Date of Graduation
12-2024
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Geosciences (PhD)
Degree Level
Graduate
Department
Geosciences
Advisor/Mentor
Aly, Mohamed H.
Committee Member
Cothren, Jackson D.
Second Committee Member
Peter, Brad G.
Keywords
Biomass; Explainable AI; Forest; Fusion; Machine Learning; Remote Sensing
Abstract
Arkansas has a diverse landscape and rich biodiversity and is home to extensive forested areas, particularly in the Ozark and Ouachita regions. These forests play a vital role in carbon sequestration, biodiversity conservation, and ecosystem services. However accurate, high-resolution forest maps and current estimates of Aboveground Biomass (AGB) are lacking. This study addresses these needs by integrating multisource remote sensing data and advanced machine learning techniques to enhance forest monitoring and AGB estimation.
The research first evaluates the utility of multisource satellite imageries data fusion for forest monitoring by incorporating a systematic literature review of 72 articles. The study highlights the growing application of fusion techniques that combine optical, radar, and LiDAR data to improve classification accuracy in forest monitoring. The study also emphasizes the application of different machine learning models and key evaluation metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Overall Accuracy (OA). The study underscores the effectiveness of these methods across varied forested regions. This part of the study helped outline a possible framework for the second of the study.
The second study addresses Arkansas’s need for a high-resolution tree species distribution map by applying machine learning classifiers within the Google Earth Engine (GEE) platform. Here, different combinations of fused datasets are created from Sentinel-1, Sentinel-2, Landsat-8, and NAIP (National Agriculture Imagery Program) data to determine which combination yields the best outcome. The study includes commonly used algorithms such as Random Forest (RF), Gradient Tree Boosting (GTB), Support Vector Machine (SVM), and K-Nearest Neighbors (K-NN) to assess which model is most suitable for forest classification in Arkansas. The analysis finds that fused data from Landsat-8 and Sentinel-1 achieve the highest classification accuracy (0.8875), with RF outperforming the other algorithms. By incorporating Shapley Additive Explanations (SHAP), this study enhances model transparency and provides insights into the significance of key features, such as elevation and reflectivity characteristics of vegetation. Furthermore, the study produces a 10-meter resolution map by applying an ensemble-based model, which demonstrates greater accuracy than the individual models.
The third part of the study fused Sentinel-1, Sentinel-2, and GEDI (Global Ecosystem Dynamics Investigation) data on the GEE platform to develop a 10-meter resolution AGB model using RF regression. Here, a total of 34 variables (out of 154 variables available) representing topographical, spectral, and textural variables were selected for the final model. The model achieved an R-squared of 0.95 for training and 0.75 for validation, showing robust predictive performance. Furthermore, historical biomass data from 2015 to 2023 were also extrapolated using Landsat-8, revealing biomass dynamics with values ranging between 100 and 200 Mg/ha.
Overall, this research demonstrates the effectiveness of data fusion and machine learning in generating accurate, region-specific forest monitoring and AGB estimation tools. The results provide valuable insights for biodiversity conservation, fire risk and related emissions estimation, and sustainable forest management, with scalable applications for broader geographic areas.
Citation
Saim, A. (2024). Machine Learning and Satellite Data Fusion for Monitoring Forest Changes and Modeling the Above Ground Biomass in Arkansas. Graduate Theses and Dissertations Retrieved from https://scholarworks.uark.edu/etd/5586
Included in
Geographic Information Sciences Commons, Natural Resources Management and Policy Commons
