Published November 18, 2026 | Version v1

Do Wind-Directional Fire Features Improve Haze-Visibility Forecasts? A Multi-Season Evaluation in Borneo and Sumatra

Description

Satellite fire detections, weighted by wind direction, are commonly used in operational smoke prediction but are rarely evaluated on their own merit. Using public METAR archives, we created a dataset of twelve airports on Sumatra and Borneo where visibility was limited to less than 3 km due to hazy or smoky conditions (henceforth haze-limited visibility events). Then we investigated whether including wind-directional features of satellite fire detections (from the VIIRS active fire product) could improve 6 and 24 hour forecasts of haze-limited visibility events. To that end we built an eleven-year dataset (2015-2025) and used leave-one-year-out cross-validation to assess model skill. As controls, we used: time-shifted wind directions; randomised (permuted) wind directions; categories of wind reporting; and a model without wind-directional features (matched for missing data). Uncertainty was estimated with a block bootstrap at station and regional level. In single model runs, the directional model appeared to improve 24-hour forecasts substantially (+0.057 average precision), but this was driven by a lot of seed-to-seed variation in one extreme season, where different runs of the same model differ by almost 0.1 average precision. When we did five-seed ensembles, the directional model was only a few thousandths of average precision better than the permuted-wind placebo (+0.004) and the controls, with no consistently larger effects with vegetation fires only. The only noticeable benefit was that the fire count predictors helped 24-hour forecasts a little bit. Overall, we recommend seed ensembles and mask-preserving placebos for studies that claim directional wind-aware fire features are helpful.

Files

Ananda-2026-directional-fire-preprint.pdf

Files (321.7 kB)

Name Size
md5:5ae7ee6f5884bc84e77511d5e6648b05
321.7 kB Preview Download

Additional details

Software