FlameTransNet: Advancing Forest Flame Segmentation with Fusion and Augmentation Techniques

Abstract

Forest fires pose severe risks, including habitat loss and air pollution. Accurate forest flame segmentation is vital for effective fire management and protection of ecosystems. It improves detection, response, and understanding of fire behavior. Due to the easy accessibility and rich information content of forest remote sensing images, remote sensing techniques are frequently applied in forest flame segmentation. With the advancement of deep learning, convolutional neural network (CNN) techniques have been widely adopted for forest flame segmentation and have achieved remarkable results. However, forest remote sensing images often have high resolutions, and relative to the entire image, forest flame regions are relatively small, resulting in class imbalance issues. Additionally, mainstream semantic segmentation methods are limited by the receptive field of CNNs, making it challenging to effectively extract global features from the images and leading to poor segmentation performance when relying solely on labeled datasets. To address these issues, we propose a method based on the deeplabV3+ model, incorporating the following design strategies: (1) an adaptive Copy-Paste data augmentation method is introduced to learn from challenging samples (Images that cannot be adequately learned due to class imbalance and other factors) effectively, (2) transformer modules are concatenated and parallelly integrated into the encoder, while a CBAM attention mechanism is added to the decoder to fully extract image features, and (3) a dice loss is introduced to mitigate the class imbalance problem. By conducting validation on our self-constructed dataset, our approach has demonstrated superior performance across multiple metrics compared to current state-of-the-art semantic segmentation methods. Specifically, in terms of IoU (Intersection over Union), Precision, and Recall metrics for the flame category, our method has exhibited notable enhancements of 4.09%, 3.48%, and 1.49%, respectively, when compared to the best-performing UNet model. Moreover, our approach has achieved advancements of 11.03%, 9.10%, and 4.77% in the same aforementioned metrics as compared to the baseline model.