News
Events
Doctoral Meeting: 'Deep learning techniques for accurate and efficient biodiversity monitoring wi...

Doctoral Meeting: 'Deep learning techniques for accurate and efficient biodiversity monitoring with limited data'

Vegetation biodiversity is essential for maintaining healthy ecosystems, yet it faces growing threats from human activities such as land transformation, pollution, and climate change. Automated large-scale biodiversity monitoring has become essential for informed decision-making, offering insights such as tracking changes in species diversity and detecting the spread of invasive species.

Remote sensing imagery, including UAVs and satellites, enables efficient monitoring of large areas. Equipped with modern multispectral sensors, these platforms capture detailed spectral data that goes beyond traditional RGB imagery. While deep learning has shown promise in processing remote sensing data across various fields, its application in biodiversity monitoring faces key challenges:

  • Limited labeled data due to labor-intensive field collection.
  • Imbalanced datasets caused by the natural dominance of certain species.
  • Difficulty distinguishing species with similar spectral patterns using only spectral data.

This research aims to advance deep learning methods to maximize the utility of limited biodiversity data for accurate and efficient monitoring. Two approaches will be presented, leveraging additional data dimensions beyond spectral information, and employing various techniques to enhance model performance: (1) ResBaGAN, which integrates spectral and spatial information by extracting vegetation textures from high-resolution UAV imagery, and (2) ResNeTS, which incorporates spectral and temporal data to identify seasonal growth patterns from Sentinel-2 time series. Future research aims to integrate spectral, spatial, and temporal dimensions, paving the way for more advanced biodiversity monitoring solutions.