Aligning Tree Locations from VHR Images with LiDAR Data Using Deep Learning Models
Makian, Hamed (2025)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-202503114067
https://urn.fi/URN:NBN:fi:amk-202503114067
Tiivistelmä
Deep learning, particularly computer vision algorithms, has gained significant attention in remote sensing for solving various object detection problems. The development of efficient object recognition techniques has been a longstanding challenge, leading to the proposal of numerous unique methods over time to address this issue.
In this master thesis Computer vision algorithms were used to align VHR images with their corresponding Lidar canopy height images, in the area near Lappeenranta in Finland. The models had been designed to predict the positions of trees, tree boles (top of trees) and canopy height, with a primary focus on accurately identifying real tree locations, accounting for tilt effects in satellite imagery caused by the satellite’s viewing angle during image capture.
To achieve this, convolutional neural networks (CNNs) were combined with transformer architectures, leveraging pre-trained models to extract weights from large-scale datasets, thereby enhancing performance and reducing training time. Through rigorous experimentation, the best-performing model — the DeepLabV3 architecture with the ADOPT optimizer, attention mechanisms — achieved:
• Canopy height estimation: Lowest Validation Regression Loss (RMSE) of 3.065 meters, indicating accurate continuous canopy height predictions.
• Tree cover classification: Highest validation accuracy of 89.735%, with Class 0 Accuracy of 85.344% and Class 1 Accuracy of 91.353%, demonstrating balanced and reliable tree cover mapping.
• Tree centroid classification: Centroid classification accuracy of 77.285%, confirming the model’s ability to identify tree canopy centers with reasonable precision.
These results highlight the feasibility of using deep learning approaches for forest structure analysis, with CNN-based architectures outperforming transformer-based models for fine-grained spatial prediction tasks.
In this master thesis Computer vision algorithms were used to align VHR images with their corresponding Lidar canopy height images, in the area near Lappeenranta in Finland. The models had been designed to predict the positions of trees, tree boles (top of trees) and canopy height, with a primary focus on accurately identifying real tree locations, accounting for tilt effects in satellite imagery caused by the satellite’s viewing angle during image capture.
To achieve this, convolutional neural networks (CNNs) were combined with transformer architectures, leveraging pre-trained models to extract weights from large-scale datasets, thereby enhancing performance and reducing training time. Through rigorous experimentation, the best-performing model — the DeepLabV3 architecture with the ADOPT optimizer, attention mechanisms — achieved:
• Canopy height estimation: Lowest Validation Regression Loss (RMSE) of 3.065 meters, indicating accurate continuous canopy height predictions.
• Tree cover classification: Highest validation accuracy of 89.735%, with Class 0 Accuracy of 85.344% and Class 1 Accuracy of 91.353%, demonstrating balanced and reliable tree cover mapping.
• Tree centroid classification: Centroid classification accuracy of 77.285%, confirming the model’s ability to identify tree canopy centers with reasonable precision.
These results highlight the feasibility of using deep learning approaches for forest structure analysis, with CNN-based architectures outperforming transformer-based models for fine-grained spatial prediction tasks.