The development of image processing algorithms has enabled the detection of forest fires through cameras and other visual data sources. Techniques such as color analysis, motion detection, and smoke pattern recognition are employed. However, these methods can be prone to false alarms due to environmental factors, like fog or dust. In response to this, numerous experts have conducted various studies. Smoke detection methods based on image processing primarily fall into two categories: traditional image processing techniques and deep learning-based image processing approaches.

Conventional image processing for smoke detection methods primarily rely on the spectral characteristics of smoke. These methods include visual interpretation, multi-threshold techniques, pattern recognition algorithms, and other similar methods. Visual interpretation employs three spectral bands of a satellite sensor, representing red, green, and blue channels, to generate true-color or false-color composite images, enabling manual visual discrimination of smoke. For instance, the true-color RGB imagery synthesized from MODIS bands 1, 4, and 3 has been used in conjunction with the false-color imagery composed of bands 7, 5, and 6 [46]. For seasoned individuals, visual interpretation serves as an effective technique for identifying smoke. However, this method has a significant drawback in that it cannot automatically process vast amounts of data. The multi-threshold method retrieves the localized optimal thresholds of reflectance or brightness temperature (BT) from established spectral bands based on historical data. These thresholds are subsequently amalgamated to eliminate cloud classes and certain ground objects, ultimately enabling the identification of smoke. For example, Li et al. [47] proposed a targeted identification approach using Himawari-8 satellite data, incorporating a connectivity domain distance weight based on multi-threshold discrimination to detect fog beneath clouds. This method exhibits high accuracy in the detection of sea and land fogs and, with limited error introduction, can effectively discern some instances of fog beneath clouds. While this approach can be effective in local areas, it poses challenges in determining the optimal threshold due to the variability of spatio-temporal information. As a result, small smoke ranges are prone to being overlooked, leading to a decrease in the promptness of fire alarms. In addition, Jang et al. [48] analyzed the variations in light scattering distributions of different colored smokes, assessing the color classification methods of smoke particles entering the smoke detectors to extract color information from the smoke, enabling the detection of fire smoke. Nevertheless, this approach overlooks the fact that certain smoke colors (such as black or gray) resemble the background environment (e.g., clouds and dust). The smoke detection method that uses a pattern recognition algorithm is an image processing technique that leverages the spectral features of smoke and typical ground objects to categorize smoke images and identify smoke pixels. Asiri et al. [49] developed a new feature space to represent visual descriptors extracted from video frames in an unsupervised manner. This mapping aims to provide better differentiation between smoke-free images and those depicting smoke patterns. This method employed training samples from a few classes, such as cloud and water, in addition to smoke. Despite its utility, the effectiveness and applicability of these smoke detection methods may be diminished when applied to diverse and intricate categories found in UAV imagery. This limitation becomes particularly evident in areas such as mountains and forests, where only a limited number of standard ground object categories are taken into account.

Most conventional image-based smoke detection algorithms utilize a pattern identification process that involves manual feature extraction and classification, where features are manually extracted and recognizers are designed. Following the extraction of candidate regions, static and dynamic smoke features are employed for smoke identification. Extracting the most crucial smoke features is challenging, and the detection process is relatively sluggish.

In recent years, the domain of deep learning has witnessed notable advancements owing to progress in hardware capabilities, the capacity to handle extensive datasets, and substantial enhancements in network architectures and training methodologies. Deep learning-based smoke detection algorithms can be classified into two-stage methods and one-stage methods. Two-stage methods include well-known representatives, such as R-CNN [28] and Faster R-CNN [26]. On the other hand, one-stage methods are exemplified by algorithms like SSD [27] and the YOLO series [29,30,31,32]. The development of these deep learning technologies has provided a solid foundation and technical support for UAV-based forest fire smoke detection.

Numerous deep learning-based techniques have been utilized to discern smoke in UAV-based scenarios. Alexandrov et al. [50] employed two one-stage detectors (SSD and YOLOv2) as well as a two-stage detector (Faster R-CNN) for smoke detection purposes. YOLOv2 outperformed Faster R-CNN, SSD, and traditional hand-crafted methods when evaluated against a large dataset of genuine and simulated images. Ghali et al. [51] introduced a novel approach based on model ensemble, combining EfficientNet and DenseNet for accurately identifying and classifying forest fire smoke with UAV-based imagery. Mukhiddinov et al. [52] proposed an early detection system for forest fire smoke using UAV imagery, employing an enhanced variant of YOLOv5. Additionally, several methods for small target detection in UAV-based settings have been proposed. Zhou et al. [53] devised a small-object detector tailored specifically for UAV-based imagery, where the YOLOv4 backbone was modified to accommodate the characteristics of small-object detection. This adaptation, combined with adjustments made to the positioning loss function, yielded improved performance in small-object localization. Jiao et al. [54] proposed a UAV aerial image forest fire detection algorithm based on YOLOv3. Initially, a UAV platform for forest fire detection was developed; subsequently, leveraging the available computational power of the onboard hardware, a scaled-down Convolutional Neural Network (CNN) was implemented utilizing YOLOv3. While these approaches demonstrate promising outcomes in object detection, they have yet to integrate real-time capabilities with high accuracy in the realm of forest fire smoke detection. Xiao et al. [55] introduced FL-YOLOv7, a lightweight model for small-target forest fire detection. By designing lightweight modules and incorporating Adaptive Spatial Feature Fusion (ASFF), they enhanced the model’s capability to detect targets of various scales and its real-time performance. However, this method did not specifically target improvements for small-scale objects but rather improved the overall accuracy of evaluation results through feature fusion. Additionally, the evaluation metrics presented in their study were limited, lacking differentiated assessment indicators for targets of varying scales.