat detail) and reconstruct historical land-cover records with minimal manual intervention. We have utilized this approach to quantify decades of agricultural expansion and forest fragmentation in Mato Grosso, Brazil, and to track annual cropland changes across the complex landscapes of Zambia.
​
Relevant Papers:
  • Lee, J., Cardille, J.A., and Coe, M.T. (2018). BULC-U: sharpening resolution and improving accuracy of land-use/land-cover classifications in Google Earth Engine. Remote Sensing, 10(9), 1455.
  • Lee, J., Cardille, J.A., and Coe, M.T. (2020). Agricultural expansion in Mato Grosso from 1986–2000: A Bayesian time series approach to tracking past land cover change. Remote Sensing, 12(4), 688.
  • Xiong, S., Baltezar, P., Crowley, M.A., Cecil, M., Crema, S.C., Baldwin, E., Cardille, J.A., and Estes, L. (2022). Probabilistic tracking of annual cropland changes over large, complex agricultural landscapes using Google Earth Engine. Remote Sensing, 14, 4896. (Applied a modified BULC-U workflow, but also in SAR section because this paper also combined SAR and optical sensors). (Link)