Multimodal classification of forest biodiversity potential from 2D orthophotos and 3D airborne laser scanning point clouds

Assessment of forest biodiversity is crucial for ecosystem management and conservation. While traditional field surveys provide high-quality assessments, they are labor-intensive and spatially limited. This study investigates whether deep learning-based fusion of close-range sensing data from 2D orthophotos and 3D airborne laser scanning (ALS) point clouds can reliable assess the biodiversity potential of forests. We introduce the BioVista dataset, comprising 44378 paired samples of orthophotos and ALS point clouds from temperate forests in Denmark, designed to explore multimodal fusion approaches. Using deep neural networks (ResNet for orthophotos and PointVector for ALS point clouds), we investigate each data modality’s ability to assess forest biodiversity potential, achieving overall accuracies of 76.7% and 75.8%, respectively. We explore various 2D and 3D fusion approaches: confidence-based ensembling, feature-level concatenation, and end-to-end training, with the latter achieving an overall accuracies of 82.0% when separating low- and high potential forest areas. Our results demonstrate that spectral information from orthophotos and structural information from ALS point clouds effectively complement each other in the assessment of forest biodiversity potential.

💡 Research Summary

This paper addresses the challenge of scaling forest biodiversity assessments by leveraging close‑range remote sensing data—high‑resolution 2‑D orthophotos and 3‑D airborne laser scanning (ALS) point clouds—and deep learning. The authors introduce the BioVista dataset, comprising 44 378 paired samples collected from temperate forests across Denmark. Each sample covers a 30 m diameter (≈ 707 m²) area and includes a four‑channel orthophoto (RGB + NIR at 12.5 cm/pixel) and a dense ALS point cloud (≈ 8 points / m²). Ground‑truth labels are derived from the High Nature Value (HNV) forest map, which aggregates 11 proxy features (e.g., canopy height variation, large‑tree presence, coastal proximity). For the purpose of this study the HNV scores are collapsed into a binary classification problem: low biodiversity potential (HNV 1‑3) versus high biodiversity potential (HNV 7‑10).

The methodological pipeline consists of three stages: (1) single‑modality baselines, (2) multimodal fusion strategies, and (3) extensive evaluation. For the 2‑D baseline the authors fine‑tune a ResNet‑50 pretrained on ImageNet, feeding the four‑channel orthophoto patches and applying class‑balanced cross‑entropy loss together with standard augmentations (rotation, scaling, color jitter). For the 3‑D baseline they adopt PointVector, a recent transformer‑based architecture that directly consumes unordered point sets with associated intensity values. Both baselines achieve comparable performance (ResNet 76.7 % accuracy, 0.81 ROC‑AUC; PointVector 75.8 % accuracy, 0.79 ROC‑AUC), confirming that spectral information and 3‑D structure each contain useful signals for biodiversity potential.

Three fusion approaches are explored. (i) Confidence‑based ensembling simply averages the softmax probabilities of the two single‑modality models, yielding modest gains. (ii) Feature‑level concatenation extracts the penultimate ResNet feature vector (2048‑D) and the global PointVector embedding (1024‑D), concatenates them, and passes the result through a two‑layer fully‑connected classifier. (iii) End‑to‑end multimodal fusion integrates the two branches early in the network and introduces a cross‑attention module that lets the image stream attend to point‑cloud features and vice‑versa. This third strategy delivers the best results: 82.0 % overall accuracy and 0.87 ROC‑AUC, with a notable reduction in confusion between the low and high classes. Detailed per‑class metrics show a recall of 0.84 and precision of 0.80 for the high‑potential class, indicating that the model is reliable for identifying priority conservation areas.

The authors also provide a thorough error analysis. Misclassifications often occur in mixed‑structure zones where spectral signatures resemble high‑potential forests but structural metrics are ambiguous, highlighting the complementary nature of the modalities. The study acknowledges limitations: the dataset is geographically confined to Danish temperate forests, and the HNV proxy labels, while correlated with biodiversity, are not direct species counts. Consequently, external validation on other biomes and a comparison with field‑based species inventories are required before operational deployment.

Future work is outlined along four dimensions: (1) incorporation of multi‑temporal aerial imagery and hyperspectral data to capture phenological dynamics; (2) refinement of point‑cloud processing to include fine‑scale branch and understory geometry; (3) collection of species‑level ground truth to enable multi‑class (rather than binary) biodiversity modeling; and (4) application of domain adaptation techniques to transfer the learned models to forests in different climatic zones.

In summary, this research demonstrates that deep‑learning‑based multimodal fusion of orthophotos and ALS point clouds significantly outperforms single‑modality approaches for assessing forest biodiversity potential. The publicly released BioVista dataset, together with the demonstrated fusion architectures, provides a valuable foundation for scalable, data‑driven forest conservation planning.