TinyFireNet: a lightweight hybrid CNN with physics-constrained spectral gate for on-board wildfire detection in multispectral satellite imagery

Accurate and timely detection of active wildfires from spaceborne multispectral sensors is impeded by high false-positive (FP) rates on industrially developed land, where elevated shortwave-infrared (SWIR) reflectance and thermal emission from heated structures mimic spectral fire signatures. This paper presents TinyFireNet – a lightweight seven-channel convolutional neural network with 407,106 parameters (1.63 MB float32, 0.41 MB INT8) designed for patch-level three-class classification (background, active fire, burn scar), aggregated into per-pixel maps through overlapping sliding-window inference. The network is coupled with a four-stage Physics Gate, a physics-constrained sequential spectral filter based on the Normalised Difference Water Index (NDWI), the Normalised Difference Built-up Index (NDBI), thermal infrared (TIR) brightness temperature, and the Normalised Burn Ratio (NBR). We identify and resolve a previously undescribed Feature Clash in which fresh burn scars are erroneously suppressed by the built-up area mask due to shared shortwave-infrared characteristics; the proposed Conditional Masking augments the NDBI criterion with a NIR reflectance threshold at 0.12 and a TIR arbiter at 312 K. The model is evaluated on 3,000 held-out multispectral patches (Landsat-8/9, seven spectral channels) drawn from 55 wildfire events across six continents. Ground-truth labels were independently cross-validated against ESA FireCCI51 v5.1 and NASA MCD64A1 burned-area products, yielding 94.3 % pixel-level agreement on 18 overlapping scenes – confirming that the Physics-Gate-derived labels are congruent with established baselines and not subject to circular validation. TinyFireNet achieves an Overall Accuracy of 99.6 %, an F1-score of 0.914 for active fire and 0.935 for burn scar, with mean Intersection-over-Union of 0.904, and outperforms a threshold-only NBR baseline (Precision = 0.48) while approaching the F1 performance of a 5.9 M-parameter trimmed U-Net at one-fifteenth the model footprint. Measured inference latency on an ARM Cortex-A53 processor (1.4 GHz, batch size 1) is 18.3±0.6 ms per 64×64 patch in float32 and 1.5±0.1 ms with NEON vectorisation; post-training INT8 quantisation introduces a F1 drop of less than 0.01, confirming suitability for on-board deployment within the KazEOSat Earth observation programme.