Abstract
Accurate monitoring of palm populations in natural tropical forests is critical for conservation but hindered by complex canopy structures and computational constraints. This paper presents PalmDSNet, a deep learning framework designed for real-time detection, segmentation, and counting of palms using Unmanned Aerial Vehicle (UAV) imagery. By integrating state-of-the-art object detection models with a novel Poisson-Gaussian bimodal reproduction algorithm, the study enables both precise mapping and ecological simulation of palm spatial distributions.
Introduction
Monitoring tropical forests is notoriously difficult. Unlike structured palm oil plantations, natural forests are characterized by overlapping crowns, heterogeneous terrain, and complex lighting conditions. These factors create "false" features that challenge traditional remote sensing methods. Furthermore, the need for real-time processing on-board UAVs necessitates lightweight yet accurate models that can operate in chaotic environments.
This research addresses these challenges by providing a robust dataset of over 7,000 expert-annotated palms across diverse ecological gradients in Ecuador and a flexible framework that bridges high-accuracy deep learning with practical field ecology.
Methodology
The proposed solution consists of two primary components:
- PalmDSNet Framework: A modular architecture that utilizes detection backbones—including YOLOv8, v9, v10, and RT-DETR—to localize palms. This is followed by the Segment Anything Model (SAM) for precise crown delineation. To ensure model interpretability, Grad-CAM saliency maps were employed to visualize the features (e.g., leaf patterns) driving the AI's predictions.
- Bimodal Reproduction Algorithm: To move beyond simple counting, the authors developed a simulation model to analyze spatial distribution. This algorithm combines a Poisson process (simulating long-range animal-mediated seed dispersal) with a local Gaussian distribution (simulating seeds falling near parent trees). The simulations were validated against real-world data using Ripley’s K-series functions.
Data and Experiments
The study utilized UAV-captured imagery from 21 sites in western Ecuador, representing a gradient from everwet rainforests to dry tropical forests. The resulting dataset includes 7,356 bounding boxes and 7,603 manually marked palm centers, providing a new benchmark for the community.
Key Results and Insights
- Model Performance: RT-DETR and YOLOv9 emerged as the strongest models. RT-DETR achieved a Mean Average Precision (mAP) of 0.734 and an AP75 (a stricter accuracy metric) of 0.822 when trained on 90% of the dataset.
- Counting Accuracy: The framework demonstrated exceptional reliability, with counting accuracy exceeding 99% in environments similar to the training data and remaining above 94% in previously unseen forest types.
- Real-Time Efficiency: Inference speeds exceeded 40 frames per second (FPS) on standard hardware, confirming the feasibility of on-board UAV deployment for rapid ecological assessments.
- Ecological Findings: Spatial analysis confirmed that palms are not randomly distributed but exhibit clustering and "negative density dependence." The success of the bimodal model suggests that palm populations are shaped by a balance of local recruitment and long-distance dispersal.
Conclusion and Future Work
PalmDSNet provides a scalable, real-time solution for forest monitoring. While currently robust, future work will focus on localized parameter estimation for the bimodal algorithm and the use of raw multi-view UAV imagery to detect palms partially obscured by taller canopy layers. Expanding the dataset to include species-level identification remains a priority for tracking endangered or economically vital species across the Amazonian basin.
Summary updated