A Framework for Mining Inventories of Individual Trees

Categories: Trees

Essay, Pages 4 (813 words)

Views

Inventories of individual trees are important datasets for societal applications in urban planning (e.g., removal of Ash trees), natural resource management, vegetation disease control, etc. However, tree inventories still remain unavailable in most of urban areas due to the difficulty of manual data collection. With the increased availability of high-resolution remote sensing datasets, we aim to automate tree identification at individual levels in urban areas at a large scale. The problem is challenging due to the complexity of the landscape in urban scenarios (e.

Don't use plagiarized sources. Get your custom essay on

“ A Framework for Mining Inventories of Individual Trees ”

Get custom paper

NEW! smart matching with writer

g., a mixture of buildings, bridges, trees, etc.). In addition, the lack of ground truth data hinders the use of machine learning frameworks for object detection (e.g., convolutional neu- ral networks).

In related work, tree identification algorithms have mainly focused on controlled forest regions where the landscape is mostly homogeneous with trees alone, making the methods difficult to generalize to urban environments. We propose a two-phase framework, namely Tree Inference by Minimizing Bound-and-band ERrors (TIMBER), to find individual trees in complex urban environments.

In the first phase, TIMBER infers the locations and sizes of tree-like structures with local error minimization. Leveraging the first-phase outputs with additional city infrastructure data, TIMBER optimizes a deep learning filter in the second phase to eliminate non-tree structures. In addition, we propose a Core Object REduction (CORE) algorithm and an image index to improve the computational efficiency of TIMBER. Experiments show that the proposed framework can significantly improve the accuracy of individual tree detection in urban environments compared to existing approaches, and the acceleration techniques can speed up of tree identification without sacrificing result quality.

Inventories of individual trees provide meaningful infor- mation for urban planning, sustainable community planning, natural resource management, etc. In recent years, the invasive emerald ash borer has expanded to many countries (e.g., US, Canada, England) and caused tree deaths in millions. Since ash trees account for more than 20% of the urban forest in many US cities, the cost of this ash borer effect has been estimated to be over 10 billion US dollars. To respond to the threat, many state and city governments have begun to identify and treat (or remove) every individual ash tree in their management zones. In addition, increasingly many cities are eager to acquire and use fine-scale tree inventories (e.g., with locations, canopy sizes, heights) in their green infras- tructure management, which is a critical part for sustainable community planning [3]. However, unlike building footprints and other man-made architectures, inventories of individual trees are rarely collected or archived in most of the urban areas. It is also difficult and time-consuming to manually locate individual trees and measure their canopy sizes (e.g., limited GPS signals under canopies, the large population of trees).

We aim to automate the generation of individual tree inven- tories using high-resolution (e.g., one meter or lower) remote sensing datasets that are available at large-scales. There are two building blocks for this problem. The first is the localiza- tion and measuring of individual trees, which yields general spatial information of the trees (e.g., coordinates, canopy diameter, height, etc.). The second is to use the individual tree mapping to collect and aggregate other properties (e.g., visible and non-visible spectral properties, shape, temporal changes) at individual tree level to identify the signatures of genus, species (e.g., ash) and conditions (e.g., disease) of the trees. In this paper, we target the first building block, that is, identifying the locations and sizes of individual trees in an urban area. The type of remote sensing data that we use is the Normalized Height Model (NHM), which is a single band image whose pixel values represent surface heights. NHM is a LiDAR-derivative that has been collected and made publicly available at large scales (e.g., state-level or major urban areas across the US).

The tree identification problem has the following three challenges. First, unlike trees in rural areas and forests, trees in urban environments are often mixed with buildings, low- vegetation, lawns, towers, etc. These structures may share sim- ilar geometric or visual properties in remote sensing datasets. Second, it is common for trees to appear in groups, where their canopies heavily overlap with each other, making it difficult to distinguish individual trees. Third, individual tree inventories that contain information about locations and canopy sizes are rarely collected at large scales in urban areas or shared in public, making it difficult to gather ground truth data to train machine learning algorithms for object detection (e.g., convolutional deep neural networks).

In the literature of data mining, Gaussian mixture models (e.g., k-means, expectation-maximization) is related to the tree detection problem. In normalized height models, the canopies of trees are dome-shaped, which makes them similar to mixtures of Guassian distributions. However, k-means or its expectation-maximization generalization often rely on an input number of clusters, which is unknown in the tree detection problem.