AI Algorithms for Identifying Patterns of Tree Failures and Abnormal Tree Stability Conditions
Supporting the data analytics work for the PSTS Project with the title of “Outdoor IoT Sensing Network
and Data Management Platform for Site-specific Tree Management”
1 Introduction
The soon-to-be available large-scale tree-specific and site-specific dataset collected in continuous
manner is a sole enabling factor to make use of dynamic tree growth data together with varying close
surrounding environmental conditions for identifying patterns of tree failures and abnormal tree
stability conditions. Such patterns may facilitate GLTMS to gain insight of how those dynamic time
series tree-related datasets lead to possibility of failures.
2 Initial Datasets Planned for Proposed project
This project initiates the data collection in three different levels – tree-specific, site-specific, and city-
wide regional levels. This project provides devices and collection methods for tree-specific and site-
specific datasets. Details of individual types of sensor data at each level are described below.
1. Tree-specific dataset:
a. Time-series dataset of tree truck tilt angle and tilt direction
b. Tree-specific tilt angle and direction of growth and fall of individual trees
i. This dataset provides a longitudinal tree growth and derived tree health
statuses for each of the trees tagged with sensor devices.
c. Data are collected at least once per day.
2. Site-specific datasets:
a. Time-series datasets from following sensors:
i. Wind direction
ii. Wind speed
iii. Air pressure
iv. Relative humidity
b. Site-specific data may be different from infrastructure-based (or regional) data collected
by Hong Kong Observatory (HKO).
c. Site-specific data may capture sensor data directly reflecting characteristics of
surrounding landscape or cityscape. For example, wind speed is affected by whether
wind tunnel effect exists. Another example is that amount of sunshine may affect
relative humidity of a specific site.
d. Site-specific data provides a set of sensor datasets having more direct causality to tree
growth and failure issues.
e. Like tree-specific dataset, data are collected at least once per day.
3. City-wide regional weather-related datasets
a. Time-series datasets of rain volume data from Hong Kong Observatory (HKO) and Civil
Engineering and Development Department (CEDD)
b. Rain volume data collection locations may be different from tree site locations.
c. The coverage area of each rain volume data collection location may cover at least one
tree site location.
d. The rain volume data collection frequency may vary between sites from different
departments.
e. The frequency of rain volume data may be different from that of tree-specific and site-
specific data.
f. Data are collected based on what individual providers are able to provide.
3 AI Algorithms for Analyzing Time-series and Sequence
Data
3.1 Time-series Datasets Related to Tree Growth
A coarse-grained view to a tree life (or sequential tree growth process over a time span) may slowly
reveal changes of tree main truck tilt angle and tilt direction. A fine-grained view of a short time span,
such as duration of gust, may reveal sudden vibrations or changes of tilt angle and direction. Under
normal circumstances, tilt angle and direction may mostly return to its usual states immediate after such
a sudden event. These coarse-grained and fine-grained views can be commonly observed from trees
around us.
In contrast, a fallen and broken tree may most likely demonstrate a permanent, non-reversible sudden
tilt angle and direction changes. Thus, the historical tilt angle and direction data along with
corresponding weather-related environmental data may encode the details of tree growth statuses.
3.2 Time-series Granularity and Asynchronous Sequences Across
Datasets
Because vibrations and changes happen normally to a tree throughout its life, the granularity, in terms
of time span and weather condition types, for time-series and sequence analysis may reveal different
information. The data collection frequencies vary across all 3 types of tree growth related datasets –
tree-specific, site-specific, and regional. Without any prior data-driven longitudinal study on changes of
tree statuses, sequence analysis will perform on different levels of abstract with the combined
asynchronous data from tree-specific, site-specific, and regional datasets.
3.3 Long Short-Term Memory (LSTM) Sequence Clustering Algorithms
Long short-term memory (LSTM) is an artificial recurrent neural network (RNN) architecture. LSTM
network architecture has feedback connections and memory handling capability. LSTM is able to handle
entire sequences of data, like documents (sequence of words or characters), speech (sequence of
acoustic signals), video (sequence of images or objects in images), weather changes (sequences of
different weather and environmental data), and historical data (sequences of growth data about life,
economy, etc.). Thus, the memory part of the LSCM networks are able to encode the dependencies
between elements in a data sequence.
3.4 LSTM for Clustering Multivariate Tree Growth Sequence Data
In order to cluster a set of multivariate tree growth time-series data, LSTM network is used to memorize
sequence data from different data streams in the form of sequence relationship pattern. A collection of
those sequence relationship patterns coming from trees at different sites in various regions. The
following figure illustrates how a set of multivariate data forming a data vector for a particular time
point is turned into an element of a sequence as input to the LSTM network. The LSCM units will then
learn sequence information and output computed results into corresponding memory vectors. All
memory vectors are flattened into a sequence relationship pattern.
The collection of sequence relationship patterns will be the input for a clustering process to group
similar sequences together. Larger group size represents commonly found sequences in the dataset.
Smaller group size represents relatively unique sequence characteristics. In the context of tree growth
statuses, trees having sequences in smaller groups may carry a status different from the same kind of
tree showing commonly statuses. Each sequence groups can be further analyzed for what tree types
and number of those trees carrying that similar sequence. This analysis can extend to different site-
specific and regional weather elements. A neural network-based self-organizing map (SOM) algorithm
can perform unsupervised learning to produce a low-dimensional (typically 2-dimensional), called a
map, with similar sequences being grouped together. The map also serves as a visualization tools to
view high-dimensional sequence data.
4 R&D Tasks
1. LSTM for Deriving Sequence Relationship Patterns
a. Experimenting the elements for composing a data vector
b. Experimenting the number of elements to form a sequence
c. Experimenting a configuration for LSTM network suitable for the tree growth sequence
data
2. Using a clustering algorithm such as SOM to groups sequence data
a. Experimenting a configuration for clustering algorithm
b. Analyze whether the similarity of sequences in each group having makes sense
intuitively