PalmProbNet: A Probabilistic Approach to Understanding Palm Distributions in Ecuadorian Tropical Forest via Transfer Learning

Kangning Cui, Zishan Shao, Gregory Larsen, Victor Paúl Pauca, Sarra M. Alqahtani, David Segurado, João Pinheiro, Manqi Wang, David A. Lutz, Robert J. Plemmons, Miles SilmanACM Southeast Regional Conference 2024 · DOI · OpenReview

TL;DR

This paper introduces PalmProbNet, a novel deep learning approach for accurately detecting and mapping palm trees in dense tropical rainforests using high-resolution UAV imagery. By employing a probabilistic, dual-scale patch-based transfer learning strategy, PalmProbNet achieves impressive accuracy, offering a valuable tool for ecological research and forest management.

What Problem Does This Paper Address?

Accurately identifying and locating palm trees in tropical ecosystems is crucial for understanding biodiversity, soil health, and supporting local communities and wildlife. However, this task is significantly challenging due to dense vegetation, overlapping canopies, variable lighting, and the inherent difficulty of obtaining comprehensive labeled data in such environments. Traditional methods often struggle with the complexity of these landscapes, and even deep learning approaches face hurdles with data scarcity and the need for precise feature extraction. This paper aims to bridge this gap by developing a robust and scalable automated method for palm detection in challenging tropical forest settings.

Methods

The core of PalmProbNet lies in its innovative use of transfer learning and a dual-scale patch-based approach.

  • Transfer Learning with ResNet-18: The authors leverage a pre-trained ResNet-18 model, a powerful convolutional neural network known for its ability to extract hierarchical features from images. They fine-tune the final layers of ResNet-18 to adapt it to the specific characteristics of their palm tree detection task, while keeping most of the earlier layers frozen to preserve the learned general features. This significantly reduces the need for massive amounts of labeled data.
  • Dual-Scale Patching: To address the challenge of varying palm visibility and canopy density, PalmProbNet utilizes two sizes of image patches: small (40x40 pixels) and large (100x100 pixels).
  • Small Patches: These are used to capture fine-grained details of palm leaves, ensuring that even partially occluded palms can be identified.
  • Large Patches: These provide broader contextual information about tree crowns, which is essential for distinguishing palms from other tree species and handling more complex scenarios.
  • Probabilistic Approach and Heatmaps: Instead of a simple binary classification, PalmProbNet generates probability heatmaps. These heatmaps visually represent the likelihood of palm presence across the entire orthomosaic image, offering a more nuanced understanding of palm distribution and abundance.
  • Sliding Window and Ensemble: A sliding window technique is applied to the orthomosaic image to extract patches. The predictions from both small and large patch models are then aggregated and averaged to create a more robust probability estimate for each region.

Datasets

The study utilizes UAV-derived orthomosaic imagery from a natural reserve in the Northwest region of Ecuador, known as the Ecuadorian Choco forest.

  • Raw Data: High-resolution UAV imagery collected using a DJI Phantom 4 RTK drone.
  • Orthomosaic: A photogrammetrically orthorectified image product generated from the raw imagery.
  • Labeled Data:
    • Fine-scale labeled data: 6,000 small (40x40 px) patches of palm features and 6,000 small patches of non-palm features.
  • Coarse-scale labeled data: Approximately 3,367 palm and 4,091 non-palm patches of larger (100x100 px) size, generated through a semi-manual process involving initial detection with small patches.

Key Insights and Ideas

  • Dual-scale patching is crucial: The combination of small patches for detailed features and large patches for context significantly improves detection accuracy, especially in dense and complex forest environments.
  • Probabilistic heatmaps offer richer information: Moving beyond simple detection, probability heatmaps provide a continuous representation of palm distribution, which is more informative for ecological studies and resource management than discrete bounding boxes.
  • Transfer learning is highly effective for limited data: Leveraging pre-trained models like ResNet-18 dramatically reduces the annotation burden and computational cost, making advanced deep learning feasible for specialized remote sensing tasks.
  • UAVs are ideal for high-resolution forest monitoring: The high spatial resolution and low altitude operation of UAVs provide the necessary detail for fine-grained analysis of forest canopies, complementing traditional satellite imagery.

Key Results

PalmProbNet demonstrated remarkable performance on the Ecuadorian tropical forest dataset:

  • Accuracy: Achieved an accuracy of 97.32% and a Cohen's κ of 94.59% on the testing set.
  • Model Configurations: The best performance was observed with a 256-node configuration for the MLP, particularly for the large patch size, yielding an accuracy of 0.9732 and Cohen's κ of 0.9459.
  • Heatmap Visualization: The generated probability heatmaps effectively visualized palm distributions, showcasing the scalability and adaptability of the approach across different forest densities.

Limitations and Future Work

  • Limitations:
    • The paper notes that some non-palm tree crowns were erroneously classified as palms, and areas with poor image quality or missing data (artifacts from image processing) were prone to misclassification.
  • While the 256-node configuration showed a slight edge, the performance across different node configurations was broadly similar, suggesting potential for further fine-tuning.
  • Future Work:
    • Refining heatmaps by including training samples with edge cases and employing post-processing to filter out patches near blank areas.
  • Exploring segmentation networks using the generated heatmaps.
  • Investigating superpixel-based methods for faster and more precise localization of individual palms.
  • Integrating UAV imagery with satellite data (e.g., Planet, WorldView) to leverage the complementary strengths of local detail and broad-scale context for improved detection accuracy.

Summary updated

Abstract

Palms play an outsized role in tropical forests and are important resources for humans and wildlife. A central question in tropical ecosystems is understanding palm distribution and abundance. However, accurately identifying and localizing palms in geospatial imagery presents significant challenges due to dense vegetation, overlapping canopies, and variable lighting conditions in mixed-forest landscapes. Addressing this, we introduce PalmProbNet, a probabilistic approach utilizing transfer learning to analyze high-resolution UAV-derived orthomosaic imagery, enabling the detection of palm trees within the dense canopy of the Ecuadorian Rainforest. This approach represents a substantial advancement in automated palm detection, effectively pinpointing palm presence and locality in mixed tropical rainforests. Our process begins by generating an orthomosaic image from UAV images, from which we extract and label palm and non-palm image patches in two distinct sizes. These patches are then used to train models with an identical architecture, consisting of an unaltered pre-trained ResNet-18 and a Multilayer Perceptron (MLP) with specifically trained parameters. Subsequently, PalmProbNet employs a sliding window technique on the landscape orthomosaic, using both small and large window sizes to generate a probability heatmap. This heatmap effectively visualizes the distribution of palms, showcasing the scalability and adaptability of our approach in various forest densities. Despite the challenging terrain, our method demonstrated remarkable performance, achieving an accuracy of 97.32% and a Cohen's κ of 94.59% in testing.
BibTeX
@inproceedings{cui2024palmprobnet,
  author = {Kangning Cui and Zishan Shao and Gregory Larsen and Victor Paúl Pauca and Sarra M. Alqahtani and David Segurado and João Pinheiro and Manqi Wang and David A. Lutz and Robert J. Plemmons and Miles Silman},
  title = {{PalmProbNet: A Probabilistic Approach to Understanding Palm Distributions in Ecuadorian Tropical Forest via Transfer Learning}},
  booktitle = {ACM Southeast Regional Conference},
  year = {2024},
  doi = {10.1145/3603287.3651220},
}
Figure