🤖 AI

FLORA: A deep learning approach to predict forest attributes from heterogeneous LiDAR data

The paper introduces FLORA, a deep learning framework that leverages an octree-based backbone and a late-fusion gating mechanism to robustly predict six key forest attributes from heterogeneous, multi-condition LiDAR data across France, achieving high accuracy and improved cross-season performance compared to existing models.

Original authors: Emilie Vautier, Clément Mallet, Cédric Vega

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Emilie Vautier, Clément Mallet, Cédric Vega

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 guess the weight, height, and number of trees in a forest just by looking at a 3D digital map made of laser dots (called LiDAR). This is what the paper FLORA is about.

The researchers wanted to build a "super-forecaster" that could look at these laser maps from all over France and tell us exactly how much wood is in the trees, how tall they are, and how many there are. But there was a huge problem: the laser maps weren't uniform. Some were taken in summer when trees were full of leaves, some in winter when they were bare, some by fast planes, some by slow ones, and some from high up, some from low down.

It's like trying to learn to recognize a friend's face by looking at photos taken in different lighting, with different filters, and sometimes with them wearing a hat and sometimes without. Most computer programs get confused by this and fail when the conditions change.

Here is how the FLORA team solved it, using simple analogies:

1. The Problem: The "Seasonal Amnesia"

The paper found that if you train a computer model only on summer photos (leaf-on), it gets terrible at guessing the forest's structure in winter (leaf-off), and vice versa.

The Analogy: Imagine a student who only studies for a math test using a calculator. If you take the calculator away on test day, they fail. Similarly, a model trained only on "leafy" data doesn't know how to interpret "bare" trees, and a model trained on "bare" trees gets confused by the dense "leafy" canopy.
The Result: When they tried to use a "summer-only" model on winter data, the errors exploded. The model would guess the forest was huge when it was actually small, or vice versa.

2. The Solution: The "All-Seasons Chef"

Instead of training separate chefs for summer and winter, the researchers trained one single model (FLORA) on both summer and winter data at the same time.

The Analogy: Think of this model as a master chef who has tasted the soup in both summer and winter. They know that the ingredients change (leaves appear or disappear), but the core recipe (the tree trunks and branches) stays the same.
The Magic: By feeding the model mixed data, it learned to ignore the "noise" of the seasons and focus on the permanent structure of the forest. This single model worked better than any of the specialized "season-only" models, even when tested on the specific season it wasn't "favorite" for.

3. The Secret Sauce: "Context Clues"

The model doesn't just look at the laser dots; it also gets a "cheat sheet" of extra information.

The Analogy: Imagine you are trying to guess the price of a house. You could look at the photos (the laser dots), but it helps to know the neighborhood, the year it was built, and the type of soil (the auxiliary data).
The Gating Mechanism: The paper describes a clever "gatekeeper" inside the AI. For some questions (like "How tall is the tree?"), the gatekeeper says, "Ignore the cheat sheet, just look at the laser dots." But for other questions (like "How much wood is in the pine trees specifically?"), the gatekeeper says, "Look at the cheat sheet! The location and season tell us a lot about pine trees."
The Result: This helped the model get much better at guessing specific types of wood (like pine vs. oak), which is usually very hard to do with just laser dots.

4. The Overlap Bonus: "The Double-Check"

Sometimes, the same forest plot was scanned twice: once in summer and once in winter.

The Analogy: This is like having a friend describe a room to you, and then having them describe it again after the furniture was moved. You can compare the two descriptions to understand the room's true shape better.
The Result: Using these "double-scanned" plots helped the model get better at guessing the volume of broadleaf trees (like oaks), but surprisingly, it didn't help much with pine trees. The researchers suspect this is because pine trees look similar year-round, so the "double-check" didn't add much new info, whereas broadleaf trees change so much that seeing both versions was a huge help.

The Bottom Line

The paper concludes that you don't need to build a different model for every season, sensor, or forest type. If you have enough data and train one big, robust model (FLORA) on a huge, messy mix of everything, it learns to handle the chaos on its own.

What it got right: It predicted tree height very accurately (almost 90% accuracy) and total wood volume quite well.
What was hard: It still struggles a bit with counting exactly how many individual tree trunks there are (stem density) and separating the exact volume of pine vs. oak, though the "cheat sheet" data helped improve those guesses.

In short, FLORA is a deep-learning tool that proves you can build a "universal translator" for forests, capable of reading the messy, inconsistent laser maps of a whole country and turning them into reliable forest statistics, without needing to manually fix the data first.

Technical Summary: FLORA – A Deep Learning Approach for Forest Attribute Prediction from Heterogeneous LiDAR Data

Problem Statement
National Forest Inventories (NFIs) rely on auxiliary remote sensing data, particularly airborne LiDAR, to estimate forest attributes at a national scale. However, operational "wall-to-wall" prediction remains challenging due to the heterogeneity of LiDAR acquisition conditions. National programs often involve varying flight altitudes, sensor types, scan angles, and, critically, phenological conditions (leaf-on vs. leaf-off). Existing models, often trained on local data with uniform acquisition parameters, struggle to generalize across these diverse conditions. Specifically, applying models trained on leaf-on data to leaf-off acquisitions (and vice versa) introduces significant bias and error, particularly in deciduous stands where foliage presence drastically alters point cloud structure. Furthermore, traditional machine learning approaches relying on hand-crafted statistical features often fail to capture the full 3D structural complexity of forests or to learn robust, transferable representations across heterogeneous datasets.

Methodology
The authors propose FLORA (Forest LiDAR Octree Regression with Auxiliary Data), a deep learning framework designed to predict six forest attributes (dominant height, total volume, deciduous volume, coniferous volume, basal area, and stem density) directly from heterogeneous LiDAR point clouds.

Architecture: FLORA utilizes a voxel-based approach, converting raw point clouds into octree structures to preserve 3D structural information while maintaining computational efficiency. The backbone is an Octree Convolutional Neural Network (OCNN) adapted with a High-Resolution Net (HRNet) architecture. This design processes feature maps at multiple spatial resolutions in parallel, allowing the model to capture both fine-grained details and coarse structural context.
Input Data: The model operates primarily on the 3D spatial coordinates (XYZ) of points. While intensity and echo attributes were considered, they were excluded during hyperparameter optimization as they did not yield consistent performance gains and introduced sensor-specific noise.
Auxiliary Data Integration: To address acquisition heterogeneity, FLORA incorporates ecological and spatiotemporal auxiliary variables (e.g., sylvo-ecoregion, canopy cover class, acquisition date, and temporal gap to field survey). These are encoded via learnable embeddings and concatenated with geometric features.
Fusion Mechanism: A "late-fusion" strategy is employed where auxiliary data is combined with LiDAR features. Crucially, a target-specific gating mechanism is introduced. This learns a coefficient ( $g_k$ ) for each prediction target to dynamically interpolate between geometry-driven features and fused representations. This allows the model to adaptively weigh contextual information based on the specific attribute being predicted, providing a degree of interpretability regarding the contribution of auxiliary data.
Training Strategy: The model was trained on 32,052 National Forest Inventory (NFI) plots across mainland France, covering diverse forest types and acquisition conditions (both leaf-on and leaf-off). The training involved systematic hyperparameter optimization using Bayesian optimization (Optuna) to adapt the architecture to the specific regression task.

Key Contributions

Heterogeneity Handling: FLORA demonstrates that a single deep learning model trained jointly on data acquired under contrasting phenological conditions (leaf-on and leaf-off) can generalize effectively across seasons and forest types, outperforming season-specific models.
Robust National-Scale Baseline: The study establishes a robust baseline for large-scale forest attribute estimation using a uniquely large and heterogeneous dataset (French LiDAR HD program), addressing the gap in existing literature regarding generic models for national-scale, multi-condition data.
Interpretable Fusion: The introduction of the gating mechanism allows for the quantification of how much contextual auxiliary data contributes to specific predictions, revealing that the utility of such data is attribute-dependent.
Phenological Insights: The research quantifies the asymmetry in cross-season generalization, showing that leaf-off models degrade severely when applied to leaf-on data, whereas leaf-on models transfer more robustly to leaf-off conditions.

Results

Performance: The full model achieved strong performance for dominant height (rRMSE ≈ 12.3%, $R^2$ ≈ 0.88) and competitive results for total volume (rRMSE ≈ 39%, $R^2$ ≈ 0.74). Stem density remained the most challenging attribute ( $R^2$ < 0.47), consistent with broader literature.
Seasonal Generalization: A single model trained on both seasons outperformed models trained on specific seasons, even within the calibration domain of those seasons. Cross-season testing revealed that leaf-off models applied to leaf-on data suffered severe degradation (e.g., volume rRMSE increased by ~24%), while leaf-on models applied to leaf-off data maintained acceptable performance.
Class Imbalance: Balancing the training set (enforcing equal leaf-on/leaf-off representation) resulted in only moderate performance degradation, suggesting the original imbalance did not critically bias the model.
Overlapping Acquisitions: Using plots surveyed in both seasons as a training strategy improved predictions for total and deciduous volume but degraded performance for coniferous volume, likely due to the limited geographic diversity of the overlap subset.
Auxiliary Data Impact: Auxiliary variables provided modest overall gains but revealed attribute-dependent contributions. They significantly improved predictions for species-specific volumes (coniferous and deciduous) and stabilized cross-season generalization. However, they offered negligible improvement for dominant height and sometimes degraded stem density predictions in same-season contexts.

Significance
The paper positions FLORA as a significant step toward operational, large-scale forest monitoring. It demonstrates that deep learning can effectively harmonize heterogeneous LiDAR data without requiring explicit stratification by season, sensor, or forest type. The findings suggest that training on diverse, national-scale datasets is a viable strategy to overcome the limitations of local, condition-specific models. The work highlights that while species-specific volume prediction remains challenging, a generic DL approach can provide robust estimates for key structural attributes, supporting the integration of national LiDAR programs into Multisource Forest Inventories (MSNFIs). The study concludes that the ability to handle acquisition heterogeneity is critical for future national-scale applications where data coverage is spatially and temporally fragmented.