Enhanced YOLOv8 for UAV-Based Wildfire Smoke Detection

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Oct 16 2023

In an article recently published in the journal Sensors, researchers proposed a new approach for enhanced wildfire smoke detection based on unmanned aerial vehicles (UAV) images and You Only Look Once version 8 (YOLOv8).

Background

Forest fires are one of the deadliest and costliest natural disasters, posing a serious risk to both human properties and lives and adversely affecting the natural ecological balance of forests. Accurate and timely detection of wildfire-induced smoke is crucial for implementing forest fire fighting measures, early forest fire alert systems, and reducing fire-induced damage.

Although satellite-based monitoring provides broad coverage of forest fires, this method has several limitations, such as susceptibility to interference by cloud cover and weather, dependence on orbital cycles, insufficient spatial resolution of satellite imagery, and challenges in real-time forest fire monitoring.

Drones or UAVs are used extensively for forest fire detection due to their adaptability, precision, and rapidity. Drones can carry different sensors and cameras that can detect several spectral ranges, including infrared radiation, which enables the recognition of latent heat sources beyond the visual perception of humans.

Moreover, UAVs equipped with real-time communication systems can provide accurate information about the fire’s trajectory, positioning, and parameters, ensuring quick responsiveness by firefighters. Although several forest fire smoke detection studies have displayed the efficacy of different smoke detection models, the difficulties related to smoke feature extraction and the complex forest environment background result in multiple early detection challenges.

The proposed approach

In this paper, researchers proposed an enhanced YOLOv8 model for forest fire smoke detection using UAV images to address the existing challenges and attain improved early forest fire smoke detection accuracy in various weather conditions, including cloudy, hazy, and sunny conditions.

Cameras in UAVs capture and send images to a control station, where an AI is utilized to detect fire or smoke. Researchers in this study developed a method that employs a deep neural network to accurately obtain smoke regions' localization through a robust processor for rapid image processing in real time.

Researchers initiated the YOLOv8 model using pre-trained weights as foundational parameters for the underlying backbone network. Subsequently, network architecture parameters were adjusted to optimize the efficacy of the conventional YOLOv8 model. Integrating this fine-tuned network architecture into a forest fire smoke dataset can enable accurate recognition of smoke.

Researchers incorporated the Wise-IoU (WIoU) v3 method as a regression loss for bounding boxes, which involved using a non-monotonic, dynamic approach to create a strategy to allocate gradient gains with enhanced rationality that prioritizes samples of common quality. WIoU v3 can effectively adjust gradient gains for both low and high-quality samples, improving localization precision and overall capacity for model generalization.

Additionally, researchers introduced the BiFormer attention mechanism, which is a dynamic sparse attention design, into the backbone network to increase the computational efficiency and address the challenge of inadequately capturing the relevant forest fire smoke features within complex wooded settings.

This mechanism enables the model to emphasize crucial information within the feature map more effectively by strategically directing the attention of the model towards the forest fire smoke feature intricacies and simultaneously suppressing the influence of non-target, irrelevant background information, leading to an improvement in the model’s object detection ability.

Moreover, researchers employed the Ghost Shuffle Convolution (GSConv) mechanism to replace the conventional convolutional process within the intermediate neck layer to establish rapid pyramid pooling modules. This strategic substitution reduces model parameters and accelerates model convergence, facilitating a faster amalgamation of smoke features with a decreased computational load while processing the smoke images.

A dataset containing 6,000 images of both forest fire smoke and non-wildfire scenarios was used in this study. The dataset was split into a training dataset containing 4800 images and a testing dataset containing 1200 images. The training and testing datasets comprised 2600 and 650 smoke images, and 2200 and 550 non-smoke images, respectively.

Experimental evaluation and findings

Researchers performed comprehensive quantitative evaluations to determine the effectiveness of the proposed enhanced YOLOv8 model using documented Microsoft Common Objects in Context (COCO) benchmarks. Average precision (AP), recall, and precision were the key metrics in these evaluations.

A comparative analysis was performed between the proposed forest fire smoke detection method and several multi-stage and single-stage object detection techniques. Ablation analyses were also performed by substituting the WIoU loss module with Complete-IoU (CIoU), Distance-IoU (DIoU), and Generalized-IoU (GIoU) loss modules to evaluate the efficacy of various bounding box regression loss modules.

The improved YOLOv8 model demonstrated the highest AP, AP small (APS), AP medium (APM), and AP large (APL) of 79.4%, 71.3%, 78.5%, and 92.6%, respectively, during the comparative evaluation of the proposed model against multi-stage object detectors on the forest fire smoke dataset.

Similarly, the proposed model displayed the best performance when compared with several single-stage object detectors, including the baseline YOLOv8, using the same dataset. The proposed enhanced YOLOv8 model showed a 3.3%, 2.8%, and 5% improvement in AP, AP50, and AP75, respectively, over the baseline YOLOv8 model. Moreover, the WIuOv3 consistently displayed superior performance in the ablation studies with an AP50 of 85.1% compared to the AP50 of 84.6% and 84.5% in GIuO and DIoU, respectively.

To summarize, the findings of this study demonstrated that the enhanced YOLOv8 model can be used feasibly for more effective wildfire smoke detection.