🤖 machine learning

Explainable few-shot learning workflow for detecting invasive and exotic tree species

This paper proposes an explainable few-shot learning workflow that integrates a Siamese network with XAI to effectively detect invasive and exotic tree species in Brazil's Atlantic Forest using UAV images, achieving high accuracy with minimal labeled data while providing transparent visual explanations for its predictions.

Original authors: Caroline M. Gevaert, Alexandra Aguiar Pedro, Ou Ku, Hao Cheng, Pranav Chandramouli, Farzaneh Dadrass Javan, Francesco Nattino, Sonja Georgievska

Published 2026-09-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Caroline M. Gevaert, Alexandra Aguiar Pedro, Ou Ku, Hao Cheng, Pranav Chandramouli, Farzaneh Dadrass Javan, Francesco Nattino, Sonja Georgievska

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the dense, vibrant Atlantic Forest of Brazil, a silent battle is underway between native trees and invasive species that threaten to displace them. For decades, ecologists have relied on walking the forest floor to identify these intruders, a slow and labor-intensive process that leaves vast areas unmonitored. Today, the tools of observation have shifted skyward. Unmanned aerial vehicles, or drones, can now capture high-resolution photographs of the forest canopy, offering a bird's-eye view that reveals the shape and color of individual trees. However, turning these images into useful knowledge requires artificial intelligence. The challenge lies in training these computer systems. Most modern AI models are like students who need to study thousands of examples of every subject before they can pass a test. In remote forests, where data is scarce and experts are few, gathering such massive libraries of labeled images is often impossible. Furthermore, even when these models make a guess, they often act as "black boxes," offering a result without explaining how they reached it, which makes it hard for forest managers to trust the answer.

A team of researchers has developed a new approach to bridge this gap, creating a system that can learn to spot invasive trees with very few examples and, crucially, show its work. Published in a study focusing on the Bororé Natural Park in São Paulo, the research introduces a workflow that combines a specific type of artificial intelligence with a method for visual explanation. Instead of requiring thousands of photos to learn what an invasive palm tree looks like, the system uses a "few-shot" learning strategy. This allows the model to be trained on a handful of known species first, and then quickly adapted to recognize new, rare, or invasive species using only a few sample images. To ensure the system is not just guessing, the researchers added a layer of transparency. When the computer identifies a tree, it does not just output a name; it displays the specific images from its memory that led to that conclusion, allowing a human expert to verify if the visual match makes sense.

The study focused on the Seafórtia palm, an invasive Australian species that has taken root in the Brazilian parks, competing with native flora and altering the local ecosystem. To build their system, the researchers started by gathering a small set of training data. They used existing datasets and new drone flights to identify five common tree species, creating a library of "known" examples. They then trained a neural network—a computer program designed to mimic the way the brain processes visual patterns—on these five species. The researchers tested two different versions of this network: one built from scratch with simple layers, and another based on a lightweight, pre-existing design called MobileNet, which is known for being efficient on smaller computers. Both versions were designed to compare two images and decide how similar they are, rather than simply memorizing a list of species.

Once the system was pre-trained, the researchers introduced the real challenge: identifying the invasive Seafórtia palm and other new species using only a tiny number of examples. This is the "few-shot" part of the process. They showed the system just one, two, or three photos of the new target species and asked it to learn the difference. The results were striking. The system using the MobileNet backbone, which had been pre-trained on a vast library of general images, proved far more effective than the simpler version. When given just three examples of a new species, the MobileNet-based system correctly identified the invasive trees with a high degree of accuracy, significantly outperforming the simpler model. This demonstrated that a sophisticated, pre-trained system could adapt to a new, data-poor environment much faster than starting from zero.

Perhaps the most significant contribution of this work, however, is how it handles the "why" behind the decision. In many artificial intelligence applications, the reasoning is hidden. Here, the researchers made the reasoning visible. When the system classifies a tree, it pulls up the few example images it used to make that decision and displays them alongside the new tree. It also provides a set of metrics to evaluate the quality of this explanation. For instance, it checks for "correctness" to see if the examples it chose actually support the final label, and "continuity" to ensure that a slight change in the input image doesn't cause the system to pick a completely different set of examples. The study found that the explanations were highly consistent and that the system often selected support images that were visually very similar to the target, giving human experts a clear, visual trail to follow.

The researchers tested their workflow on the invasive Seafórtia palms found in the Bororé Natural Park. In one specific test case, the system correctly identified a palm tree and showed the user the three reference images it used for comparison. In some instances, the system flagged a potential error by showing a reference image that looked quite different from the target, even though they were labeled as the same species. This visual feedback allowed the researchers to see exactly where the model was struggling, something a standard "black box" model would never reveal. The study concluded that this combination of few-shot learning and visual explanation offers a practical, reliable tool for forest management. It proves that it is possible to deploy advanced AI in conservation efforts even when labeled data is scarce, provided the system can explain its logic in a way that humans can understand and verify. This approach opens a new path for using drones and artificial intelligence to protect biodiversity, turning the vast, complex data of the forest into actionable insights for conservationists.

Technical Summary: Explainable Few-Shot Learning Workflow for Detecting Invasive and Exotic Tree Species

Problem Statement
Deep learning models for Earth Observation, particularly for identifying invasive and exotic tree species, typically require extensive labeled datasets for training and evaluation. This dependency creates a significant barrier in practical scenarios, especially in Low and Middle-Income Countries (LMICs) where labeled data is scarce. Furthermore, existing few-shot learning (FSL) approaches often fail to provide sufficient explanations for their predictions, leaving users without confidence in the model's decision-making process. The specific challenge addressed in this study is the detection of invasive species (specifically Archontophoenix cunninghamiana, or Seafórtia) in the Atlantic Forest of São Paulo, Brazil, where manual field inspections are labor-intensive and data availability is limited.

Methodology
The authors propose an end-to-end workflow integrating Unmanned Aerial Vehicle (UAV) imagery, few-shot learning, and Explainable AI (XAI). The methodology consists of three primary processing steps:

Data Preparation and Pre-training:
- Input: UAV orthomosaics and a benchmark dataset (Reforestree).
- Object Detection: Individual tree canopies are automatically extracted from UAV imagery using a Netflora workflow or manually labeled. These are combined with benchmark data (banana, cacao, fruit species) to form a base training set.
- Model Architecture: Two Siamese networks are pre-trained: a shallow Convolutional Neural Network (CNN) and a lightweight MobileNet backbone (pre-trained on ImageNet). Both networks utilize a similarity metric (Euclidean distance) to compare image pairs.
- Training Strategy: The networks are trained on balanced pairs of similar (same species) and dissimilar (different species) cutouts (128x128 pixels) generated via data augmentation.
Few-Shot Learning Refinement:
- Scenario: The pre-trained models are refined to recognize new target species (invasive and native) using a minimal number of labeled samples (1-shot to 3-shot).
- Process: The base models are fine-tuned using a small support set of manually labeled samples from the study area.
- Classification: Two prediction methods are evaluated: (1) assigning the class with the highest average similarity score, and (2) using K-Nearest-Neighbor (K-NN) logic based on the top $K$ similarity scores.
Explainable AI (XAI) Integration:
- Case-Based Explanation: Instead of generating heatmaps (e.g., Grad-CAM), the workflow provides visual, case-based explanations. For a query image, the system displays the most similar support set images from the training data.
- Evaluation Metrics: The quality of these explanations is quantitatively evaluated using three Co-12 content properties:
  - Correctness ( $C_{cor}$ ): Measures if the explanation (majority class of similar samples) aligns with the model's final classification.
  - Continuity ( $C_{cty}$ ): Assesses stability by checking if the explanation changes when the input image is slightly perturbed.
  - Contrastivity ( $C_{cst}$ ): Quantifies the diversity of support images selected for explanations using Shannon Entropy.

Key Contributions

Integrated Workflow: The paper presents a novel workflow that combines Siamese networks for few-shot learning with a case-based XAI strategy, addressing both the data scarcity and interpretability challenges simultaneously.
Lightweight Architecture: The study demonstrates that a MobileNet-based Siamese network, while larger than a shallow CNN, remains lightweight enough for deployment on devices with limited computational capacity, achieving superior performance.
Quantitative XAI Evaluation: The authors apply a rigorous framework (Co-12) to quantitatively evaluate the quality of example-based explanations, moving beyond qualitative visual inspection.

Results

Base Model Performance: The MobileNet-based Siamese network significantly outperformed the shallow CNN in terms of Recall, F1-score, and Accuracy, achieving an F1-score of 0.97 and Accuracy of 0.97 on the base classification task.
Few-Shot Learning: Refinement with few-shot learning improved performance over zero-shot learning. The MobileNet-based model achieved the best results with 3-shot learning, reaching a Precision of 0.90 and an F1-score of 0.86. The performance gap between MobileNet and the shallow CNN widened as the number of support samples increased (from 1-shot to 3-shot).
Explanation Quality:
- Correctness: Improved significantly after model refinement (e.g., MobileNet $C_{cor}$ increased from 0.15 to 0.60 in 3-shot learning).
- Continuity: Both models showed high continuity (values around 0.80–0.87), indicating stable explanations under minor input perturbations.
- Contrastivity: High values (near 1.00) were observed, indicating diverse selection of support images for explanations.
- Visual Case Studies: The paper provides visual examples showing that when the model makes a prediction, the associated support images often visually corroborate the decision, though false positives can be identified by low similarity scores in the explanation list.

Significance
The paper claims that this approach opens new avenues for using AI and UAVs in forest management and biodiversity conservation, particularly for rare or under-studied species. By enabling the classification of tree species with minimal labeled data while providing visual, case-based explanations, the workflow addresses the practical limitations of data availability in LMICs. The study emphasizes that example-based explanations align better with human reasoning (contrastive and simple) than complex saliency maps, thereby increasing user confidence in automated forest monitoring systems. The use of lightweight backbones like MobileNet further suggests the feasibility of deploying these solutions in resource-constrained environments.

More like this