BiodiverseNet: Multitask Learning on Fused Multispectral and Radar Data for Scalable Ecosystem Monitoring

Monitoring global biodiversity is a pressing challenge given the accelerating rates of ecosystem degradation and loss. Earth observation (EO) offers a scalable solution, but existing methods struggle with heterogeneity in ecosystems, label sparsity, and modality limitations. We propose Bio-diverselvet11https://github.com/TransformerTitan/BiodiverseNet, a multitask learning framework that leverages fused multispectral (Sentinel-2) and radar (Sentinel-1) imagery to jointly predict key biodiversity metrics: canopy cover, habitat fragmentation, and landscape connectivity. Our model employs a Vision Transformer (ViT) backbone with DINOv2 pre-training, combined with task-specific heads and auxiliary objectives including land cover classification and NDVI prediction. Evaluated on a global benchmark spanning 15 ecoregions across five continents, BiodiverseNet achieves competitive performance with R2 scores of 0.76 for canopy cover, 0.62 for fragmentation, and 0.67 for connectivity, showing modest but consistent improvements of 2-2.5% over transformer baselines. The model demonstrates reasonable robustness across diverse biomes, though performance varies with ecosystem complexity and data availability.

• Publication year

  2025

• Venue

  2025 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW)

• Publication date

  2025-10-19

• Fields of study

  Not labeled

• Identifiers
  DOI 10.1109/ICCVW69036.2025.00309
• External record

  Open on Semantic Scholar

• Source metadata

  Semantic Scholar

• No claims are published for this paper.

• No concepts are published for this paper.

• DINOv2: Learning Robust Visual Features without Supervision

  2023cited by this paper
• Vision Transformers for Remote Sensing Image Classification

  2021cited by this paper
• TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation

  2021cited by this paper
• Continuous Damping Control for Rollover Prevention with Optimal Distribution Strategy of Damping Force

  2020cited by this paper
• An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

  2020cited by this paper
• Land Cover Maps Production with High Resolution Satellite Image Time Series and Convolutional Neural Networks: Adaptations and Limits for Operational Systems

  2019cited by this paper
• Priorities to Advance Monitoring of Ecosystem Services Using Earth Observation.

  2017cited by this paper
• Multi-task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics

  2017cited by this paper
• Deep Learning Classification of Land Cover and Crop Types Using Remote Sensing Data

  2017cited by this paper
• Beyond RGB: Very High Resolution Urban Remote Sensing With Multimodal Deep Networks

  2017cited by this paper
• An Overview of Multi-Task Learning in Deep Neural Networks

  2017cited by this paper
• Combining satellite imagery and machine learning to predict poverty

  2016cited by this paper
• Deep Learning for Remote Sensing Data: A Technical Tutorial on the State of the Art

  2016cited by this paper
• Satellite remote sensing for applied ecologists: opportunities and challenges

  2014cited by this paper
• The biodiversity of species and their rates of extinction, distribution, and protection

  2014cited by this paper
• High-Resolution Global Maps of 21st-Century Forest Cover Change

  2013cited by this paper
• Morphological segmentation of binary patterns

  2009cited by this paper
• Using circuit theory to model connectivity in ecology, evolution, and conservation.

  2008cited by this paper
• Overview of the radiometric and biophysical performance of the MODIS vegetation indices

  2002cited by this paper
• Tropical Deforestation and Habitat Fragmentation in the Amazon: Satellite Data from 1978 to 1988

  1993cited by this paper
• Red and photographic infrared linear combinations for monitoring vegetation

  1979cited by this paper

• No citing papers are available for this paper.