Scaling Deep learning in forestry: Innovative training data strategies to enhance model performance and application

Deep learning has transformed industries like transportation and healthcare, yet its adoption in forestry – a field vital for ecological and economic stability – remains limited, partly due to challenges in generating scalable, high-quality training data. This dissertation overcomes these barriers by developing innovative deep learning models and frameworks integrated with remote sensing for forest management, enhancing model performance, generalizability, and practical utility across diverse forested landscapes. Guided by four objectives – establishing efficient training data generation, improving model adaptability, quantifying training data impacts, and demonstrating large-scale applications – this work advances deep learning applications in forestry using publicly available datasets such as the National Agriculture Imagery Program (NAIP), the 3D Elevation Program (3DEP), and the National Land Cover Database (NLCD). The dissertation comprises three studies addressing distinct training data challenges. The first study maps invasive Callery pear across New York City using high-resolution, single-source 4-band multispectral imagery. It successfully identifies and maps the species with an F1 score of 88.2%, while negative training data further boosts CNN performance, reducing false positives. This approach enhances invasive species management by leveraging phenological traits with accessible imagery. The second study compares the impact of high- and low-resolution training data on deep learning model performance. It reveals a minimal 2.7% F1 performance gap despite significant differences in training data generation costs between 1.5-meter 3DEP LiDAR and 30-meter NLCD data. This finding highlights cost-effective strategies for scalable training data generation in forestry. The third study employs a mixture of experts model with transfer learning for stem enumeration. Scaling medium-resolution data, it improves accuracy, reducing RRMSE from 34.6 to 14.9 in high-density forests. This provides a robust tool for large-scale stem inventory across diverse environments. Collectively, these studies advance scalable training data strategies, model adaptability, and practical forestry applications, offering tools for invasive species management, forest inventory, and carbon accounting. The findings underscore deep learning’s potential to transform forestry through research advancements in model development and practical management solutions, providing a foundation for future research into broader applications and refined methodologies to improve global forest resource management.