A Rapid Simulation System for Decision Making in Intelligent Forest Management

INTELLIGENT DECISION MAKING

A Rapid Simulation
System for Decision
Making in Intelligent
Forest Management
Jing Fan, Tianyang Dong, Xinxin Guan, and Ying Tang, Zhejiang University of
Technology, China

T

sive evaluation of forest ecosystems—their function, structure, biological

Balancing different

diversity, adaptability, succession laws, and management efficiency. Therefore,

forest management

deciding on which evaluation criteria influence the forest economic, ecological,


goals related
to economic,
ecological, and
social sustainability,
a rapid simulation
system uses
CPU+GPU
heterogeneous
patterns to accelerate
the computation of
dynamic processes,
supporting quick and
intelligent decision
making.
2

he forest management decision-making process is built on the comprehen-

and social sustainability can have a tremendous impact on the final selection of forest
management plans.1,2 The cycle of forest

succession is very long—which includes germination, seed dispersion, juvenile growth,
juvenile mortality, adult growth, and adult
mortality—so forest managers don’t have
the required time to complete actual experiments and verify the effect of each forest
management plan. Instead, they use succession models to quickly predict the long-term
effects of forest management practices.
Spatially explicit forest succession models can describe the detailed interactions
among trees both vertically and horizontally, and they provide a realistic and accurate representation of a forest ecosystem.3,4
However, the incorporation of this explicitness into forest succession models leads to
great costs in terms of computation time.
Because hundreds of plans are generated to
reach all kinds of forest management goals,
it can  take several months to complete the

simulation and evaluation of each alternative to make a scientific and reasonable
selection. The existing simulation method
simply can’t satisfy the simulation requirements of large-scale plots in forest management decision making. This problem
has also seriously restricted the ability of
forest managers to quickly make relevant
decisions.

To speed up the simulation and evaluation of forest management plans, we present
a rapid system that uses GPUs to accelerate
the computation of dynamic forest succession
processes, simulate alternative forest management plans, and support intelligent selection
of forest management plans.

1541-1672/13/$31.00 © 2013 IEEE

IEEE INTELLIGENT SYSTEMS

Simulation Process
and System Architecture
Forest management decision making needs
to find more intelligent ways to strike a
balance between different forest management goals, such as the maximum amount

Published by the IEEE Computer Society

Initial scene


Initial distribution

IEEE Computer Society
Publications Office

Choices of key decision parameters
No thinning

Moderate
thinning

Threshold
diameter thinning

Plan 2

Plan 1

Classical clearcut system


10 years

Determination
of candidate
plans

...

Plan 3

10 years

...

...

Los Alamitos, CA 90720-1314

Lead Editor
Brian Kirk

[email protected]

10 years

10 years

20 years

10662 Los Vaqueros Circle, PO Box 3014

Plan n

Simulation of
management
plans

...
20 years

IEEE


Initial state

Management
knowledge

20 years

20 years

...

...

Editorial Management
Tammi Titsworth
Manager, Editorial Services
Jenny Stout
Publications Coordinator
[email protected]

Director, Products & Services
Evan Butterfield

150 years

Largest
harvest

150 years

Biological
diversity

Forest
distribution

Senior Manager, Editorial Services
Robin Baldwin

150 years


150 years

Ecological
environment

...

Intelligent ranking method of forest management plans

Management
goal

Digital Library Marketing Manager
Georgann Carter
Senior Business Development Manager
Sandra Brown
Senior Advertising Coordinator
Marian Anderson
[email protected]


Management
knowledge

Recommended
plan
Target scene

Target distribution

Figure 1. Simulation to support forest management decision making. It’s an
intelligent way to strike a balance between different forest management goals,
such as the maximum amount of acceptable deforestation, biological diversity, and
spatial distribution.

of acceptable deforestation, biological diversity, and spatial distribution.
Forest management plans also need
to factor in the most common selective cutting methods, such as no thinning, moderate thinning, threshold
diameter thinning, and classic clearcut. Figure 1 shows the simulation
process.

To rapidly simulate alternative forest management plans and quickly determine the optimal solution, the rapid
simulation system is divided into four
SEPTEMBER/OCTOBER 2013

layers (see Figure 2): data, strategy,
computing, and application.
The data layer is the system’s foundation. It stores and manages the
environment model, the forest succession model, any knowledge related to
forest management, and the data in the
simulation scene that are necessary for
simulating dynamic succession. Besides
helping with selection, the strategy layer
automatically identifies key decision
parameters and comprehensively considers both forest management goals
www.computer.org/intelligent

Submissions: For detailed instructions
and formatting, see the author guidelines
at www.computer.org/intelligent/author.
htm or log onto IEEE Intelligent Systems’
author center at Manuscript Central
(www.computer.org/mc/intelligent/
author.htm). Visit www.computer.org/
intelligent for editorial guidelines.

Editorial: Unless otherwise stated,
bylined articles, as well as product and
service descriptions, reflect the author’s
or firm’s opinion. Inclusion in IEEE
Intelligent Systems does not necessarily
constitute endorsement by the IEEE or the
IEEE Computer Society. All submissions
are subject to editing for style, clarity, and
length.

3

INTELLIGENT DECISION MAKING

Simulation of forest
dynamic
succession

Thread meshing Calculation of the
of seed
shading rate of
dispersion
adult tree

Selection of forest
succession
model

Calculation of
neighborhood
competition

Identification of
key decision
parameters

Forest succession
model

Calculation of
tree productivity

Forest
management
methods

Management
knowledge

Management of
knowledge and
models

Visualization and
statistical analysis of
simulation data

Forest
management
goals

Environment
model

Data layer

Forest scene
data

Intelligent ranking
for management
plans

Application layer Computing layer Strategy layer

Management of
candidate
solutions

Figure 2. Structure of the rapid forest simulation system, which is split into four
layers: data, strategy, computing, and application. The data layer forms the
foundation of the system where the models and related simulation data are stored;
the strategy layer helps to select the proper models and identify the key decision
parameters; the computing layer is responsible for the CUDA-accelerated calculation
of the forest succession model; and the application layer provides the interface to
support efficient management.

and methods. The computing layer is a
bridge between the strategy and application layers. It uses CUDA-based accelerated calculation to determine seed
dispersion, the neighborhood competition index, adult trees’ shading rates, and
tree productivity. Finally, the application layer provides a user interface for
managing plans, knowledge, succession
models, and so on. It can also calculate
a dynamic succession model based on
GPU acceleration, rapidly realize the
visualization and statistical analysis of
simulation data from each solution,
and intelligently rank management
plans to get to the optimal solution.

GPU-Based Accelerated
Calculation
As a spatially explicit dynamic model,
the SORTIE model takes into consideration the interactions among
tress both vertically and horizontally,
making it more realistic and accurate
than other forest models in simulating the evolution of large-scale forests
4

(www.bvcentre.ca/sortie-nd). We adopt
the SORTIE model in this work to simulate forest succession; the process is cyclic, and the simulation flow of one time
step is divided into six submodels: germination, juvenile growth, juvenile
mortality, adult growth, adult mortality, and seed dispersion. We can compute each submodel independently but
sequentially, with the computation depending on the results of the previous
submodel.
The computation process of dynamic forest succession is highly
complex: in a 100 × 100 plot area
with 2,500 trees, the simulation of a
year can take at least five hours on
a computer with an Intel Xeon CPU
E5506 2.13-GHz processor. Because
there are all kinds of forest management goals, decision making can produce hundreds of plans, thus when the
forest plot is much larger (at least 1
km2) and the simulation is run for a
longer duration (at least 100  years),
it can take several months to simulate
www.computer.org/intelligent

all the options. The main objectives of
our research are to determine how to
quickly simulate each forest management plan based on a spatially
explicit forest succession model, and
how to get more accurate simulation
data that are visual and quantitative
to support decision making.
Few studies improve the computation time of forest succession. Sathish Govindarajan and his colleagues
presented an efficient simulator that
adopted an optimized calculation
method to show how the seeds are
dispersed.5 However, the preprocess
is time-consuming for larger-sized
plots, and it’s no longer applicable
when the seed dispersion doesn’t
have obvious spatial-temporal variation characteristics. Govindarajan
and his colleagues also exploited the
model to calculate the distribution of
light resources in forest and proposed
a novel method to compute the distribution of understory light in the
forest based on the graphics card.6
To calculate the competition between
trees more efficiently, an implementation of the SORTIE model reduced
the calculation load by building the
list structure according to the circle
of competition influence.
Although our system has adopted
an algorithm that appears elsewhere,6
to speed up computing the distribution
of understory light, the large amount
of forest data involved in the simulation still leads to a large cost in computational time. From our experimental results’ statistics and analysis,
based on the accelerated computation for the distribution of understory
light, the calculation of adult growth
and seed dispersion submodels account
for 90 percent of the entire calculation,
making them the most important optimization objects. The submodels
of adult growth and seed dispersion
are computationally intensive, and
the data have highly parallel features.
IEEE INTELLIGENT SYSTEMS

Thread
Dispersion

GPU

Register
Thread block

Video memory
Dispersion

Shared
memory

seeds_block
array
System memory

Global
memory

Dispersion
0 1 2

Plot coordinates
adult attributes

CPU host

Germination

Juvenile
growth

GPU

Juvenile
mortality

Adult
mortality

Adult growth

Thread

System memory

Competition
Register

Plot coordinates
adult attributes

Target tree Neighbor tree
Thread block
Competition

Video memory

Shared
memory

Target tree Neighbor trees

Seed dispersion

NCI_block array,
Shading array

Grid

Global
memory

Competition

Figure 3. Heterogeneous operation patterns of the CPU+GPU pattern. In this pattern, the CPU is the main processor; the GPU
serves as the CPU-scheduled coprocessor, which is primarily for the CUDA-based parallel computing optimization of seed
dispersion and adult growth.

Accordingly, we can split these submodels into multiple subtasks by using
a parallel method for optimization.
In the succession process, germination, juvenile mortality, and adult
mortality are primarily related to the
logically complex operations of addition, deletion, and judgment; the
SEPTEMBER/OCTOBER 2013

computation consumption of juvenile
growth is relatively small. Therefore,
a CPU can execute these four subprocesses, so we use a CPU+GPU heterogeneous pattern to optimize overall
system performance, as Figure 3
shows.7,8 In this pattern, the CPU is
the main processor responsible for
www.computer.org/intelligent

organizing and storing forest data,
controlling the succession process,
and handling serial computation for
the four subprocesses. The GPU serves
as the CPU-scheduled coprocessor,
primarily for the CUDA-based parallel computing optimization of seed
dispersion and adult growth. In each
5

INTELLIGENT DECISION MAKING

growth cycle, the CPU sequentially
schedules each submodel according
to the succession process. When the
GPU-based calculation is required, the
CPU provides data for the GPU. Once
the GPU-based calculation is completed, the CPU regains control and
accepts the results returned from the
GPU, and then starts again with the
next submodel.
The GPU-based accelerated computation of seed dispersion divides
the plots into uniform grid cells. We
can calculate the number of seeds
dispersed in each plot to get the total seed dispersion results. This calculation process has two levels of
computation parallelism: the first is
to compute the number of seeds dispersed in the plot’s grid cells, and the
other is the process by which each
adult tree produces seeds per plot
cell. Figure  3 shows the GPU optimization algorithms for seed dispersion. The two parallelisms are
mapped to the threads of a 2D grid,
where the Y dimension (defined as
grid Y) represents the plot cells, and
the X dimension (defined as grid X)
represents the array of adult trees that
generate seeds. Each thread block is
responsible for calculating the seed
contribution of a certain number of
adult trees on a plot cell—each thread
deals with a single adult tree for the
seeds dispersed to that plot cell. The
total number of seeds contributed by
all the blocks in the same Y dimensions corresponds to that plot cell’s
seed dispersal results.
In the GPU-based accelerated computation of adult growth, the shading value blocked by neighboring
trees, the neighborhood competition
index, and final growth are divided
into three separate computational
kernels according to their computational complexity. The thread meshing
of the first two factors complies with
the method of 2D thread meshing for
6

the seed dispersion. The X dimension
represents the competition impact of
neighborhood adult trees, and the Y
dimension represents the computation
of target adult trees. Therefore, the
thread blocks on the same Y dimension respond to the competitive effects
computation of the same target adult
trees. Based on the shading rate and
neighborhood competition index, we
can design a 1D thread grid to compute the growth of adult trees in parallel, where each thread corresponds
to an adult tree.

Intelligently Ranking
Forest Management Plans
It’s difficult for decision makers to select the optimal forest management
plan—something that can meet multiple objectives from a large number
of options. To address this conundrum, we use an intelligent ranking
method of forest management plans
based on multiobjective optimization.
If n is the number of forest management goals, and m is the number of
candidate forest management plans
in decision making, we have
Xi = (x1i, x2i, ..., xmi),

i = 1, 2, …, n,

where Xi is the ranking vector of
plans for the ith forest management
goal, and xji is the ranking of the jth
plan in the ith forest management
goal (j = 1, 2, …, m). For the ranking
vector of forest management plans
after composite analysis, we have
X0 = (x10, x20, ..., xm0),
where X 0 is the ranking vector
of forest management plans after
composite analysis, and xj0 is the
ranking of the jth forest management
plan in X0 (j = 1, 2, …, m).
In the comprehensive ranking of
forest management plans, we use the
following multiobjective optimization:
www.computer.org/intelligent

n

min f = min

∑ Wi (Xi X0 )2
i =1
n

= min



m



∑ Wi  ∑ (x ji − x j0 )2  ,
i =1

 j =1



where 1