Resolution-Standardized Bamboo-Pole Counting from Multi-Resolution UAV Imagery for Coastal Monitoring

Coastal erosion and mangrove degradation remain important management challenges along the muddy coasts of Thailand. The Department of Marine and Coastal Resources (DMCR) recognizes bamboo fences as one form of coastal erosion mitigation [1]. Their use reflects a broader need to combine practical shoreline protection with shared responsibility among public agencies and coastal stakeholders [2].
Bamboo coastal structures also represent a locally adapted coastal-management practice. Digital records of these structures may support long-term documentation of their construction, maintenance, and change over time [3]. However, the present study focuses specifically on the technical problem of reliable bamboo-pole counting from UAV imagery rather than evaluating cultural preservation outcomes.
When the previous model was deployed and tested by the DMCR, it was reported to perform well when processing input images on a computer. However, when the same images were submitted using other devices, such as Android or iOS smartphones, the model produced inconsistent counting results, as illustrated in Fig. 1. Therefore, this work aims to investigate and address this discrepancy.
In our previous work, Galajit et al. [4] developed a two-stage framework combining bamboo-region segmentation with individual-pole detection for bamboo counting from UAV imagery.
The segmentation model and the high-resolution detector used in the present study were inherited from this previous framework. In the present work, the segmentation model was retained without retraining, while the inherited detector was used as the baseline and as the initialization for detector fine-tuning.
The observed device-dependent inconsistency motivated the present study to investigate whether differences in input-image resolution could contribute to variations in counting output. The study focuses on improving counting consistency across controlled resolution variants through detector fine-tuning and resolution standardization. Its main contributions are: (1) detector fine-tuning for small, densely arranged bamboo-pole ends; (2) a resolution-standardized inference workflow; and (3) evaluation across 4K, 2K, and 1080p inputs.

Studies of the Upper Gulf of Thailand report that bamboo fences may promote sediment accumulation under some site conditions, although their effectiveness varies with hydrodynamics, design, maintenance, and local geomorphology [5]. Experimental work in Southeast Asia has also shown that fence porosity and arrangement influence wave transmission and wetland protection [6]. These structures are therefore context-dependent interventions rather than universally effective solutions.
Maintenance is particularly important because bamboo has a limited service life. Deteriorated or detached material can become debris and damage mangrove stems, demonstrating that installation without long-term monitoring can create additional environmental risks [7]. Repeated assessment of nature-based coastal measures is consequently necessary to distinguish short-term construction completion from sustained performance [8].

Unmanned aerial vehicles provide flexible, repeatable, and comparatively low-cost data acquisition for coastal change monitoring [9]. Their ability to capture high-resolution imagery over difficult-to-access areas makes them suitable for documenting long linear structures such as bamboo fences. In cultural-heritage applications, UAV imagery combined with deep learning has also been used to automate monitoring tasks and support authorities with limited inspection resources [10]. These studies demonstrate the broader value of remote sensing as a tool for longitudinal documentation and decision support.

Aerial object detection is challenging because objects may occupy few pixels, appear at multiple scales and orientations, and occur in dense clusters [11]. Region-focused and scale-normalized methods have been proposed to improve the detection of small aerial objects [12]. The bamboo-counting task presents similar difficulties because individual pole ends are small, visually similar, and closely packed. The system adopts the YOLO family of one-stage detectors [13] and uses the YOLO11 implementation provided by Ultralytics [14].

The complete two-stage bamboo-counting workflow is shown in Fig. 2. After optional resolution standardization, the segmentation model identifies bamboo-containing regions and limits detection to cropped regions of interest. Each region is then processed using sliced object detection. The detector identifies visible circular or near-circular bamboo-pole ends, and each accepted detection contributes one unit to the predicted count.

For the controlled resolution experiment, each original high-resolution UAV image was first resized using LANCZOS interpolation to 3840 × 2160 (4K), 2560 × 1440 (2K), and 1920 × 1080 (1080p). These three variants therefore depict identical scenes and differ only in the controlled image-resolution transformation.
Resolution standardization is a separate inference step. Before segmentation, the input resolution is checked. The 2K and 1080p variants selected for standardization are enlarged to 3840 × 2160 using LANCZOS interpolation, whereas the 4K variant requires no additional standardization. The purpose of this step is not to reconstruct visual information lost during downsampling, but to provide lower-resolution inputs at a spatial scale more consistent with the detector and sliced-inference configuration.

Predicted counts are compared with manual reference counts using mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), and mean signed error (bias), defined as follows.

The final counting evaluation used 50 high-resolution UAV images with corresponding manual reference counts. These 50 image files were not used for detector training, validation, or fine-tuning and constituted a fixed counting-evaluation set. The manual reference counts ranged from 305 to 588 poles, with a mean of 449.66 and a total of 22,483. Because the model-development dataset and its original split were inherited from the previous project, source-flight or source-video overlap between the inherited development imagery and the 50 evaluation images could not be independently verified in the present study.
For the controlled resolution experiment, each source image was resized using LANCZOS interpolation to 3840 × 2160 (4K), 2560 × 1440 (2K), and 1920 × 1080 (1080p). Thus, the three resolution sets depict identical scenes and represent paired resolution variants rather than independent image datasets. Two detection models were evaluated: the inherited detector from the previous project, used as the baseline, and the fine-tuned detector developed in the present study as part of the proposed framework. Five unique input-processing conditions were considered for each detector: 4K, 2K without standardization, 2K standardized to 4K, 1080p without standardization, and 1080p standardized to 4K. With two detectors and 50 evaluation images, this resulted in 10 unique experimental configurations and 500 unique inference results.
The same inherited segmentation model and sliced-inference procedure were used across all configurations. The segmentation and detection confidence thresholds were set to 0.60 and 0.20, respectively, and the combined segmentation mask was binarized using a threshold of 0.50. Each segmented bamboo region was expanded by 50 pixels after resolution preparation and before sliced detection. Each region of interest was processed using 640 × 640-pixel slices with horizontal and vertical overlap ratios of 0.20. Predictions from overlapping slices within each region were reconciled using SAHI non-maximum suppression (NMS), with IoU as the matching metric and a matching threshold of 0.45.
To provide a resolution-consistent post-processing criterion, the maximum accepted bounding-box dimensions were scaled relative to the 4K reference resolution. Thresholds of 30 pixels at 3840 × 2160, 20 pixels at 2560 × 1440, and 15 pixels at 1920 × 1080 were used. Standardized 2K and 1080p variants were processed at 3840 × 2160 and therefore used the 30-pixel threshold. Accepted detections from all segmented regions were aggregated to obtain the final predicted count; no additional global duplicate-removal stage was applied across different segmented regions.

The bamboo-region segmentation model was inherited from the previous framework [4] and was used without further training or modification in the present study. Therefore, the training performed in this study was limited to fine-tuning the bamboo-pole detection model.
The detector was initialized from the high-resolution bamboo-pole detection model developed by the previous project team. The existing dataset split from the previous project was retained rather than re-divided in the present study. It contained 1,164 training images, 111 validation images, and 55 test images. The separate set of 50 manually counted evaluation images described in Section IV-A was not included in these training, validation, or test sets.
Only the training split was augmented for fine-tuning. Each of the 1,164 training images was retained in its original form and used to generate six additional variants: resolution degradation to 50% and 75%, JPEG compression at quality levels 40 and 70, and sharpness factors of 0.50 and 1.50. Thus, each source training image contributed one original and six augmented samples, corresponding to a 1:6 original-to-augmented ratio. For resolution degradation, images were downsampled using bicubic interpolation and restored using LANCZOS interpolation. This increased the training set from 1,164 to 8,148 images, while the 111 validation and 55 test images were retained without this offline augmentation.
Fine-tuning was performed in Google Colab using Ultralytics 8.4.90, Python 3.12.13, PyTorch 2.11.0+cu128, and an NVIDIA L4 GPU. The model input size was 640 × 640 pixels with a batch size of 32. Training was configured for a maximum of 100 epochs with an early-stopping patience of 20. Ultralytics automatic optimizer selection chose MuSGD with an initial learning rate of 0.01, momentum of 0.9, and weight decay of 0.0005. No explicit layer-freezing strategy was specified. Training stopped after 52 epochs, and the checkpoint selected by the Ultralytics validation criterion corresponded to epoch 32. It achieved precision of 0.942, recall of 0.955, mAP@0.5 of 0.967, and mAP@0.5:0.95 of 0.626.

Table I summarizes counting performance using the resolution-scaled size filter. The inherited detector generally undercounted across all evaluated resolution variants. Its lowest MAE was 53.52 poles for 2K input standardized to 4K, whereas 1080p without standardization produced an MAE of 211.84 poles. Fine-tuning substantially improved performance on the 4K resized variant, reducing MAE from 64.40 to 19.60 poles. However, fine-tuning alone did not improve the 1080p condition: without standardization, the fine-tuned detector produced an MAE of 253.60 poles compared with 211.84 poles for the inherited detector.
The largest observed improvement occurred for the fine-tuned detector at 1080p. After 4K standardization, MAE decreased from 253.60 to 19.04 poles and bias changed from −253.60 to +1.92, corresponding to approximately a 92.5% reduction in MAE. At 2K, standardization reduced MAE from 63.64 to 18.90 poles. These results show substantially improved counting consistency after standardization under the evaluated resolution conditions.

Because the original pipeline used a fixed 30-pixel maximum bounding-box dimension, an additional ablation was performed to determine whether this rule biased the resolution comparison. Three variants were evaluated using the same detections: no size filter, the original fixed 30-pixel filter, and the proposed resolution-scaled filter. For the fine-tuned detector, the resulting MAEs were similar across all five input conditions, indicating that the principal resolution-dependent trends were not caused by the size-filtering rule.

The lowest observed MAE was 18.90 poles for 2K input standardized to 4K. However, the differences from standardized 1080p input (19.04 poles) and the 4K resized variant (19.60 poles) were small. The corresponding MAPE values were 4.23%, 4.23%, and 4.42%, respectively. These three conditions are therefore treated as producing comparable counting errors unless a paired statistical comparison demonstrates a meaningful difference. Figure 4 shows that predictions for the 2K-to-4K condition generally follow the manual reference counts, although errors remain for some images.

Controlled downsampling reduces the apparent size and edge detail of bamboo-pole ends and was associated with substantial undercounting when lower-resolution variants were processed without standardization. Resolution standardization does not reconstruct visual information lost during downsampling. Instead, it increases the apparent pixel dimensions of bamboo-pole ends and presents the lower-resolution variants at a spatial scale more consistent with the detector and sliced-inference configuration.
The results suggest that scale mismatch is an important factor in the observed performance differences; however, the present experiment does not isolate this effect from segmentation, sliced inference, interpolation, and post-processing. The segmentation model, ROI extraction, and sliced-detection procedure were held fixed across conditions. Consequently, the reported counting performance represents the combined pipeline and should not be attributed solely to image resizing. The size-filter ablation further showed that the main resolution trend remained similar when the filtering rule was changed or removed.

The proposed inventory may contribute to documenting locally adapted coastal-management practices by creating systematic records of visible bamboo-pole counts. However, the present experiment evaluates counting performance only. Its usefulness for maintenance planning, ecological assessment, cultural documentation, or longitudinal monitoring requires further validation through repeated surveys and stakeholder use.
Published studies report both potential sediment-trapping benefits and important limitations, including short service life and damage caused by detached bamboo [5]–[8]. Automated counting should therefore support inspection and longitudinal records rather than act as a standalone measure of structural integrity, wave attenuation, sedimentation, or ecological impact. These outcomes require complementary engineering and environmental measurements.

The evaluation used controlled resolution variants derived from 50 images from a single project dataset and therefore does not establish robustness across different UAV cameras, flights, acquisition altitudes, compression pipelines, or environmental conditions. Because the model-development dataset was inherited from the previous project, source-flight or source-video overlap between the inherited training imagery and the 50 evaluation images could not be independently verified. Detailed information on the manual reference-counting protocol, including the number of annotators, treatment of occluded, broken, leaning, and boundary poles, disagreement resolution, and inter-annotator agreement, was not available for the present study.
Count-level metrics also cannot distinguish missed poles, false detections, duplicate detections, segmentation failures, or filtering errors. Bounding-box ground-truth annotations were not available for the 50-image evaluation set, and therefore precision, recall, and F1 score could not be calculated on this set. In addition, the individual contributions of segmentation and sliced inference were not isolated. Resolution standardization enlarges lower-resolution imagery but cannot recover information absent from the downsampled input.

This paper presented a resolution-standardized two-stage framework for counting visible bamboo-pole ends in UAV imagery. Fifty high-resolution UAV images were transformed into paired 4K, 2K, and 1080p variants to evaluate counting performance under controlled resolution changes. Using the fine-tuned detector, 4K standardization reduced the 1080p MAE from 253.60 to 19.04 poles, while standardized 2K and the 4K resized variant produced MAEs of 18.90 and 19.60 poles, respectively. These results demonstrate improved counting consistency across the evaluated controlled resolution variants but do not establish robustness across different cameras, flights, or acquisition conditions. The resulting procedure may support systematic bamboo-count inventories for coastal monitoring and should be used alongside field inspection and environmental assessment.

The authors gratefully acknowledge the Sirindhorn International Institute of Technology (SIIT), Thammasat University, and Queen Mary University of London (QMUL) for their financial support. This support was granted through the International Science Partnerships Fund (ISPF) to promote academic collaboration between QMUL and SIIT.