A low-cost UAV-edge-cloud system for agricultural straw burning monitoring: multi-target detection and event-based alerting

1 Introduction

Crop residue burning, particularly straw burning in agricultural fields, has become a critical environmental concern worldwide. Open burning of straw will emit fine particulate matter, greenhouse gases, aerosols and other atmospheric pollutants. These pollutants will aggravate air pollution and pose serious threats to human health (Lv et al., 2024; Huang et al., 2025). In addition, burning crop residues will also damage the soil ecology, leading to the decline of farmland fertility and agricultural production capacity over time (Khangar et al., 2025; Parambil-Peedika et al., 2025). To this end, the Chinese government has introduced strict policies to curb open straw burning. But the traditional manual inspection mode is inefficient for normalized management of large-scale farmland. For example, covert straw burning has increased at nighttime to evade daytime surveillance in China (Wang and Wang, 2025). It can be seen that an effective agricultural straw burning system must be able to capture ongoing fires and identify traces of completed burning activities.

Remote sensing technology has been used to study wildfire monitoring. Satellite remote sensing platforms such as MODIS use thermal infrared bands to retrieve the distribution of active fires (Li et al., 2023). However, limited by the spatial and temporal resolution of satellites, it is difficult to capture small-scale or short-lasting straw burning events. The unmanned aerial vehicle (UAV) is flexible and can be equipped with high-definition optical sensors and thermal imaging sensors. It is very suitable for monitoring straw burning scenes (Panicker and Devidas, 2025; Chao et al., 2025).

There are limited studies on straw burning scenarios (Wen et al., 2024). Most of the relevant research focuses on forest fire and wildfire (Zhu et al., 2025; Bai et al., 2026; Ma et al., 2026). Straw burning has a distinct visual representation. In the early stage of fire, the open flame is usually very small. When the flame is weak or smoldering, smoke is the main fire indicator. After the fire is extinguished, burned residue remains in the field. If monitoring is solely based on flame, the effective monitoring time window will be very short (Zhang et al., 2026).

UAV-based visual monitoring can be formulated at different levels of spatial detail. Image-classification methods predict the presence of fire visual cues for an entire image. Such methods are computationally efficient but do not identify the number or image-plane locations of individual targets. Object-detection methods provide bounding boxes with category labels for individual instances. This paradigm retains the spatial information of each combustion feature at the object level. When combined with calibrated camera intrinsics, synchronized UAV position and attitude, and geometric projection, the spatial outputs of detection models may support more precise fire source localization through triangulation (Choutri et al., 2023).

Recent UAV-based fire detection studies have primarily focused on improving visual recognition. Representative developments include lightweight UAV smoke detection (Saydirasulovich et al., 2023), RGB-thermal fusion for straw burning recognition (Wen et al., 2024; Hu et al., 2025), and multiscale detection in complex forest environments (Zhu et al., 2025). These studies have advanced detection under small-target or multimodal imaging conditions, but most emphasize detection evaluation rather than the complete fire monitoring workflow.

Edge deployment studies are limited by the computing power and storage capacity of embedded edge devices on UAVs (Zhang et al., 2022; Dong et al., 2025). The NVIDIA Jetson hardware platform has strong inference performance, so it is widely used as a prevailing solution. However, its cost and power consumption are relatively high (Mu et al., 2025). Raspberry Pi hardware is an economical selection. Titu et al. (2024) deployed a distilled YOLOv8n model on a Raspberry Pi 5 and achieved an average processing rate of 8 FPS during field experiments. This result demonstrates the feasibility of CPU-based onboard inference on low-cost hardware. However, general-purpose CPUs may remain constrained for high throughput of UAV monitoring tasks. These studies demonstrate the feasibility of onboard inference, but they differ substantially in fire scene and hardware platform. More importantly, model evaluation or detector deployment alone does not establish a complete operational workflow for agricultural straw burning monitoring. Such a workflow should combine fire visual evidence, onboard inference and alert reporting. Whether this complete chain can operate on low-cost edge hardware with practical processing continuity remains insufficiently examined.

To address this systems gap, we developed a low-cost UAV-edge-cloud framework that integrates image acquisition, multi-target detection, onboard INT8 inference, UAV-position association, local alarms, and cloud reporting. The detector recognizes flame, smoke, and burned residue to provide complementary visual evidence across multiple combustion stages. The INT8 model runs on the dedicated NPU of the low-cost MaixCAM-Pro. The ESP32 manages GNSS-linked event transmission and local warning. Field experiments were conducted to evaluate the workflow performance of our system. The principal originality of the study lies in the field-validated integration of a complete monitoring and alert workflow on a low-cost payload. The main contributions of this study are summarized as follows.

• This straw burning monitoring system supports multi-target detection, including flame, smoke, and burned residue. Flame indicates active combustion. Smoke can reveal early-stage or smoldering combustion. Burned residue remains visible after the flame and smoke has weakened. This multi-category design broadens the temporal window for event identification compared with flame-only detection, and provides complementary visual evidence under complex farmland conditions.
• The full precision CBAM-YOLOv8s achieves 0.877 mAP@0.5 and 0.708 mAP@0.5:0.95 at an input resolution of 640 × 640. The INT8-quantized model at 320 × 224 achieved an mAP@0.5 of 0.864 and an mAP@0.5:0.95 of 0.69, with a 72% reduction in model size. Deployed on the low-cost MaixCAM-Pro platform, the INT8 model delivers 24.5 FPS during local inference, with a peak memory usage of 208 MB. Under continuous inference with event-triggered Wi-Fi uploads, the payload consumed an average of 6.7 W and reached 8.2 W during transmission.
• The implemented straw burning monitoring system closes an information-and-alert chain from local detection to remote notification. UAV field tests verified that the system achieved 91% image-level classification accuracy, a local response delay of approximately 100 ms and a cloud-reporting delay of approximately 500 ms.

2 Materials and methods

2.1 Dataset preparation

2.1.1 Image sources and screening

A dedicated drone-view dataset of straw burning detection was developed by combining multiple open source straw fire events datasets (Hu et al., 2025; Wen et al., 2024). All candidate images were screened before annotation. Images corrupted, severely blurred, irrelevant to farmland burning, or duplicated were excluded.

After screening, the final dataset contained 6,586 images. There are 5,951 images defined as positive samples that contain at least one target of flame, smoke, or burned residue, and 635 images retained as negative samples where none of the three target categories was present. Figure 1 presents representative examples from the dataset used in this study. After data collection was completed, a preprocessing workflow was applied to make the images suitable for model training. This workflow included manual annotation, dataset splitting, and data augmentation.

2.1.2 Annotation protocol

All positive images were manually annotated using LabelImg v1.8.1. The 635 fire-free images were preserved as negative samples, with no bounding boxes annotated. The annotation in this study sets three types of targets: flame, smoke and burned residue. It is common that multiple straw burning visual features appear in the same image. Multiple target categories to be labeled within a single image are allowed. The number and visual features of each labeled target category are summarized in Table 1.

TABLE 1. The number and visual features of each category in the dataset.

A single image may simultaneously contain multiple categories of targets, which leads to overlapping samples.

This study added burned residue as a contextual indicator to identify ongoing or completed straw burning events. Flame and smoke are used for real-time fire identification. Burned residue can assist subsequent field inspections in straw burning control areas. The three types of fire-related targets can achieve timely fire detection and also support post-event monitoring. When labeling flame targets, select the minimum bounding rectangle to enclose flame pixels. Smoke was annotated over the complete discernible region in which visibility was reduced by the smoke plume. Burned residue refers to the combusted material remaining on the ground after burning. Therefore, the burned-residue annotation covered the continuous area judged to contain post-combustion residue. During labeling, the surrounding vegetation, ground texture, and relationship with flame or smoke were considered to distinguish burned residue from ordinary shadow, dark soil, or unrelated background objects.

Analysis of the annotation geometry showed that the whole dataset was not restricted to near-field targets occupying most of the image. Among the 11,754 annotated objects, 2,323 (19.76%) occupied less than 10% of the image area. This proportion was 47.62% for flame and 12.23% for smoke. There were 3,515 out of the 5,951 positive images (59.07%) that contained multiple annotated objects, and 2,920 images (49.07%) that contained more than one target class. Nevertheless, 41.54% of the positive images contained at least one target occupying 50% or more of the image.

2.1.3 Dataset splitting and preprocessing

A 7:1:2 random split was adopted for training, validation, and testing, with 4,789, 644, and 1,153 images allocated to each subset, respectively. Specifically, the training, validation, and test sets contained 446 of 4,789 (9.3%), 73 of 644 (11.3%), and 116 of 1,153 (10.1%) negative images, respectively.

For the training set, data augmentation was used to strengthen generalization and prevent overfitting. The augmentation operations included mosaic augmentation, horizontal flip, random brightness adjustment, random contrast adjustment, Gaussian blur and random scaling. The validation and test images were not augmented. Images were resized to 640 × 640 pixels for model input preprocessing.

2.2 Detection model and quantization

2.2.1 Baseline detector and attention mechanisms

The detector was required to localize multiple instances of flame, smoke, and burned residue in a single forward pass. It should also be compatible with the SG2002 INT8 compilation toolchain and the MaixCAM-Pro’s 1-TOPS NPU and 256 MB memory capacity. The YOLO (You Only Look Once) family of detectors has been widely used for target detection (Zhong et al., 2026; Lin et al., 2026; Krishna and Khalaf, 2026). YOLOv8s (small) offers a satisfying balanced trade-off between efficiency and detection accuracy (Tahir et al., 2024). The model uses the C2f-based backbone to enrich feature extraction and gradient propagation. Then, the neck combines Feature Pyramid Network (FPN) and Path Aggregation Network (PANet) structures to fully fuse information from the different backbone levels. In addition, the decoupled detection head separates classification and localization tasks (Wang et al., 2024).

Nevertheless, the direct application of the YOLOv8s model to high-precision fire detection remains challenging. Early-stage flames are typically limited to a few pixels in UAV imagery. The dataset used in this study showed that 47.62% of the annotated flame boxes occupied less than 10% of the image area. Such small fire targets are easily weakened during feature down sampling or masked by background noise, which increases the probability of missed or false detections (Yun et al., 2025).

To strengthen the representation of these difficult targets, two lightweight attention mechanisms were evaluated. The Convolutional Block Attention Module (CBAM) refines a feature map through sequential channel and spatial attention to emphasize informative feature channels and spatial regions (Woo et al., 2018). Coordinate Attention (CA) aggregates features separately along the horizontal and vertical axes, preserving coordinate sensitive features in the dimensional direction and capturing cross directional long-distance feature associations (Hou et al., 2021).

Additionally, the results of the model detection task depend on where the attention module is inserted in the network architecture (Bai et al., 2026; Yu et al., 2024). In order to determine the most efficient configuration for the specific straw burning detection task of this study, the differences in detection accuracy, model size and computational efficiency were compared when they were inserted into the backbone or neck of YOLOv8s. For the backbone, the attention module was chosen to be embedded after the P3, P4 and P5 feature layers. For the neck, the selection was placed after the C2f module outputs the P3, P4, and P5 feature maps at the neck and before the multi-scale fusion features were input to the detection head.

2.2.2 Model training, quantization, and edge deployment

Model training was conducted using PyTorch 2.5.1. Transfer learning was adopted to reduce training difficulty and improve convergence efficiency. The YOLOv8s model was initialized with weights pretrained on the COCO dataset and fine-tuned on the straw burning dataset. All input images were resized to 640 × 640. The primary full-precision FP32 training phase runs for 200 epochs with a batch size of 8 using the Adam optimizer and a cosine annealing scheduler. Table 2 lists the main training parameters used for the experiments.

TABLE 2. Training hyperparameters.

The SG2002 chip on MaixCAM-Pro only provides limited computing power (1.0 TOPS under INT8 inference). The converged full-precision (FP32) model is further optimized via quantization-aware training (QAT) using the PyTorch QAT API. Fake quantization nodes are inserted into the computational graph during QAT fine-tuning. They simulate INT8 arithmetic operations in the forward pass, and enable gradient-based parameter updates to compensate for quantization errors. The model is fine-tuned for an additional 30 epochs with a reduced learning rate of 1 × 10⁻⁴ and a batch size of 8.

After QAT converges, the fine-tuned model is exported as an ONNX file containing quantization parameters. To further improve inference speed on the MaixCAM-Pro, the input resolution is respecified during export from the original 640 × 640 training size to 320 × 224, a dimension that better matches the MaixCAM-Pro’s display aspect ratio. Then, the ONNX model is converted and compiled using the chip manufacturer’s official toolchain, and a native executable format (.cvimodel) for the SG2002 NPU is generated. During conversion, the three YOLOv8s detection heads are specified as output nodes. Compiled model files and descriptor files can be generated after model quantization. These files are deployed on the MaixCAM-Pro platform to realize the real-time inference computation.

2.3 Evaluation metrics

Object-level detection performance was evaluated for flame, smoke, and burned residue using precision, recall, F1-score, mAP@0.5 and mAP@0.5:0.95. The calculation of these metrics depends on matching predicted boxes with ground-truth boxes. Intersection over Union (IoU) is the ratio between the intersection area of a predicted box and a ground-truth box and their union area. Precision (P) reflects the accuracy of the prediction results. Recall (R) reflects the completeness of target detection. The F1-score is the harmonic mean of precision and recall. Average precision (AP) reflects the overall detection effect of a certain type of target of the model. The mean average precision with an intersection ratio threshold of 0.5 (mAP@0.5) is used to evaluate the overall detection level of all categories of targets in the model. The mAP@0.5:0.95 averages AP across IoU thresholds from 0.50 to 0.95 in increments of 0.05. Detailed metric definitions are provided in Supplementary Material.

Image-level and event-level metrics evaluated the operational alarm workflow. An image was considered positive when at least one target was detected with a confidence score of 0.60 or higher. An alarm event required qualifying detections in two consecutive inference outputs. These system-level metrics assess the final warning outcome and should not be interpreted as substitutes for the object-level detection metrics.

Deployment performance was evaluated using model size, peak runtime memory, power consumption and frames per second (FPS) measured on the edge device. Peak memory was obtained from the built-in system monitoring interface of MaixCAM-Pro during continuous image acquisition and model inference. Power consumption was measured using a DC digital power monitor at the power input of the detection and communication payload. The measurement boundary included the MaixCAM-Pro, ESP32-based communication controller, and associated detection and communication electronics. It excluded the UAV propulsion system, flight controller, and battery-conversion losses outside the payload.

2.4 System architecture

The system is designed to process images at the sensor node, issue a local warning and upload compact geotagged event information. Figure 2 presents the modular architecture of the proposed system. Image inference is performed on the UAV-borne MaixCAM-Pro. Only straw burning event results need to be uploaded to the cloud platform. Related UAV-edge systems have reported responsiveness and communication advantages from onboard inference (Mu et al., 2025; Wei et al., 2026). Detailed circuit schematics, interface assignments, and component specifications are provided in Supplementary Material.

2.4.1 Perception layer

The perception layer uses the GC4653 sensor integrated in the MaixCAM-Pro to capture RGB farmland imagery. MaixCAM-Pro camera streamed video continuously at 30 fps. An external GPS receiver provides geographic coordinates through a serial interface. Images are processed locally on the same MaixCAM-Pro device, and GPS information is associated with positive detection events before reporting. The reported GNSS coordinates represent the UAV position at the time of detection rather than the geographical coordinates of the detected burning source. Thus, the perception layer supplies image and location inputs without requiring the continuous transmission of synchronized raw imagery to an external edge gateway.

2.4.2 Edge computing layer

The edge-computing layer is the local inference component of the system. Raw images from the GC4653 sensor enter the device through the MIPI interface. The SG2002 heterogeneous chip in the MaixCAM-Pro hosts the deployed model. The processor cores support operating-system tasks and preprocessing. After on-chip image-signal processing, frames are resized to the 320 × 224 model input and passed to the NPU. Then, the NPU executes INT8 model inference. After each inference cycle starts, the model fetches the latest image frame from the input buffer. If frame data accumulates in the buffer, old frames will be discarded. The target label, confidence, and bounding box information output by the model will be packaged into lightweight event data and transmitted to the communication controller. This modular architecture allows the MaixCAM-Pro to independently generate local structured detection results without relying on remote inference services.

2.4.3 Transmission and interaction layer

The MaixCAM-Pro edge computing module will send the detection results to the main controller. This study selected ESP32_WROOM_32E_N8 as the main controller. This chip integrates Wi-Fi and Bluetooth communication functions. It also has low power consumption characteristics and a complete IoT ecosystem. The selected ThingsCloud cloud platform will assign an independent Message Queuing Telemetry Transport (MQTT) access node to each access device. The ESP32 communicates directly with the cloud platform through MQTT and does not require an intermediate computer.

The ESP32 controller mainly has two tasks. (1) It drives a local alarm module composed of a buzzer and LED indicator light. When a straw burning event is identified, an on-site warning will be triggered immediately. (2) It encapsulates the event detection results and geographical location metadata, and uploads them to the cloud platform through the on-chip Wi-Fi module. Raw images and video are not included in the transmitted message. Cloud communication requires the ESP32 to connect to an external 2.4 GHz Wi-Fi access point with Internet backhaul, such as a mobile hotspot or wireless router. The firmware implements local message buffering and automatic reconnection during temporary network interruptions. ThingsCloud parses the uploaded data and stores it in a structured database. At the same time, the platform can push real-time fire alerts to users through WeChat mini programs and computer web management interfaces.

Onboard inference and local buzzer and LED alarms operate independently of cloud connectivity. In areas with limited network coverage, cloud storage and remote notifications are delayed until the connection is restored.

2.5 Field experimental design

The low-cost straw burning detection system was evaluated in controlled field experiments to assess image acquisition, onboard inference, event transmission and short-duration operational behavior. Tests were conducted in Nanyang, Henan Province, in May 2025. For safety, controlled wheat straw combustion was conducted and sustained for 15 min. Tests took place under clear, gently breezy conditions. The controlled burning was arranged in isolated farmland with fire safety safeguards.

The S500 quadrotor, a 500 mm wheelbase airframe with a glass-fiber center board and engineering-plastic arms, served as the aerial carrier. Equipped with a Pixhawk flight controller, 2,216 motors, and 4 S Li-Po batteries, the platform provided the payload capacity required for the experiment. Detection-related functions, including image acquisition, model inference, GPS association, and event transmission, were performed by the MaixCAM-Pro and ESP32 modules.

As shown in Figure 3, the survey included two scenarios: controlled wheat-straw combustion and fire-free farmland. The fire-free field test was conducted over a piece of farmland covering approximately 9,600 m² (120 m × 80 m). The UAV flew at a target altitude of 30 m above ground level, and the actual altitude ranged from 28 to 33 m. The fire-free field test included two UAV flights, with a total duration of 12 min and 10 min respectively. During the controlled wheat straw combustion experiment, two independent UAV flights were carried out for fixed-point observation. The first flight performed long-range observation from across an adjacent field, with an observation duration of 4 min. The second flight performed close observation of the burning source, with an observation duration of 5 min.

Frames were extracted from the continuous videos at fixed time intervals in both scenarios. Blurred, overexposed, duplicate, and invalid frames were removed. Finally, a total of 78 field-test images were collected. The frame selection was independent of the deployed model output. Manual annotation identified 54 positive images and 24 negative images. A positive image contained flame, smoke, burned residue, or a combination of these targets. The deployed model was also applied to the same 78 images using a confidence threshold of 0.6. The deployed model predictions were compared with the manual image-level labels to calculate TP, FN, FP, and TN.

During continuous operation, a straw burning event was confirmed only when at least one target was detected with a confidence score of 0.6 or higher in two successive inference outputs. Local alarms and cloud notifications were triggered after both conditions were met. Event-triggered snapshots were collected separately and were used only to verify the operational workflow. They were not included in the image-level performance evaluation.

3 Results

3.1 Model performance

We evaluated two lightweight attention mechanisms and compared their performance embedded in YOLOv8s′ backbone or neck to find the optimal setup for straw burning detection. All experiments are performed under identical training settings, with the baseline YOLOv8s model (without any attention module). The detection results are summarized in Table 3. Figure 4 shows the PR curves of the five comparative models. Because the maximum absolute overall mAP@0.5 improvement is 0.01, the experiment result should be interpreted as a task-specific selection outcome rather than evidence of a generally superior architecture.

TABLE 3. Detection performance comparison of attention module configurations for 640 × 640 image input.

As shown in Table 3, the effects of the attention mechanisms depended on the module type and insertion position. The C1 model achieved an overall precision of 0.879, recall of 0.798, F1-score of 0.836, mAP@0.5 of 0.877, and mAP@0.5:0.95 of 0.708. The C1 configuration also delivers the best performance on flame detection (0.845, +3.7% relative to baseline) and burned residue detection (0.923, +1.1%). This indicates that CBAM embedded at the feature extraction stage improves sensitivity to flame and burned residue. The PR curves in Figure 4d corroborate this finding that C1 achieves the largest area under the curve for both flame and burned residue categories.

At the category level as shown in Supplementary Table S2, burned residue achieved the highest performance, with an F1-score of 0.882 and mAP@0.5:0.95 of 0.811. Smoke showed intermediate performance, with an F1-score of 0.823 and mAP@0.5:0.95 of 0.697. Flame was more difficult to detect, with an F1-score of 0.804 and mAP@0.5:0.95 of 0.618.

In terms of computational efficiency, CA-based configurations (B1 and B2) introduce no additional GFLOPs and only a marginal increase in model size (+0.07 MB, +0.3%), confirming the lightweight nature of the CA mechanism. CBAM-based configurations (C1 and C2) also maintains competitive computational efficiency, with only marginal increases in computational cost (from 28.7 to 28.9 GFLOPs, +0.7%) and model size (+0.66 MB, +3.1%).

Therefore, we adopt C1 (YOLOv8s with CBAM embedded in the Backbone) as the optimized model for subsequent edge deployment. Given that early-stage flame detection is the most critical requirement for timely straw burning intervention, this trade-off is considered acceptable and practically beneficial.

3.2 On-device performance

We quantized the full-precision C1 model to INT8 and deployed it on the MaixCAM-Pro platform. As summarized in Table 4, the INT8-quantized C1 model at 320 × 224 achieved an mAP@0.5 of 0.864 and an mAP@0.5:0.95 of 0.69. The AP values were 0.827 for flame, 0.851 for smoke, and 0.914 for burned residue. The results showed that INT8 quantization largely preserved the detection accuracy.

TABLE 4. On-device performance of the INT8 C1 model for 320 × 224 image input.

SG2002 NPU INT8 for 320 × 224 image input.

FPS, is measured end-to-end on the MaixCAM-Pro, including MIPI image capture, ISP processing, resizing to 320 × 224, and NPU inference, but excluding Wi-Fi transmission to the ESP32.

Peak memory refers to the maximum memory use observed during continuous image acquisition and INT8 inference on the MaixCAM-Pro.

To distinguish the effect of input resolution reduction from that of numerical quantization, the FP32 C1 model at 320 × 224 was evaluated on the same validation dataset. With the reduction of the input resolution from 640 × 640 to 320 × 224, the FP32 C1 model showed a decrease in mAP@0.5 from 0.877 to 0.869 and in mAP@0.5:0.95 from 0.708 to 0.698, corresponding to absolute reductions of 0.008 and 0.01, respectively. At the same input resolution of 320 × 224, INT8 quantization further decreased mAP@0.5 from 0.869 to 0.864 and mAP@0.5:0.95 from 0.698 to 0.690. Therefore, the accuracy reductions attributable to INT8 quantization were 0.005 for mAP@0.5 and 0.008 for mAP@0.5:0.95.

The INT8-quantized C1 model ran at 24.5 FPS in the on-device deployment benchmark, with a model size of only 6.2 MB. The measured frame rate was higher than 15 FPS. This reference value was defined in this study for the UAV monitoring system. This INT8-quantized model occupies a small size, which greatly reduces the storage space and running memory overhead of the MaixCAM-Pro. Peak memory use was 208 MB, within the device capacity. Power measurements refer to the detection and communication payload rather than the complete UAV. Under continuous inference with event-triggered Wi-Fi uploads, the payload consumed an average of 6.7 W and reached 8.2 W during transmission.

To assess whether limited negative training samples lead to deployment false positives, the deployed INT8 C1 model was evaluated on 116 negative samples extracted from the test set. At a confidence threshold of 0.60, the deployed model produced 14 false-positive detections in 9 images. This corresponds to a per-image false positive rate of 7.76%. This detection-level result explains why our operational workflow combines image-level and event-level metrics, since temporal confirmation helps filter out isolated false-positive detections at the alarm stage.

3.3 Field detection performance

Among the total 78 collected images, 54 samples contained actual straw burning fire scenarios and 24 were fire-free backgrounds. A fire-present image could contain one or more of the three annotated target categories: flame, smoke, and burned residue. Figure 5 presents representative detection outputs produced by the straw burning monitoring system during the controlled field experiments.

An image was classified as positive when at least one target was detected at a confidence threshold of 0.6. At the image level, 49 of the 54 fire-present images produced positive image-level classifications and were classified as true positives (TP), 5 fire-present images were missed and classified as false negatives (FN). Among the 24 fire-absent images, 22 were correctly classified as true negatives (TN), 2 generated false alarms and were classified as false positives (FP), as summarized in Table 5. Overall image-level classification accuracy was 91% (71/78). Note that the 91% field accuracy is an image-level binary correctness rate (fire-present vs. fire-free), which differs from the per-category, IoU-thresholded mAP@0.5. Image-level classification performance in the field experiment is summarized in Table 6. The system achieved a recall of 90.7%, precision of 96.1%, and F1-score of 93.3%.

TABLE 5. Image-level binary classification confusion matrix for the deployed INT8-quantized C1 model in the controlled field experiments.

TABLE 6. Image-level classification performance of the deployed INT8-quantized C1 model in the controlled field experiments.

Category-specific differences were observed in the field results. Table 7 compares three cumulative detection rules at the image level. The flame-only presented the highest risk of missed detection. The flame-only rule identified 34 of the 54 positive images, resulting in a recall of 63.0% and a precision of 97.1%. Adding smoke increased the number of true positives to 42 and improved recall to 77.8%. This recovered eight images missed by the flame-only rule. However, the number of false positives increased from one to two. Smoke helped identify events when visible flames were weak or obscured. However, its diffuse boundaries are difficult to distinguish from background. Adding burned residue further increased the number of true positives to 49 and recall to 90.7%. It recovered seven additional positive images without introducing new false positives. Since burned residue remained visible after the flame and smoke had disappeared, its detections provide post-event evidence and can support targeted farmland inspection. The three target categories provided complementary information across different stages of straw burning and broadened the temporal window for image-level detection.

TABLE 7. Cumulative image-level performance of different detection rules derived from the deployed INT8-quantized C1 model.

All rules were evaluated using outputs from the deployed INT8-quantized C1 model on the same 54 positive images and 24 negative images at a confidence threshold of 0.6.

Supplementary Figure S5 showed the image-level five false negative and two false positive cases examined by manual inspection. Among the five false negative images, three contained small flames that occupied only a limited portion of the image, and the other two contained diffuse smoke with low contrast against the field background. These observations indicate that small target size and weak visual contrast were the main characteristics associated with the missed cases. The two false positive images were associated with different visual similarities. A distant field region with a color similar to fire was misidentified as flame. A hazy background region was misidentified as smoke. It was also confirmed in Table 7 that smoke detection helped to recover additional fire-present images, but it remained susceptible to smoke like background interference.

3.4 Alert and communication performance

A straw burning event was confirmed when at least one target was detected in two successive inference outputs with a confidence score of 0.60 or higher. Once confirmed, the event record was sent to the ESP32 controller. ESP32 immediately triggered on-site alarms and uploaded event records to the cloud platform, which pushed remote alarm notifications. Field experiment results showed that the model inference frame rate ranged from 19 FPS to 27 FPS, and the average was 22 FPS. The response delay for local alerts was approximately 100 ms. The transmission delay of straw burning events reporting to the cloud was approximately 500 ms.

During the field experiment, a smartphone was configured as a 2.4 GHz Wi-Fi hotspot to provide temporary Internet access for the ESP32. The ESP32 connected directly to the ThingsCloud MQTT broker through the hotspot. The serial output is shown in Supplementary Figure S4. It confirmed successful Wi-Fi and MQTT connections. There was correct reception and parsing of the straw burning detection (SBD) status from the MaixCAM-Pro and the geographical coordinates from the Air530Z. The resulting structured message contained the SBD status and its corresponding latitude and longitude. It was verified that the functional data path from the sensing modules through the ESP32 to the cloud platform worked as expected.

As shown in Figure 6, ThingsCloud marked the straw burning detection (SBD) point with a red mark on the map. As shown in Figure 7a, the cloud platform stored historical alarm data for subsequent query and statistical analysis. Figure 7b showed that the WeChat Mini Program can also receive messages at the same time. The field experiment verified the entire operational workflow from drone onboard sensing, edge-end inference to cloud remote alarm.

4 Discussion

4.1 Environmental management implications

The practical environmental significance of the proposed system lies in its potential to shorten the interval between the appearance of straw burning evidence and its communication to field managers. Rather than replacing satellite monitoring or manual patrols, the system can serve as a complementary tool for detecting small or short-duration events that require closer local observation. The three target categories provide information at different stages of a fire event to extend the system from real-time straw burning alerts to retrospective verification after the event. Flame indicates active combustion, smoke may reveal weak or smoldering combustion, and burned residue preserves post-event visual evidence. Onboard inference converts the captured imagery into compact event records without continuously transmitting raw video. Different from the traditional offline detection, when this system captures a straw burning event, it will trigger on-site alarms and remote notification. The cloud platform stores all straw burning event reports. It supports staff to conduct remote inspections around the reported UAV location and make scientific management decisions based on monitoring data.

4.2 Role and limitations of the YOLOv8s modification

The selection of YOLOv8s should be interpreted as a hardware-constrained engineering decision. YOLOv8s satisfied the memory and throughput requirements of the MaixCAM-Pro after INT8 quantization. The ablation experiment showed that backbone CBAM provided the best trade-off among the configurations evaluated in this study, but the YOLOv8s modification should not be interpreted as the main highlight of the study. CBAM in the backbone increased overall mAP@0.5 from 0.867 to 0.877, and improved flame AP from 0.815 to 0.845. However, smoke detection did not benefit to the same extent as flame. One possible explanation is that diffuse smoke has weaker boundaries and broader spatial extent than compact flame or residue regions (Keerthinathan et al., 2025), so the tested attention mechanisms may not enhance its discriminative features in the same way. Therefore, the model-level novelty is incremental and does not establish that CBAM is universally preferable for fire detection. More extensive fire detectors use lightweight backbones (Saydirasulovich et al., 2023), multi-scale neck redesign (Lei et al., 2024), alternative convolutions (Tao et al., 2026), attention modules (Zheng et al., 2025), and localization losses (Zhao et al., 2024) together. The YOLOv8s model improvement in this study is best viewed as an enabling component. Its value is demonstrated through resource-constrained system deployment.

4.3 Comparison with related edge fire-detection systems

Few studies have deployed fire-detection systems on embedded edge devices with consistent cost and evaluation benchmarks. Table 8 compares the proposed system with several representative edge-based detectors. However, these studies differ in fire scene, dataset, metric definition, hardware boundary, and test protocol. Since the overall cost of a UAV prototype varies greatly, the cost comparison only covers the core edge-computing module. It supports a low-cost and speed-oriented design comparison, not a direct ranking of detection accuracy or total UAV-system cost.

TABLE 8. Characteristics of edge-based fire-detection systems.

Multi-scene fire including forest, vehicle, industrial, accidental, residential, and indoor fires.

The hardware cost listed in this table refers to the edge-computing module or development board. It does not include components required for a complete drone system.

mAP@0.5 is used for offline evaluation, and image-level accuracy for field tests. These two metrics are derived from different datasets and evaluation protocols and thus cannot be directly compared.

The measurement boundaries for FPS vary among references.

The NVIDIA Jetson-based systems offer high offline mAP@0.5 on public datasets, but they rely on costly and power-hungry platforms. The Raspberry Pi 5-based systems use general-purpose hardware, but achieve only 8 FPS for edge inference. The proposed system follows a different design route by using a dedicated INT8 NPU and integrating inference with UAV-position association, local warning, and compact cloud reporting. Our proposed system attains mean 22 FPS inference speed on a 1-TOPS edge module costing USD 50 in field experiment. Although it does not claim the highest absolute accuracy, it offers a favorable balance among hardware cost, power consumption, inference speed, and field-ready performance. The low module cost and compact form factor make it feasible to equip multiple UAVs or distributed monitoring nodes. Thus, the system shows potential for large-scale agricultural deployment.

4.4 Limitations and future work

The proposed system indicates that local event detection and remote supervisory reporting can be combined on a low-cost payload. The system still has several limitations.

The current system associates confirmed detections with the GNSS position of the UAV. For an application that only needs event presence and platform-position reporting, a binary or multi-label classifier could provide a simpler alternative. Because geo-localization using images could enable the UAV to monitor a broader ground area visually from fewer flight paths, future work could combine detection or segmentation with camera calibration, UAV pose information, and multi-view geometry to achieve more precise fire source localization. These localization outputs could further guide adaptive flight-path planning by selecting viewpoints that improve coverage or localization geometry (Ullah et al., 2025).

Although the system required detections in two consecutive inference outputs before confirming a straw burning event, it was difficult to completely eliminate false positives. It would be beneficial to collect confusing background scenes (clouds, dust, sunlight glare) and add them to the training dataset. Stricter temporal criteria may reduce false alarms, but their effects on recall and alarm latency require further assessment. Ensemble-based detection verification could be evaluated as a potential means of suppressing persistent false positives (Renju and Brunda, 2024).

Remote notifications and cloud storage require network connectivity. The cloud latency reported by the field experiment is only applicable to the tested Wi Fi conditions. Future research will evaluate a LoRa link with an Internet-connected gateway to improve the stability of cloud data upload. There is also a lack of diversity in the sample scenarios, with the dataset consisting of only 78 selected frames from four UAV flights. The field experiment for straw burning scenarios was artificially set up and limited to a single location and a single season. The experimental combustion conditions could not cover the changes in crop types, fire scale, and smoke diffusion patterns. In addition, the burning source was established under controlled safety conditions. It may differ from naturally occurring straw burning events in fuel arrangement, duration, and occlusion. Future work should evaluate the system performance across multiple sites, seasons, crop types and meteorological conditions.

5 Conclusion

This study developed a low-cost straw burning monitoring system combining local inference with event-based alert reporting. An INT8-quantized CBAM-YOLOv8s is deployed on the MaixCAM-Pro for local inference. The system showed an image-level classification accuracy of 91.0% and an F1-score of 93.3% in field experiments. The newly added burned residue detection target extends the effective monitoring period beyond the open flame and smoke phase. The local alarm and cloud reporting modules triggered by straw burning events worked as expected.

The system can help reduce the spatial and temporal gaps of manual patrols and satellite observations. It also supports faster identification of suspected straw burning locations. However, the present evidence comes from one location and one season. Future work will include more crop types, weather conditions, regions, and natural burning events.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. During the preparation of this manuscript, the authors used ChatGPT (OpenAI, GPT-5.6) for language translation and polishing. The authors have reviewed and edited the output and take full responsibility for the content of this publication.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fenvs.2026.1966149/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Dmitry Efremenko, German Aerospace Center (DLR), Germany

Reviewed by: Rajesh N, College of Applied Sciences, Sohar, Oman

Julius Pesonen, Finnish Geospatial Research Institute, Finland