Fire Detection

Active fire and hotspot detection based on the MODIS Collection 6.1 algorithm (Giglio et al., 2003; Giglio et al., 2016). This is a simplified implementation using 4 of the ~12 tests in the full algorithm.

Input

Two co-registered brightness temperature bands stored as flat &[f32] arrays (row-major):

Band	Wavelength	MODIS Equivalent	Role
T4 (MIR)	~3.9 um	Band 21/22	Fire-sensitive; strong emission from active fires
T11 (TIR)	~11 um	Band 31	Background temperature reference

Detection Pipeline

Each pixel passes through four sequential tests. A pixel must pass all tests to be flagged as a hotspot.

Test 1: Absolute MIR Threshold

Reject if T4 < t4_abs. Eliminates all cool pixels immediately.

Test 2: Absolute Split-Window Threshold

Reject if (T4 - T11) < t4_t11_abs. Fire pixels have a large MIR-TIR difference because fire radiates strongly at 3.9 um but not proportionally at 11 um.

Test 3: Contextual MIR Anomaly

Compute background statistics from a window of radius bg_radius around the pixel (excluding the pixel itself and any pixel with T4 > bg_max_t4). Reject if the MIR anomaly above the background mean is not greater than max(dt4_threshold, 3 * std_T4).

Test 4: Contextual Split-Window Anomaly

Using the same background window, reject if the split-window anomaly (T4 - T11) above the background mean is not greater than max(dt4_t11_threshold, 3 * std_dT).

If fewer than min_bg_count valid background pixels are available, tests 3 and 4 are skipped and only the absolute thresholds apply, with confidence reduced to 0.3.

Thresholds

Default values follow MODIS Collection 6.1 (Giglio et al., 2016, Table 2).

Parameter	Day	Night	Unit
t4_abs	310	305	K
dt4_threshold	10	10	K
dt4_t11_threshold	6	6	K
t4_t11_abs	10	10	K
min_bg_count	8	8	pixels
bg_radius	5	5	pixels
bg_max_t4	325	320	K

Confidence Scoring

Confidence is the average of two terms, each clamped to [0, 1]:

$\text{confidence} = \frac{\min(\Delta T4 / 50, 1) + \min(\Delta (T4{-}T11) / 30, 1)}{2}$

Higher anomalies above the background produce higher confidence. When the contextual test is skipped (insufficient background pixels), confidence is fixed at 0.3.

FRP Estimation

Fire Radiative Power is estimated from brightness temperature using a simplified Wooster et al. (2003) approach:

$\text{FRP} = \max\big((\Delta T4)^2 \times 10^{-6},\; 0\big) \quad \text{(MW)}$

where $\Delta T4 = T4_{\text{fire}} - T4_{\text{background}}$. This is a proxy; accurate FRP requires the actual MIR spectral radiance (W/m^2/sr/um).

Output

detect_fire returns a count and fills a caller-provided &mut [Hotspot] buffer. Each Hotspot contains pixel coordinates, T4, T11, anomaly values, FRP, and confidence.