Operational pixel-wise active-fire detection over central Kalimantan from Himawari-8 Band-7 with FIRMS-aligned supervision

Geostationary satellites enable near-real-time wildfire monitoring, yet rigorous multi-month evaluations at tropical sites remain scarce. We develop and assess an end-to-end, pixel-wise active-fire detector for Central Kalimantan using Himawari-8 Band-7 (≈3.9 µm) imagery at 10-minute cadence, with supervision from NASA FIRMS (MODIS C6.1; VIIRS C2). Polar detections are temporally aligned within ±5 minutes of AHI timestamps and rasterized to a fixed 256×256 grid with a light 3×3 dilation to accommodate geolocation tolerance. The dataset spans August–October 2019 and is split with time-aware partitions. We benchmark a baseline U-Net against an Attention U-Net that gates skip connections to suppress background clutter while emphasizing thermally active regions. On the pixel-wise test set (75 timestamps), the best operating point—selected once on validation and transferred to test—yields ROC-AUC = 0.9223, PR-AUC = 0.3601, and F1 = 0.4090, improving upon the conventional τ = 0.5 setting (F1 = 0.3289) by re-balancing precision–recall under class imbalance. Head-to-head, the Attention U-Net attains higher average precision than U-Net (0.360 vs 0.324) with marginal ROC differences, indicating better retrieval of small or fragmented hotspots without inflating false alarms. Beyond the model, we contribute a reproducible dataset-construction pipeline and a transparent evaluation protocol emphasizing operational realism for geostationary fire surveillance in Indonesia.