Tree-Level Forest Monitoring with Artificial Intelligence Using Neural Networks and High-Resolution Imagery

Abstract

Mapping and monitoring tree crowns provides vital information on the earth's changing climate. We can use tree measurements to quantify the biomass and stored carbon in a forest, we can use those measurements to inform plantation management on harvest cycles, we can even use tree measurements to predict climate change. However, for these to occur there needs to be routinely measured and updated forest measurements which can be costly and take a long time to complete. There has been great success with the combination of machine or deep learning with high resolution imagery for individual tree crown mapping. However, many of these studies have had difficulty with dense deciduous canopies and most have focused on tree detection and not individual tree segmentation. In this collection of studies, I utilize sub-meter multispectral imagery with an instance segmentation convolutional neural network (CNN) to how variations of the training data and model architecture impact the model's ability to individually identify and segment trees. I chose southeastern Virginia, United States as the study area due to its unique combination of dense deciduous forests and homogeneous pine plantations, which have not been heavily studied with this model architecture. The model used was the U-Net CNN with an instance segmentation backend to separate trees from one another after they were identified by the model. I compared training data quality versus quantity, increased spatial resolution, and altered band combination across three studies, respectively. Chapter 1 assesses the impact of training data quality versus quantity for the model performance over two forest cover types and six forest age groups. Here, I used 60 cm spatial resolution imagery and hand-delineated tree crowns across mixed/deciduous forests, pine plantation forests, and a group that combined the two datasets. There was a total of 14,137 trees hand-delineated for the training data with 3,457 from the mixed/deciduous forests and 10,680 from the pine plantations. Three models were trained, one using all the data, and one for each set of forest cover specific data. Results of this study showed that the forest cover specific models performed better than the model that used all the forest cover data, indicating that training data quality is more important than quantity in this model architecture with these specific forest cover types. Chapter 2 assesses the impact of model application over the same pine plantation forests with a higher resolution image collected during the dormant season. I used 30 cm spatial resolution imagery and hand delineated tree crowns over only pine plantations to investigate the impact of the higher resolution in the model performance. The results of this study indicated better tree identification, but worse individual tree segmentation compared to the previous chapter. I found that the increased resolution performed better in the younger aged pines but worse in the older pines. Chapter 3 investigates the model's performance when using the near-infrared band and false color imagery for the training data and model architecture. I used the imagery from my previous studies but incorporated the near-infrared band into both the training and model architecture. I evaluated this adjustment by investigating the change in the training data as well as the change in model performance by forest cover type, pine plantation age, and the number of bands given to the model architecture. I found that the near-infrared band alters the training data and improves model performance by forest cover and pine plantation age. For the conclusion of this work, I compared all three studies and found that the best models for identifying and segmenting individual trees in southeastern Virginia, U.S. are with true color, 60 cm spatial resolution imagery for mixed/deciduous forests and false color, 60 cm spatial resolution imagery for the pine plantations.