Large-scale detection, mapping, and initial health assessment of date palm trees using multiplatform remotely-sensed data and deep learning techniques

Reliable and efficient large-scale detection and mapping of date palm plantations through multiplatform remote sensing are crucial to developing palm tree inventories and sustainable management of the date palm industry. Given the increasing availability of unmanned aerial vehicle (UAV) images with limited spectral information, the high intraclass variance of palm trees, the variations in spatial resolutions of data, and the complexity of image contexts, accurate and automatic large-scale mapping of date palm trees from multisource and multidate data remains a challenge. Recent studies on date palms using convolutional neural network (CNN)-based object detection methods have primarily focused on limited study sites and relied on a single data source. In-depth investigations on the generalizability and transferability of semantic and instance segmentation models for mapping date palm trees from multiplatform remotely sensed data are lacking. Moreover, no effort has been exerted to assess the feasibility of evaluating the health of date palm trees from large-scale UAV-based images. This study aims to provide an end-to-end, efficient, and transferable deep learning architecture for the large-scale mapping of date palm trees and initial health assessment from UAV-based images. Considering the ability of deep vision transformers to capture robust local–global feature representations, this research hypothesizes that transformer-based models can deliver effective outcomes for large-scale mapping and assessment of date palm trees. This thesis evaluates various deep vision transformers and presents an efficient and cost-effective transformer-based framework to identify, quantify, monitor, and evaluate the overall well-being of palm trees using large-scale multiplatform images. This framework integrates a mask region CNN, a hierarchical Swin transformer, a feature pyramid network, and slicing-aided hyperinference to efficiently undertake large-scale instance segmentation of individual date palm trees, subsequently converting the results into a vector representation. Experimental results show that the examined deep vision transformers for the semantic segmentation of date palm trees are comparable to several CNNbased models and achieve satisfactory results in mapping date palm trees from UAV images. The SegFormer model, followed by the UperNet-Swin transformer, outperforms all the evaluated CNN-based models in the multiscale testing dataset and the additional unseen UAV test dataset. Moreover, the SegFormer model can be fine-tuned to delineate date palm trees using VHR WorldView-3 satellite imagery. The performance of the proposed instance segmentation framework surpasses that of several CNN-based models, demonstrating effective detection and delineation of individual date palm trees with F-scores of 94% and 93%, respectively. The proposed framework also exhibits great generalizability in detecting and mapping individual date palm trees from different UAV images with diverse spatial resolutions. The transformer-based architecture is fine-tuned through transfer learning to differentiate between healthy and unhealthy date palm trees. The potential generic condition of date palm trees is predicted with mAP50 of 80.2%. In sum, the proposed framework provides an efficient tool for accurately detecting and mapping individual date palm trees from multiscale and multidate UAV images, thereby building and updating geospatial databases and enabling consistent monitoring of date palm trees. It is suitable for individual tree crown delineations and other Earth-related applications.

Text 115761.pdf Download (1MB) Official URL or Download Paper: http://ethesis.upm.edu.my/id/eprint/18249

Actions (login required)

View Item