Smoke characterization for incipient wildfire detection from Sentinel-2 sensor based on sigmoid activation function and momentum gradient descent optimizer

Abstract Wildfires are considered a natural factor which leaves detrimental effects on the environment. In this study, the occurrence of wildfire smoke coincided with the occurrence of clouds, and this underscored the need to separate the wildfire smoke from the clouds. The sigmoid activation function, coupled with momentum gradient optimizer (MGD) optimizer, was applied to spectrally reconfigure selected Sentinel-2 bands to smoke plumes. Bartlett’s k-comparison of equal variance statistical was applied to determine spectral radiance properties of smoke plumes and clouds across selected Sentinel-2 bands. The Relative Operation Characteristics (ROC) was used to evaluate the performance of the performance of the sigmoid activation function with MGD in characterizing smoke plumes. Bartlett’s test revealed variations in the radiance properties of smoke and clouds across the selected spectral bands of Sentinel-2 sensor, with the p-value of < 0.001 for both smoke and clouds. The mean radiance values for smoke plume were noted to be lower than that of the clouds across all the selected spectral channels besides the shortwave infrared (SWIR) cirrus channel for both original and calibrated image, where smoke and clouds had similar radiance properties. The relative operation characteristics (ROC) results confirmed the calibrated blue and green spectral bands to be effective in detecting smoke plume, with area under curve (AUC) value of 0.81 and 0.73 respectively. This research emphasized the significance of integrating machine learning and multispectral remote sensing in mitigating wildfire disaster. Because wildfire is an unpredictable incident, the findings of this study were not validated with ground-based data.