Introduction
In modern forestry, "mounding" is a critical soil preparation technique where mechanical excavators create specific microsites, or mounds, to optimize seedling growth. Accurate counting of these mounds is essential for logistics, inventory management, and ensuring planting density targets are met. Traditionally, this task relies on manual field surveys or human interpretation of aerial imagery—both of which are labor-intensive, costly, and prone to subjective error.
Automating this process using Unmanned Aerial Vehicle (UAV) imagery presents significant challenges. Mounds are often obscured by woody debris, eroded by water, or hidden in shadows, making them difficult for standard object detectors to identify reliably.
Methodology: The Detect-then-Correct Framework
To address the limitations of standard detection, the authors introduce MoundCount, a hybrid two-stage framework that combines local object detection with a global regression module.
1. Local Mound Detection
The system processes large UAV orthomosaics by tiling them into smaller patches. A YOLO (You Only Look Once) detector is employed to identify clearly visible mounds. To enhance robustness against varying backgrounds, the training process incorporates data augmentation techniques, specifically scaling and translating bounding boxes to provide better environmental context.
2. Global Count Estimation
Recognizing that physical occlusions lead to undercounting, MoundCount includes a global estimation module. This module utilizes Ridge Regression to refine the initial detection count. By incorporating global features such as the total area of the planting block and the expected average mound density, the regressor "fills in the blanks" for mounds missed by the local detector due to environmental factors.
Dataset
The framework was validated using a newly constructed dataset consisting of 18 UAV orthomosaics from plantation blocks in Southern Quebec, Canada. The dataset includes over 111,000 annotated mounds. High-resolution images (approx. 20,000 × 30,000 pixels) were subdivided into 608 × 608 patches for model training and evaluation.
Key Results and Insights
MoundCount significantly outperforms standard detection and density-estimation models.
- Accuracy: The framework achieved a Mean Relative Precision (MRP) of 87.43%, compared to 82.29% for the best density-estimation model (MCCNN) and only 69.34% for standard YOLO detection.
- Error Reduction: The Mean Absolute Error (MAE) was reduced to 608, a substantial improvement over MCCNN's MAE of 1670.
- Correction Power: The global module proved vital in challenging conditions. In one test case (Block 12), the regression module improved the final count by 70% over the detector's output alone.
The study also established a Human-Level Performance (HLP) benchmark, finding that human error rates varied between 4% and 12% depending on terrain difficulty, highlighting the complexity of the task even for expert annotators.
Conclusion and Future Work
MoundCount demonstrates that a hybrid approach—treating detection as a preliminary step followed by global context correction—is highly effective for counting objects in "wild" environments where physical occlusion is common.
While the current system uses RGB imagery, future work aims to incorporate Thermal Infrared (IR) imaging. Since mounds consist of exposed mineral soil, they exhibit distinct thermal signatures compared to surrounding vegetation and debris, potentially allowing for even more accurate counts in heavily obscured areas.
Summary updated