<b>AI-Driven Tree Species Identification: Building the CentralBark Dataset and Decoding Environmental Effects Leading to Classification Accuracy</b>

<p dir="ltr">Tree species classification is a foundational task in forestry, underpinning efforts in sustainable forest management, biodiversity monitoring, and climate resilience strategies. Historically, this process has relied on manual identification by trained experts, a method that, while effective, is labor intensive, and limited by the scalability of expert availability. There is a need for faster, more reliable tools to catalog and manage tree populations. Artificial intelligence (AI), particularly deep learning, has emerged as a transformative approach to automate this process, offering the potential to achieve high accuracy and efficiency. However, the success of such AI systems depends on two critical factors: access to high quality, diverse training data and the ability to perform reliably under the variable conditions encountered in natural environments. This thesis tackles these challenges through two interconnected contributions, advancing the field of digital forestry.</p><p dir="ltr">The first chapter is the development of the CentralBark dataset, a robust collection of over 19,000 bark images representing 25 hardwood species native to the Central Hardwood and Appalachian regions of the United States. These species, including economically and ecologically significant ones like walnut, white oak, cherry, hickory, maple, etc. were selected to reflect the diversity of the central hardwood region. The dataset was carefully designed to capture real world variability, incorporating images of trees with different diameters, taken under a range of lighting conditions (morning, afternoon, and evening), and in varying states of bark moisture (wet and dry). Each image is accompanied by detailed metadata, such as GPS coordinates, timestamp, and environmental notes, making CentralBark a uniquely comprehensive resource. To assess its utility, we benchmarked three cutting edge deep learning models: EfficientNet-b3, ResNet-50, and MobileNet-V3-small. The EfficientNet-b3 model outperformed the others, achieving a classification accuracy of 83.21%, a result that highlights the dataset’s potential to support the creation of accurate, AI-driven tree identification tools. This contribution not only fills a gap in the availability of bark specific datasets but also establishes a foundation for future research and practical applications in automated forestry.</p><p dir="ltr">The second chapter is a detailed environmental sensitivity analysis, which explores how external factors influence the performance of AI-based tree species classification. Using a curated subset of the CentralBark dataset (referred to as CBDS_Small). CBDS_Small is comprised of the top, middle, and bottom performers (northern red oak, hackberry, and bitternut hickory respectively). We examined the effects of three variables: time of day, bark moisture, and cardinal direction (the side of the tree photographed). Our analysis revealed that bark moisture is a particularly significant factor: dry bark conditions improved classification accuracy by 8.19% compared to wet conditions, likely due to the enhanced visibility of texture and color features when water is absent. We also identified species specific responses to lighting conditions tied to the time of day. Northern red oak performed best in afternoon light, benefiting from even illumination, whereas bitternut hickory consistently struggled across all times, suggesting inherent challenges in its bark features. Cardinal direction, however, showed a negligible impact, with accuracy varying by only 2.4% across north, south, east, and west facing sides. These findings provide actionable insights into the conditions under which AI models perform optimally, as well as the limitations that must be addressed for field deployment.<br></p>