The process of identifying and recognizing early indicators of FLF, such as localized temperature spikes and the appearance of smoke from fire sources, through the use of visual analysis technology is commonly referred to as FLF detection. By analyzing images to assess the size and density of fire and smoke in forest and land areas, FLF detection has become a critical component of early disaster warning systems and global forest conservation. Indonesian FLF Prevention Patrol System [15] serves as a real-world example of early FLF detection implementation.

Li et al., in their study, combined AFN and DANN models to improve the detection accuracy of small-sized smoke in complex forest scenarios [8]. This method showed significant improvements in both detection accuracy and generalization ability, particularly in reducing false alarms. Such capability is critical to ensure detection systems can identify potential fires even from the early appearance of smoke before it spreads and becomes uncontrollable. This is especially crucial in tropical forests, like those in Indonesia, where vegetation is dense.

In addition, the YOLO model has become one of the most widely adopted techniques for developing real-time FLF detection systems [16]. Utilizing YOLOv5, v6, and v8 for smoke localization, Goncalves et al. [2] still encountered challenges in detecting small fires under visually disturbed conditions caused by haze and sunlight. Caixiong Li et al. [9] further explored YOLOv8’s capability in recognizing varying fire sizes, although the method required high computational resources. This can hinder FLF detection efficiency in real-time scenarios with limited computing power.

On the other hand, Venâncio et al. [5] addressed this issue by combining YOLOv4 with pruning filters, which demonstrated YOLO’s strong potential in FLF detection. However, this still needs further testing under underexplored data conditions. A clear example would be maximizing early-stage fire detection when smoke or fire visuals are faint and scattered, while simultaneously avoiding false alarms. To support effective early warning systems and FLF mitigation, developing an approach that can maintain accuracy under such conditions becomes highly crucial.

Histogram Equalization (HE) is an image enhancement technique that improves visual quality by equalizing contrast through the redistribution of pixel intensity values. It has been successfully applied in medical and forestry imaging, particularly in improving Signal-to-Noise Ratio (SNR) metrics [11,17]. However, HE has limitations in scenes with uneven lighting, where global adjustments may fail to enhance local details.

To overcome this, Adaptive Histogram Equalization (AHE) was introduced, operating locally on sub-regions of the image. Yet, AHE often excessively amplifies noise. CLAHE (Contrast Limited Adaptive Histogram Equalization) improves upon AHE by applying a clip limit to the histogram, thus preserving visual stability while enhancing local contrast [18]. CLAHE is particularly useful in handling varying illumination levels and textured backgrounds, making it more suitable for complex forest scenes compared to traditional HE or Retinex-based methods.

Recent studies have also explored hybrid techniques such as Fuzzy Contrast Enhancement (FCE) and learning-based histogram models to further improve detail visibility in low-resolution or noisy images [19]. Applications in plant disease detection have demonstrated that HE and CLAHE can improve classification accuracy by clarifying subtle patterns and textures during preprocessing [20,21].

To address dynamic scene complexity and preserve fine details, DBST-LCM (Dynamic Block Size Technique-Local Contrast Modification) was developed. This method adapts enhancement parameters by dynamically selecting block sizes based on image features and applying localized adjustments. It then performs CLAHE, followed by a feedback-driven quality check to ensure contrast clarity and sharpness in complex backgrounds [13]. Unlike conventional methods, DBST-LCM provides both adaptability and structure-aware enhancement, making it especially effective for detecting subtle smoke or fire signatures in challenging FLF conditions.

These three techniques—HE, CLAHE, and DBST-LCM—were chosen for this study due to their progressive improvements in enhancing visual cues critical for fire and smoke detection under varying illumination and environmental conditions. Their performance will be evaluated comprehensively to determine their suitability for real-time FLF detection.

YOLOv11 is the latest evolution in the YOLO (You Only Look Once) architecture series, developed to improve real-time object detection by enhancing both accuracy and computational efficiency. Key innovations in this version include the C3K2 block and the C2PSA attention module, which are designed to optimize feature extraction and attention to critical image regions. These improvements allow YOLOv11 to achieve better inference speed and precision compared to its predecessors—YOLOv8, YOLOv9, and YOLOv10—without significantly increasing model complexity. In this research, the official open-source implementation of YOLOv11 provided by Ultralytics (available at https://docs.ultralytics.com/models/yolo11/, accessed on 17 June 2024) has been adopted to ensure reproducibility and consistency with the original architecture.

Here are the key features of YOLOv11: (1) Efficient Feature Extraction with the C3K2 Block: The C3K2 block is an enhancement of the Cross Stage Partial (CSP) architecture used in previous versions. By using two small convolutions (3 × 3 kernel) instead of one large convolution, the C3K2 block maintains feature extraction performance while reducing the number of parameters and computational load [22]. (2) Improved Spatial Attention with C2PSA: C2PSA (Cross Stage Partial with Spatial Attention) is a new module that introduces a spatial attention mechanism to help the model focus more on major areas in the image, such as small objects or partially occluded objects. This improves the model’s sensitivity to spatial variations in the image [23]. (3) Multi-Scale Feature Combination through SPPF: Like previous versions, YOLOv11 retains the Spatial Pyramid Pooling Fast (SPPF) module, which combines features from various scales to enhance the detection of both small and large objects [24]. (4) CBS Blocks for Inference Stability: YOLOv11 also uses a Convolution-BatchNorm-SiLU (CBS) arrangement in the head section to ensure stable and effective data flow, supporting more accurate bounding box prediction and classification [23].

As part of the Single Shot Detector (SSD) architecture family, YOLOv11 performs object detection in a single forward pass, eliminating the need for region proposal stages found in two-stage detectors like Faster R-CNN. This makes YOLOv11 highly efficient and well-suited for real-time applications, including wildfire detection, where rapid response is critical. Combined with its enhanced modules—such as C3K2, C2PSA, and SPPF—YOLOv11 achieves robust performance in detecting fire and smoke under challenging conditions like low light, varied object scales, and partial occlusion. These improvements make it a robust and practical solution for early forest and land fire monitoring systems, particularly when deployed in real-time surveillance setups using UAVs (drones) or edge devices in high-risk areas such as peatlands or remote conservation forests.

Focused on effective early detection of fire and smoke in real-time conditions, this experiment is based on real images representing fire and smoke events from various environmental conditions. These include scenarios with only fire, only smoke, a combination of both, and negative examples without fire or smoke but with visual elements that might be misinterpreted. Based on these four scenarios, Venancio et al. [25] developed the D-Fire dataset, as outlined in Table 2. D-Fire consists of 21,527 labeled images categorized accordingly. Although the fire category contains fewer images, it has a higher annotation density, with an average of 2.52 fire objects per image. In contrast, smoke objects in the smoke and fire-and-smoke categories have an average of 1.13 annotations per image. In total, the dataset includes 26,557 bounding boxes: 14,692 labeled as fire and 11,865 as smoke, as shown in Table 3.

To support variation, the images were collected from several sources, including internet searches, legal fire simulations at the Technological Park of Belo Horizonte (Brazil), surveillance camera footage from Universidade Federal de Minas Gerais (UFMG), and Serra Verde State Park [26]. Additionally, some synthetic images were generated using montage techniques by overlaying artificial smoke onto green landscape backgrounds with photo editing software to simulate real forest conditions. Fig. 1 visually represents the diversity within the dataset, highlighting different instances of fire and smoke along with their corresponding ground-truth labels. The dataset captures a broad spectrum of scene types—such as forests, parks, and semi-urban environments—as well as variations in camera angles, smoke density, and lighting conditions.

This diversity allows for a comprehensive evaluation of image enhancement methods, especially in challenging scenarios such as diffused smoke, reduced visibility, and fluctuating lighting conditions. As a result, D-Fire is particularly well-suited for testing contrast-based techniques aimed at enhancing feature clarity in complex visual environments.

The D-Fire dataset is initially divided into two parts with an 80:20 ratio for training and testing. We then combined these two folders before randomly splitting the dataset again into a 70:20:10 ratio for training, validating, and testing. Afterward, we applied 3 image enhancement methods and categorized our dataset into 3 groups: X Dataset, Y Dataset, and Z Dataset. X Dataset consists of images that were applied to HE (Histogram Equalization), which works by equalizing the pixel intensity distribution in the image to flatten the image contrast. The Y dataset contains images that have been applied to CLAHE (Contrast Limited Adaptive HE). Improving upon HE, CLAHE divides the image into smaller blocks and limits contrast amplification to avoid excessive noise in the same areas. Finally, the Z Dataset consists of images that have been applied DBST-LCM CLAHE (Dynamic Block Size Technique-Local Contrast Modification), where, before CLAHE, a combination of noise reduction based on shift transformation (DBST) and local contrast modification (LCM) is applied. Fig. 2 shows examples of images that have undergone image enhancement with (a) HE, (b) CLAHE, and (c) DBST-LCM CLAHE.

We implemented the YOLOv11x variant for forest and land fire detection, utilizing a configuration of depth_multiple = 1.00, width_multiple = 1.50, and max_channels = 512. These values were not the result of empirical tuning but were adopted directly from the official Ultralytics YOLOv11 model configuration as the standard settings for the YOLOv11x variant. This ensures consistency with the original reference implementation and benchmark results. The model processes 640 × 640 × 3 input images through a deep convolutional pipeline. It features down-sampling stages via alternating Conv and C3 modules, which progressively reduce spatial resolution while increasing feature depth. The SPPF block aggregates multi-scale context before passing features to the neck, where they are unsampled and concatenated to enhance semantic richness. Final detection heads operate at resolutions of 80 × 80, 40 × 40, and 20 × 20 to detect small, medium, and large-scale objects, respectively. The architecture’s combination of C3K2, SPPF, and C2PSA modules improves feature extraction efficiency and attention precision, capabilities particularly valuable for detecting subtle visual cues such as smoke. Fig. 3 illustrates the research workflow that is based on YOLOv11 architecture with customized depth, width, and max channels [22].

Model training was conducted on Google Collaboratory using an NVIDIA A100-SXM4-40GB GPU. The training process spanned 100 epochs with a batch size of 6 and an input resolution of 640 × 640 pixels. The selection of these hyperparameters was based on hardware availability and training stability considerations, in line with practices found in related studies [25].

Table 4 shows the results of training the YOLOv11x model on the D-Fire dataset, which has been processed using three different image enhancement techniques: Histogram Equalization (HE), CLAHE, and DBST-LCM CLAHE. During the training process, X Dataset (HE) achieved the best performance with an mAP50 score of 0.771, followed by Z Dataset (DBST-LCM CLAHE) with 0.770, and Y Dataset (CLAHE) with 0.759. These results provide an initial indication that the HE method offers strong support in the object detection model training process, even when compared to more complex methods like DBST-LCM CLAHE.

The evaluation of detection results is based on the Intersection over Union (IoU) value, which calculates the overlap ratio between the predicted box and the ground truth box. The formula can be seen in Eq. (1): (1)IoU=Areapred∩AreagtAreapred∪Areagt

The prediction outputs are then classified into True Positive (TP), False Positive (FP), and False Negative (FN). This classification serves as the basis for calculating Precision and Recall, as shown in Eqs. (2) and (3) [27]. Another evaluation indicator is the F1 Score, which is the harmonic mean of Precision and Recall, as shown in Eq. (4). (2)Precision(P)=TPTP+FP (3)Recall(R)=TPTP+FN

Based on the training results on the X, Y, and Z datasets, the model showed consistent detection performance, especially for the Smoke class. The highest mAP50 score for this class was achieved with the Z Dataset, reaching 0.846. In terms of Precision and Recall, the Smoke class in the Z Dataset recorded the highest Precision (0.839), while the highest Recall (0.791) was obtained from the X Dataset. On the other hand, for the Fire class, the highest mAP50 score was achieved with X Dataset (0.703), with the highest Precision (0.741) recorded in Z Dataset and the highest Recall (0.615) in X Dataset. Another evaluation index, F1, is shown in Eq. (4). (4)F1=2×Precision×RecallPrecision+Recall

The training processes for each enhanced dataset are visualized in Fig. 4 (X Dataset for HE), Fig. 5 (Y Dataset for CLAHE), and Fig. 6 (Z Dataset for DBST-LCM CLAHE). Across more than 30 epochs, the loss values for all datasets consistently declined, indicating successful convergence of the model parameters. Throughout training, the key components of the YOLOv11x loss function—namely, box loss (localization error), classification loss, and distribution focal loss (bounding box refinement)—exhibited a steady downward trend, reflecting stable and effective optimization during training. These loss components correspond to the model’s efforts in improving bounding box localization, object confidence prediction, and class classification performance, which are essential for accurate fire and smoke detection in various environmental conditions.

While the overall trends were similar, minor differences in the loss values were observed across the three datasets. The model trained with the Z Dataset showed slightly lower overall loss values, particularly in the validation box and classification losses, suggesting better optimization. The Y Dataset achieved moderate loss reductions, whereas the X Dataset maintained slightly higher losses throughout training. These differences highlight the potential benefits of advanced image enhancement techniques in improving model performance. (5)ℒ=λcoord∑i=0S2∑j=0Bıijobj[(xi−x^i)2+(yi−y^i)2]+λcoord∑i=0S2∑j=0Bıijobj[(wi−w^i)2+(hi−h^i)2]+∑i=0S2∑j=0Bıijobj(Ci−C^i)2+λnoobj∑i=0S2∑j=0Bıijnoobj(Ci−C^i)2+∑i=0S2ıijobj∑cεclasses(pi(c)−p^i(c))2

As described in Eq. (5), the YOLOv11x model adopts a loss (ℒ) function structure based on the basic YOLO architecture. This loss function minimizes errors in predicting bounding box parameters. (x,y,w,h), object confidence C, and class probabilities p(c) [27]. Here, the symbols x^,y^^,w^,h^,C^,p^ represent the predicted values, while the corresponding ground truth labels are denoted without a hat. The indicator ıijobj identifies whether an object exists within a cell i and whether the j-th bounding box predictor is responsible for that detection. In the training configuration, the hyperparameter λcoord was set to 0.5, assigning moderate weight to localization errors. Similarly, λnoobj was set to 0.5 to reduce the impact of prediction in grid cells without objects, preventing distraction from large background regions.

YOLOv11x introduces significant advancements in real-time object detection, emphasizing both efficiency and precision. Its architecture integrates modules such as the C3K2 block for lightweight feature extraction and the C2PSA module for enhanced spatial attention. During training, YOLOv11x employed a combination of data augmentation techniques, including HSV augmentation, random rotation, translation, perspective transformation, scaling, and both vertical and horizontal flipping. Additionally, advanced augmentations such as Mosaic and MixUp were applied to improve generalization across diverse environmental conditions, with optional use of CutMix further enhancing training robustness.