Wildfires are a worldwide natural disaster causing important economic damages and loss of lives. Experts predict that wildfires will increase in the coming years mainly due to climate change. Early detection and prediction of fire spread can help reduce affected areas and improve firefighting. Numerous systems were developed to detect fire. Recently, Unmanned Aerial Vehicles were employed to tackle this problem due to their high flexibility, their low-cost, and their ability to cover wide areas during the day or night. However, they are still limited by challenging problems such as small fire size, background complexity, and image degradation. To deal with the aforementioned limitations, we adapted and optimized Deep Learning methods to detect wildfire at an early stage. A novel deep ensemble learning method, which combines EfficientNet-B5 and DenseNet-201 models, is proposed to identify and classify wildfire using aerial images. In addition, two vision transformers (TransUNet and TransFire) and a deep convolutional model (EfficientSeg) were employed to segment wildfire regions and determine the precise fire regions. The obtained results are promising and show the efficiency of using Deep Learning and vision transformers for wildfire classification and segmentation. The proposed model for wildfire classification obtained an accuracy of 85.12% and outperformed many state-of-the-art works. It proved its ability in classifying wildfire even small fire areas. The best semantic segmentation models achieved an F1-score of 99.9% for TransUNet architecture and 99.82% for TransFire architecture superior to recent published models. More specifically, we demonstrated the ability of these models to extract the finer details of wildfire using aerial images. They can further overcome current model limitations, such as background complexity and small wildfire areas.

In this paper, we address the problem of forest fires’ early detection and segmentation in order to predict their spread and help with fire fighting. Techniques based on Convolutional Networks are the most used and have proven to be efficient at solving such a problem. However, they remain limited in modeling the long-range relationship between objects in the image, due to the intrinsic locality of convolution operators. In order to overcome this drawback, Transformers, designed for sequence-to-sequence prediction, have emerged as alternative architectures. They have recently been used to determine the global dependencies between input and output sequences using the self-attention mechanism. In this context, we present in this work the very first study, which explores the potential of vision Transformers in the context of forest fire segmentation. Two vision-based Transformers are used, TransUNet and MedT. Thus, we design two frameworks based on the former image Transformers adapted to our complex, non-structured environment, which we evaluate using varying backbones and we optimize for forest fires’ segmentation. Extensive evaluations of both frameworks revealed a performance superior to current methods. The proposed approaches achieved a state-of-the-art performance with an F1-score of 97.7% for TransUNet architecture and 96.0% for MedT architecture. The analysis of the results showed that these models reduce fire pixels mis-classifications thanks to the extraction of both global and local features, which provide finer detection of the fire’s shape.

In this paper, we propose an object detection methodology applied to Global Wheat Head Detection (GWHD) Dataset. We have been through two major architectures of object detection which are Faster R-CNN, and EfficientDet, in order to design a novel and robust wheat head detection model. We emphasize on optimizing the performance of our proposed final architectures. Furthermore, we have been through an extensive exploratory data analysis, data cleaning, data splitting and adapted best data augmentation techniques to our context. We use semi supervised learning, precisely pseudo-labeling, to boost previous supervised models of object detection. Moreover, we put much effort on ensemble learning including test time augmentation, multi-scale ensemble and bootstrap aggregating to achieve higher performance. Finally, we use weighted boxes fusion as our post processing technique to optimize our wheat head detection results. Our solution has been submitted to solve a research challenge launched on the GWHD Dataset which was led by nine research institutes from seven countries. Our proposed method was ranked within the top 6% in the above-mentioned challenge.

With the widespread use of deep learning in leading systems, it has become the mainstream in the table detection field. Some tables are difficult to detect because of the likely figure layout or the small size. As a solution to the underlined problem, we propose a novel method, called DCTable, to improve Faster R-CNN for table detection. DCTable came up to extract more discriminative features using a backbone with dilated convolutions in order to improve the quality of region proposals. Another main contribution of this paper is the anchors optimization using the Intersection over Union (IoU)-balanced loss to train the RPN and reduce the false positive rate. This is followed by a RoI Align layer, instead of the ROI pooling, to improve the accuracy during mapping table proposal candidates by eliminating the coarse misalignment and introducing the bilinear interpolation in mapping region proposal candidates. Training and testing on a public dataset showed the effectiveness of the algorithm and a considerable improvement of the F1-score on ICDAR 2017-Pod, ICDAR-2019, Marmot and RVL CDIP datasets.