MAPPING FOREST ABOVEGROUND BIOMASS IN THE BRAZILIAN AMAZON USING AIRBORNE LIDAR, LANDSAT IMAGERY, AND DEEP LEARNING

Simple item page
dc.contributor.advisor Chen, Qi
dc.contributor.author Zhang, Qian
dc.contributor.department Geography
dc.date.accessioned 2022-03-03T19:52:27Z
dc.date.issued 2021
dc.description.degree Ph.D.
dc.embargo.liftdate 2024-03-02
dc.identifier.uri http://hdl.handle.net/10125/81609
dc.subject Remote sensing
dc.subject Artificial intelligence
dc.subject Geographic information science and geodesy
dc.subject Brazilian Amazon
dc.subject Deep learning
dc.subject Forest aboveground biomass
dc.subject LiDAR
dc.subject Remote sensing
dc.subject Tropical forest
dc.title MAPPING FOREST ABOVEGROUND BIOMASS IN THE BRAZILIAN AMAZON USING AIRBORNE LIDAR, LANDSAT IMAGERY, AND DEEP LEARNING
dc.type Thesis
dcterms.abstract The Amazon forest is playing a critical role in the global carbon cycle and implementation of Reduce Deforestation and Forest Degradation (REDD+). However, the range of possible carbon emissions in this region is broad. Most carbon in the Amazon forest is stored in biomass and biomass can be the potential carbon emission when disturbances occur (e.g., deforestation, degradation, and fires). Therefore, an accurate estimation of biomass can help better predict carbon emissions in the Amazon forest. The biomass estimations of previous studies show little agreement on their values and spatial distributions in this region. In addition, deforestation and degradation in the Brazilian Amazon have changed significantly from large-scale patterns to fine-scale patterns since the early 2000s. However, existing biomass maps for the Brazilian Amazon forests are limited in capacity to capture fine-scale biomass variations due to their coarse spatial resolutions. Besides, due to the high level of biomass and heterogeneity of tropical forests, the commonly used regression models perform worse in tropical forests compared to boreal and temperate forests. Deep learning is a promising way to improve the accuracy of biomass estimations, which are increasing in success across a variety of remote sensing tasks. The application of deep learning models in estimating forest biomass is still in a nascent stage. Given the aforementioned research gaps, this research proposed a deep learning framework to estimate and map aboveground biomass on a fine-scale for the Brazilian Amazon with inventory data, airborne LiDAR data, and Landsat imagery. Three stages are involved in the framework development.In the first stage, a multiplicative power model was developed to link airborne LiDAR metrics with biomass inventory data. To determine the best fitting approach to estimate parameters for the multiplicative power model, three multiplicative power models fitted by nonlinear least-square (NLR), linear ordinary least-square (OLSR), and weighted linear least-square (WLSR) were compared by ANOVA and Tukey’s Test. The results show that significant performance differences existed among the three models at a 99% confidence level. More extreme predictions and lower accuracies were produced by NLR compared to OLSR or WLSR. OLSR had the most accurate prediction performance. Accordingly, OLSR was used to fit the LiDAR-based model that was used in the subsequent stages to calculate biomass for each LiDAR transect in the Brazilian Amazon forests. In the second stage, a deep feedforward fully connected neural network (DNN) model was developed to estimate and map aboveground biomass with airborne LiDAR data and Landsat 8 imagery. The effects of hyperparameter values on the DNN model performances were comprehensively investigated. The results show that the model with Scaled Exponential Linear Unit (SELU) had the best performance compared to other activation functions. Besides, both too large and too small learning rates could not achieve optimal results. The learning rate of 0.001 was chosen for the Adam optimizer. The DNN model with these optimal hyperparameters significantly outperformed the Random Forest model, Support Vector Regression model, and Linear regression model with the R2 of 0.64 and RMSE of 55.7 Mg/ha. This stage provides new insight into the application of deep learning in estimating forest biomass. In the last stage, Landsat time-series imagery was utilized to enhance the relationship between Landsat spectral reflectance and biomass. An RNN-FNN model integrating the long short-term memory network (LSTM) and the fully connected neuron network (FNN) was proposed to capture time dependencies in Landsat time-series data. The RNN-FNN model was compared to the Random Forest model and linear regression model implemented with single-date predictors. The results indicate that the RNN-FNN model significantly outperformed the Random Forest model and linear regression model. The RNN-FNN model yielded an R2 of 0.63 and RMSE of 25.5 Mg/ha with 10-year time-series data (2004-2013). At last, the RNN-FNN model was used to generate a map of biomass density for the study area, which demonstrated the practical value of the proposed model. The proposed framework that bridges inventory data, airborne LiDAR data, and Landsat imagery provides an effective way for forest managers to estimate and understand the spatial distribution of aboveground biomass in the Brazilian Amazon forests. In addition, this research illustrates the value of deep learning in estimating forest biomass and provides practical guidance for future studies on biomass estimations with deep learning.
dcterms.extent 93 pages
dcterms.language en
dcterms.publisher University of Hawai'i at Manoa
dcterms.rights All UHM dissertations and theses are protected by copyright. They may be viewed from this source for any purpose, but reproduction or distribution in any format is prohibited without written permission from the copyright owner.
dcterms.type Text
local.identifier.alturi http://dissertations.umi.com/hawii:11194
Files
Original bundle
Now showing 1 - 1 of 1
No Thumbnail Available
Name:
Zhang_hawii_0085A_11194.pdf
Size:
9.06 MB
Format:
Adobe Portable Document Format
Description:
Download
Collections
Ph.D. - Geography