Harkat, HoudaNascimento, JoseBernardino, Alexandre2020-11-232020-11-232020-09-20HARKAT, H.; NASCIMENTO, José; BERNARDINO, A. – Fire segmentation using a Deeplabv3+ architecture. In Proc. SPIE 11533, Image and Signal Processing for Remote Sensing XXVI. United Kingdom: SPIE, 2020. ISBN 978-151-063-879-2. Pp. 115330M-1-115330M-12978-151-063-879-2http://hdl.handle.net/10400.21/12391In the last decade, the number of forest _res events is growing due to the fast change of earth's climate. Hence, more automatized _re _ghting actions had become necessary. Deep learning had drawn interesting results for pixel level classi_cation for smoke detection, but few systems are proposed for _re ame detection. In this paper, a semantic segmentation approach using Deeplabv3+ architecture for wild_re detection is proposed. The network uses Deeplabv3 architecture as encoder and Atrous Spatial Pyramid Pooling (ASPP) which allows to encode multi-scale information and boost the network performance. In fact, the ASPP block concatenates a stack of parallel Atrous convolutions with graduating rates, which produces a multi-scale feature map that is further resized. The tests were performed on a public dataset, Corsican _re dataset, which contains 1135 RGB images and 640 infrared pictures. The experiments were conducted on two customized datasets, one using the whole dataset within a single channel information (red and infra-red), and another using only the RGB image set that contains information coded in 3 channels. The used dataset is unbalanced, which could induce high precision with very low sensitivity. Therefore, to measure the performance, Dice similarity and Tversky loss functions with cross-entropy are adopted. The capability of Deeplabv3+ was tested with two di_erent backbones, ResNet-18 and ResNet-50, and compared to a very simple Convolutional Neural Network (CNN) architecture with dilated convolution. Four di_erent metrics were used to evaluate the segmentation capability: Accuracy, mean Intersection over Union (IoU), Mean Boundary F1 (BF) Score, and Mean Dice coe_cient. The experimental results demonstrate that the Deeplabv3+ with ResNet-50 backbone and a loss function type Dice or Tversky can accurately detect the _re ame. The given results are very encouraging for further study using deep learning approaches.engWildfireRGB imagesSegmentationDilated convolutionAtrous Spatial Pyramid Pooling (ASPP)conference object10.1117/12.2573902