Status : Verified
| Personal Name | Narciso, Gilson Andre M. |
|---|---|
| Resource Title | Comparison of Segment Anything Model and U-Net in developing an automatic coconut tree crown delineation model for optical unmanned aerial vehicle (UAV) remote sensing orthomosaics |
| Date Issued | 14 February 2025 |
| Abstract | The coconut tree, often called the "tree of life," is renowned for its diverse applications, serving as a vital source of food and as a renewable energy resource in the form of biodiesel. With the growing emphasis on sustainable and climate-sensitive development, the demand for coconuts is projected to rise, potentially leading to conflicts between its use for food and energy. This highlights the urgent need for effective resource management and systematic monitoring of national coconut resources. In the Philippines, however, the integration of modern technologies into natural resource management remains slow. To address this challenge, this study explores the enhancement of UAV remote sensing through the application of deep learning techniques, particularly for mapping coconut tree crowns. Two models, the Segment Anything Model (SAM)—a Vision Foundation Model trained on extensive web-scale data—and U-Net, a conventional convolutional neural network, were evaluated for their ability to perform automatic and unsupervised coconut tree crown delineation through a binary semantic segmentation process. Fine-tuning strategies for SAM, as well as multiscale and multimodal training methods were implemented for both models to assess their capabilities for this task. Results indicate that U-Net outperformed SAM, achieving higher segmentation accuracy with average F1-Scores and IoU values of 0.889 and 0.848, respectively, compared to SAM’s 0.772 and 0.73 using probability threshold value of 0.5. U-Net also demonstrated superior performance in multiscale training, with an average increase of 0.128 in F1-Score and 0.194 in IoU. However, multimodal training yielded suboptimal results for both models mainly due to the lack of model optimization for such data. For SAM, its image encoder’s bias towards RGB datasets can explain its low segmentation performance using the multimodal data. The lack of further preprocessing and normalization of the multimodal data could have also |
| Degree Course | Geomatics Engineering |
| Language | English |
| Keyword | UAV remote sensing; coconut trees; semantic segmentation; deep learning; Segment Anything Model (SAM); U-Net |
| Material Type | Thesis/Dissertation |
Preliminary Pages
991.98 Kb
Category : P - Author wishes to publish the work personally.
Access Permission : Limited Access