Forestry information extraction using high spatial resolution remote sensing imagery

Thumbnail Image



Journal Title

Journal ISSN

Volume Title


University of New Brunswick


This PhD research focuses on the development of reliable and efficient methods for individual tree crown delineation and tree species classification, which provide essential information for modern forestry management and climate change monitoring. The primary objective of this research is to develop robust methods that can accurately delineate individual tree crowns and classify tree species using high spatial resolution remote sensing imagery. By achieving this objective, the research aims to enhance the reliability and efficiency of forestry management practices and contribute to the field of remote sensing applications in forestry. For tree crown delineation task, existing tree crown delineation methods are not suitable for large areas applications, because they need highly experienced experts to manually assign suitable parameters to control the delineation results, which is time-consuming, inaccurate, and not suitable for normal users. To address this issue, this dissertation presents a tree crown delineation method utilizing marker-controlled watershed segmentation specifically designed for high spatial resolution multispectral WorldView-3 satellite imagery. To reduce the difficulty of assigning parameters, the proposed method incorporates an automated supervised search process to determine the threshold. Moreover, an enhanced definition of spatial local maximum is employed to mitigate false treetops, thereby enhancing the accuracy of treetop detection. For tree species classification task, although deep learning methods based on convolutional neural networks (CNN) have achieved promising results, challenges remain in hyperparameter tuning and the requirement for large number of labeled training samples, limiting their applicability in real-world applications. This dissertation addresses these challenges by proposing three models based on the concept of deep forest to enhance the tree species classification from high spatial resolution hyperspectral imagery. All the three proposed models only require two hyperparameters that are easy to be determined by users. To optimize the classification accuracy, two different ways to combine both fixed-size patches and shape-adaptive superpixels are proposed to fully exploit spectral-spatial information within the high spatial resolution hyperspectral imagery. To reduce the demand for labeled training samples, the active learning (AL) is perfectly integrated into the multilayer cascaded random forests classification model.