🤖 AI

Tighnari v2: Mitigating Label Noise and Distribution Shift in Multimodal Plant Distribution Prediction via Mixture of Experts and Weakly Supervised Learning

Tighnari v2 proposes a multimodal fusion framework that leverages both sparse, high-quality Presence-Absence data and abundant, noisy Presence-Only data for plant distribution prediction by utilizing a geographic-based pseudo-labeling strategy and a Mixture-of-Experts approach to mitigate label noise and distribution shifts.

Original authors: Haixu Liu, Yufei Wang, Tianxiang Xu, Chuancheng Shi, Hongsheng Xing

Published 2026-02-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Haixu Liu, Yufei Wang, Tianxiang Xu, Chuancheng Shi, Hongsheng Xing

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

Imagine you are trying to create a "Master Map" of every plant species in Europe. To do this, you have two types of helpers, but both have flaws.

The Two Types of Helpers

The Professional Scientists (PA Data): These are like highly trained detectives. They go to a specific patch of forest, look under every rock, and write down exactly what they see. If they don't see a rare orchid, they can confidently say, "It’s not here." This data is perfect but tiny. Because it’s so expensive and slow, they only have a few "detective reports."
The Citizen Scientists (PO Data): These are like thousands of hikers with smartphone apps. They take amazing photos of beautiful flowers, but they often ignore "boring" weeds or common grass. If a hiker doesn't take a photo of a plant, it doesn't mean the plant isn't there—it just means the hiker didn't notice it. This data is huge but messy. It covers the whole continent, but it’s full of "false negatives" (missing things that actually exist).

The Problem: If you try to mix the "Perfect Detective Reports" with the "Messy Hiker Photos" into one big pile, the messiness of the hiker data "pollutes" the accuracy of the detectives. Your map becomes blurry and unreliable.

The Solution: The "Tighnari v2" Strategy

The researchers created a smart system to use the best of both worlds without the pollution. Here is how they did it:

1. The "Smart Grouping" Trick (Pseudo-Labeling)

Instead of trusting every single hiker's "missing" plant list, the researchers looked at satellite images. They said: "If ten different hikers all took photos in this same 600-meter square, let's treat that square like a mini-detective report." By grouping hikers together, they turned messy individual observations into a much more reliable "neighborhood report."

2. The "Multimodal Brain" (The Architecture)

The model doesn't just look at one thing; it has different "senses" working together:

Eyes: It looks at satellite photos (Swin Transformer).
Memory: It looks at how the weather and colors change over time (Temporal Swin Transformer).
Facts: It looks at spreadsheets of soil and climate data (TabM).
The Fusion: They built a "Serial Cross-Attention" mechanism. Think of this like a relay race. Instead of all the senses shouting at once, the "Eyes" pass their info to the "Memory," which then passes its updated info to the "Facts." This ensures the most important information isn't drowned out by the noise.

3. The "Expert Committee" (Mixture of Experts)

The researchers noticed a huge problem: the "Detectives" mostly worked in Western Europe, but the "Test Map" required predicting plants in Eastern Europe (where the detectives had never been). This is called a Distribution Shift.

To fix this, they used a Mixture of Experts (MoE) approach. Imagine you have two specialists:

Expert A (The Local): Trained on the high-quality, perfect detective data.
Expert B (The Explorer): Trained on the massive, messy hiker data to learn what plants look like in new territories.

When a new location comes up for prediction, the system asks: "Is this place near a detective's old route?"

If YES, it lets Expert A take the lead because we want perfect accuracy.
If NO, it lets Expert B take the lead because even though they are a bit messy, they actually know what the rest of the world looks like.

The Result

By using this "Expert Committee" and cleaning up the hiker data, the model didn't just perform well—it actually beat previous records. It was able to predict plants in entirely new geographic regions where it had never seen "perfect" data before, making it a much more powerful tool for protecting biodiversity.

Technical Summary: Tighnari v2: Mitigating Label Noise and Distribution Shift in Multimodal Plant Distribution Prediction

1. Problem Statement

Predicting plant species distribution is vital for biodiversity conservation but is hindered by two primary data constraints:

Data Imbalance and Noise: Presence–Absence (PA) data are high-quality and noise-free but expensive and sparse. Presence-Only (PO) data (from citizen science) are abundant and provide great spatial coverage but suffer from severe label noise; specifically, "negative" samples in PO data are unreliable because observers often ignore common or unrecognized species rather than confirming their absence.
Distribution Shift: There is a significant geographic out-of-distribution (OOD) gap between training data and test data. Models trained on specific regions (e.g., Western Europe) struggle to generalize to new regions (e.g., Eastern Europe) due to ecological differences.
Multimodal Fusion Challenges: Integrating heterogeneous data (satellite imagery, tabular environmental data, and time-series bioclimatic data) is difficult because different modalities contribute unequally to predictive accuracy, and standard fusion methods often dilute the influence of dominant modalities.

2. Methodology

The authors propose an enhanced multimodal framework that leverages both PA and PO data through several innovative components:

Weakly Supervised Pseudo-Labeling: To mitigate PO label noise, the authors implement a geographic aggregation strategy. They aggregate species labels from all PO survey samples located within the same $640\text{m} \times 640\text{m}$ satellite image patch. This effectively transforms sparse PO data into a "pseudo-quadrat" survey, reducing negative label noise without introducing false positives. They also use a reserved-sample strategy to ensure rare species are not lost during aggregation.
Advanced Backbone Architectures:
- Satellite Imagery: Upgraded to a Swin Transformer Base to better capture low-frequency environmental information.
- Tabular Data: Utilizes the TabM network (based on BatchEnsemble) for superior feature representation of environmental variables.
- Time-Series: Retains a Temporal Swin Transformer for modeling seasonal bioclimatic changes.
Stackable Serial Tri-Modal Cross-Attention: Instead of simple concatenation, the authors designed a serial attention mechanism. It uses a shared Key/Value ($KV$) projection across all modalities to unify semantic spaces and an iterative update process (Modality A updates B and C, then B updates A and C, etc.). This allows the model to capture higher-order dependencies and prevents dominant modalities from being diluted.
Asymmetric Loss (ASL): To handle the extreme class imbalance (many more absent species than present) and noisy negatives, they replaced Binary Cross-Entropy with ASL. ASL uses asymmetric focusing parameters and negative sample clipping to prevent noisy "absent" labels from dominating the gradient.
Mixture of Experts (MoE) for Distribution Shift: To handle OOD test samples, the authors partition test samples based on their proximity to PA training sites. They train two specialized models:
1. An In-Distribution model trained on high-quality PA data.
2. An Out-of-Distribution model trained via a two-stage process (pre-training on pseudo-labeled PO data, then fine-tuning on clean PA data).

3. Key Contributions

A novel pseudo-labeling rule that aligns PO data with the geographic coverage of remote sensing features, minimizing label noise.
A stackable serial tri-modal cross-attention module that enhances the fusion of heterogeneous environmental modalities.
A two-stage training procedure that balances the broad coverage of PO data with the precision of PA data.
An MoE-inspired inference strategy that uses geographic proximity to select the most appropriate model for a given test sample, effectively mitigating distribution shifts.

4. Results

Superior Performance: The model achieved high rankings in the GeoLifeCLEF 2025 challenge.
OOD Robustness: The MoE approach significantly outperformed single-model approaches in OOD scenarios. While a model trained on both PA and PO data suffered from noise, the MoE model maintained high accuracy for both in-distribution and out-of-distribution samples.
Benchmark Comparison: The MoE approach surpassed the 2nd-place score of the 2024 leaderboard (where data was identically distributed) and demonstrated superior performance on the 2025 leaderboard (where geographic shifts were pronounced).

5. Significance

This work provides a robust blueprint for applying deep learning to real-world ecological monitoring. By treating label noise and distribution shift as primary architectural concerns rather than secondary issues, the authors demonstrate how to effectively bridge the gap between professional scientific surveys (PA) and massive, noisy crowdsourced data (PO). This is highly significant for large-scale biodiversity conservation efforts and climate change impact assessments.