Advances In Multi- and Hyper-Spectral Imagery Analysis for Improved Species Classification, Ecological Integrity Assessment, and Aboveground Biomass Mapping

ABSTRACTGlobal forests are declining at large geographical scales and at rates that outpace our ability to monitor them. The use of field methods alone is limited in scope and can be labour-intensive when utilized to capture multi-temporal changes in biodiversity at a landscape scale. However, it is possible to overcome these constraints using remote sensing, because large geographic areas can be assessed consistently and repeatedly. There is, therefore, a need for innovative, operational remote sensing-based approaches for accurate, reliable, and repeatable monitoring of changes in forest ecosystems at large geographic scales to support conservation initiatives. The aim of this dissertation is to develop novel remote sensing approaches to assess forest ecological integrity, species composition, and aboveground biomass at multiple spatial scales. The study employs multi- and hyperspectral (HSI) images collected from temperate and tropical forest ecosystems. The dissertation makes three key contributions, which are elaborated upon in three distinct chapters as follows.The first analytical chapter develops a new metric for ecological integrity from optical satellite imagery, referred to as the phenospectral similarity (PSpecM) metric and demonstrates its implementation in Google Earth Engine to help identify potential terrestrial stand-level forests of high ecological integrity (EI). The PSpecM approach was tested in a temperate deciduous forest in Quebec, Canada, and it was found that, 2,700 km2 (22%) of the study area were commonly delineated as potential areas of high EI. As Canada leads the implementation of the IUCN standard for Key Biodiversity Area (KBA) identification, this metric can be deployed for rapid assessments of forest stands to delineate potential areas of high EI.The second analytical chapter introduces a new approach to mangrove classification based on the concept of spectral complexity, which measures the degree of variation in spectral information within a set of hyperspectral images. The study evaluates two metrics of complexity, namely the mean information gain (MIG), a measure of spatial heterogeneity, and the marginal entropy (ME), a measure of aspatial heterogeneity together with spectral signatures for mangrove classification. It was found that while the highest accuracy separating mangroves from terrestrial forests in the Visible and Near-Infrared was for the reflectance data (Overall Accuracy = 98.8%), the classification accuracy of both MIG and ME outperformed the original spectral reflectance in the shortwave Infrared by 2.7%. The results from this chapter indicate that MIG and ME can extract features from HSI to improve mangrove classification accuracy.The final analytical chapter demonstrates the utility of machine learning and wavelet-based approaches for modelling aboveground biomass (AGB) using airborne HSI from tropical and temperate forest ecosystems. Spectral features were extracted from HSI using a wavelet transform attained an R2 of 0.72 and an RMSE of 45 Mg/ha. In contrast, the extraction of spectral and spatial features simultaneously from HSI using a Deep Convolutional Neural Network (3D CNN) produced the best modeling performance improving upon the wavelet-based AGB predictions by ~ 50%. This finding suggests that 3D-CNN can improve the reliability of forest AGB estimates in tropical and temperate forest ecosystems.Overall, the dissertation advances multi- and hyperspectral data analysis and their application in diverse types of forest ecosystems and presents novel methods to assist in forest ecosystem assessments and conservation efforts

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.