A forestry investigation: Exploring factors behind improved tree species classification using bark images

Abstract

Novel ground-based remote sensing approaches have demonstrated high potential for accurate and detailed mapping and monitoring of forest ecosystems. These methods enable the measurement of various tree parameters important for forest inventory or ecological research, such as diameter at breast height, tree height and volume, and crown parameters. One crucial piece of information is tree species, which is essential for various reasons and challenging to implement within ground-based technology workflows. This study investigates why researchers often focus on segment-specific bark images for tree species classification via deep neural networks rather than large or entire tree images. Additionally, the aim is to determine the most effective algorithmic approaches for efficient tree species classification from bark images and to make these methods more accessible to interdisciplinary researchers. The findings reveal that segment-specific datasets with more overlaps provide better accuracy across various algorithms. Additionally, pre-processing techniques such as scaling can enhance accuracy to a certain extent. Convolutional Neural Networks (CNNs) consistently deliver the highest accuracy, even with diverse datasets, but fine-tuning these algorithms poses significant challenges for interdisciplinary researchers. To address this, we developed Windows-based research software, CNN Parameter Tuner 1.0, which allows the import of various data formats (jpg and png) and efficiently conducts parameter tuning by selecting parameters and values from the menu options.

Keywords

• Tree species classification
• Machine learning
• Close-range remote sensing
• Grid search
• CNN parameter tuning