Monitoring early-successional trees for tropical forest restoration using low-cost UAV-based species classification

Jonathan Williams, Toby D. Jackson, Carola‐Bibiane Schönlieb, Tom Swinfield, Bambang Irawan, Eva Achmad, Muhammad Zudhi, Habibi Habibi, Elva Gemita, David A. CoomesFrontiers in Forests and Global Change 2022 · DOI

TL;DR

Researchers have developed SLIC-UAV, a low-cost monitoring pipeline that uses consumer-grade drones and "superpixel" segmentation to identify tree species in restoring tropical forests. By focusing on key early-successional species as indicators, the system achieves up to 91.7% accuracy, providing a scalable alternative to expensive field surveys or high-end aerial sensors.

What Problem Does This Paper Address?

Tropical forest restoration is critical for carbon sequestration and biodiversity, but we are currently terrible at monitoring it at scale. Logged forests cover roughly four million square kilometers of the tropics. While we can use satellites to see if a forest is "green," it is incredibly difficult to tell what is growing there.

Traditional field surveys are prohibitively expensive and slow in dense tropical terrain. On the other hand, high-end remote sensing (like aerial LiDAR or hyperspectral imaging) requires specialized aircraft and million-dollar sensors. There is a massive "missing middle": a need for a low-cost, high-resolution way to identify specific tree species that indicate whether a forest is actually recovering or just being overtaken by invasive weeds.

Methods

The authors introduce the SLIC-UAV pipeline, which stands out by moving away from traditional pixel-by-pixel analysis.

  1. Data Acquisition: They used standard consumer-grade UAVs (3DR Solo and DJI Phantom 4) to collect RGB, Multispectral (MS), and Digital Surface Model (DSM) data over a 100-hectare site.
  2. Superpixel Segmentation: Instead of classifying individual pixels, they used Simple Linear Iterative Clustering (SLIC) to group neighboring pixels into "superpixels"—blobs of roughly 0.5 $m^2$ that share similar colors and textures. This massively reduces the computational load and allows the model to "see" patterns.
  3. Feature Extraction: For every superpixel, the team extracted both spectral features (color, brightness, vegetation indices) and textural features (using tools like Gray-Level Co-occurrence Matrices and Local Binary Patterns). Texture is key here; it captures the "jaggedness" or "smoothness" of different leaf arrangements.
  4. Machine Learning: They compared three models—Lasso Regression, Random Forests (RF), and Support Vector Machines (SVM)—to classify the superpixels into species categories.

Datasets

The study was conducted at Hutan Harapan in Sumatra, Indonesia, an ecosystem restoration concession.

  • Training Data: 11,996 manually labeled superpixels derived from 328 individual tree crowns identified in the field with local experts.
  • Target Species: Four key species were tracked:
  • Bellucia pentamera (Invasive pioneer)
    • Macaranga gigantea (Early-successional)
    • Alstonia scholaris (Long-lived)
    • Endospermum malaccense (Long-lived)
  • Additional Classes: Oil palm, "other" trees, and non-vegetation (roads, buildings).

Key Insights and Ideas

  • Texture Trumps Color: One of the "aha" moments was that adding textural features significantly improved accuracy. In a tropical canopy, many trees share similar shades of green, but their leaf structures and branching patterns (texture) are distinct signatures.
  • Superpixels as a Bridge: By using superpixels, the researchers bridged the gap between noisy pixel-level data and complex crown-level geometry. This allowed the model to be flexible with regards to the spatial distribution of training data—essential when working in unpredictable field conditions.
  • Indicator Species Strategy: Rather than trying to map every one of the 600+ tree species in the region, the authors focused on "indicator" species. By tracking early-successional trees vs. long-lived ones, they could create a "heat map" of forest maturity across 100 hectares.
  • The "Low-Cost" Win: The entire setup used a ~$3,000 camera system, proving that sophisticated forest management doesn't always require government-level budgets.

Key Results

  • High Accuracy: The SVM model performed best. It achieved 74.3% accuracy for the full four-species task, which jumped to 91.7% when the model focused specifically on the most distinct early-successional species.
  • Structural Value: Adding DSM (height) data improved accuracy across the board, as it helped the model distinguish between low-lying invasive shrubs and tall, established canopy trees.
  • Scalability: The team successfully extended the model to produce wall-to-wall dominance maps for the entire 100-hectare project area, identifying exactly where restoration was lagging (dominated by Bellucia) versus where it was thriving.

Limitations and Future Work

The authors acknowledge a few hurdles:

  • Visual Similarity: Long-lived species (Alstonia and Endospermum) look very similar from above, leading to more frequent misclassifications between them.
  • Overfitting: While multispectral data helped, it also introduced a risk of overfitting if not carefully calibrated for lighting conditions.
  • Manual Labor: The pipeline still requires manual labeling of tree crowns by experts to train the initial model.

Future Work: The researchers suggest that Convolutional Neural Networks (CNNs) might further improve results by automatically learning even more complex features, though this would require a much larger labeled dataset than currently available for most restoration projects.

Summary updated

Abstract

Logged forests cover four million square kilometers of the tropics, capturing carbon more rapidly than temperate forests and harboring rich biodiversity. Restoring these forests is essential to help avoid the worst impacts of climate change. Yet monitoring tropical forest recovery is challenging. We track the abundance of early-successional species in a forest restoration concession in Indonesia. If the species are carefully chosen, they can be used as an indicator of restoration progress. We present SLIC-UAV, a new pipeline for processing Unoccupied Aerial Vehicle (UAV) imagery using simple linear iterative clustering (SLIC)to map early-successional species in tropical forests. The pipeline comprises: (a) a field verified approach for manually labeling species; (b) automatic segmentation of imagery into “superpixels” and (c) machine learning classification of species based on both spectral and textural features. Creating superpixels massively reduces the dataset's dimensionality and enables the use of textural features, which improve classification accuracy. In addition, this approach is flexible with regards to the spatial distribution of training data. This allowed us to be flexible in the field and collect high-quality training data with the help of local experts. The accuracy ranged from 74.3% for a four-species classification task to 91.7% when focusing only on the key early-succesional species. We then extended these models across 100 hectares of forest, mapping species dominance and forest condition across the entire restoration project.
BibTeX
@inproceedings{williams2022monitoring,
  author = {Jonathan Williams and Toby D. Jackson and Carola‐Bibiane Schönlieb and Tom Swinfield and Bambang Irawan and Eva Achmad and Muhammad Zudhi and Habibi Habibi and Elva Gemita and David A. Coomes},
  title = {{Monitoring early-successional trees for tropical forest restoration using low-cost UAV-based species classification}},
  booktitle = {Frontiers in Forests and Global Change},
  year = {2022},
  doi = {10.3389/ffgc.2022.876448},
}
Figure