Active Fire Detection from Landsat-8 Imagery Using Deep Multiple Kernel Learning

Active fires are devastating natural disasters that cause socio-economical damage across the globe. The detection and mapping of these require efficient tools, scientific methods, reliable observations. Satellite images have been widely used for active fire (AFD) during past years due to their nearly global coverage. However, accurate AFD in satellite imagery is still a challenging task remote sensing community, which mainly uses traditional methods. Deep learning (DL) methods recently yielded outstanding results applications. Nevertheless, less attention has given them imagery. This study presented deep convolutional neural network (CNN) “MultiScale-Net” Landsat-8 datasets at pixel level. proposed had two main characteristics: (1) several convolution kernels with multiple sizes, (2) dilated layers (DCLs) various dilation rates. Moreover, this paper suggested an innovative Fire Index (AFI) AFD. AFI was added inputs consisting SWIR2, SWIR1, Blue bands improve performance MultiScale-Net. In ablation analysis, three different scenarios were designed multi-size kernels, rates, input variables individually, resulting 27 distinct models. quantitative indicated model AFI-SWIR2-SWIR1-Blue as variables, using sizes 3 × 3, 5 5, 7 simultaneously, rate 2, achieved highest F1-score IoU 91.62% 84.54%, respectively. Stacking led fewer false negative (FN) pixels. Furthermore, our qualitative assessment revealed models could detect single pixels detached from large zones by taking advantage kernels. Overall, MultiScale-Net met expectations detecting varying shapes over test samples.