Detecting and counting trees from satellite images

Abstract

This thesis addresses the challenge of detecting and counting trees from satellite images using deep learning techniques. Accurate tree detection and counting are essential for environmental management, urban planning, and biodiversity conservation. Leveraging the Yosemite Tree Dataset, this study develops a U-Net-based model to estimate density maps for tree counting, refining the standard density estimation methodology through iterative design and testing.

The research explores three phases: establishing a baseline model with static standard deviation, revisiting the loss function to improve density map accuracy, and introducing a learnable standard deviation to enhance adaptability. Results demonstrate the effectiveness of the proposed model, achieving competitive performance against state-of-the-art benchmarks, particularly excelling in regions with moderate tree density.

Key contributions include innovative use of adaptive Gaussian kernel parameters and a multi-component loss function to balance count accuracy and spatial resolution. This work highlights the potential of density estimation techniques in remote sensing applications and identifies challenges, such as dataset variability and computational constraints, that warrant further exploration. The developed model provides a foundation for future research in tree detection and other object counting tasks in complex environments.

The full implementation developed in Pytorch and the model weights and code are available at: https://doi.org/10.5281/zenodo.14274482