, Part 7B
In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
309
DETECTION OF FOREST MANAGEMENT OPERATIONS USING BI-TEMPORAL
AERIAL PHOTOGRAPHS
P. Hyvonen a *, J. Heinonen a , A. Haara b
a Finnish Forest Research Institute, P.O. Box 68, FI-80101 Joensuu, Finland - (pekka.hyvonen,
j aakko .heinonen)@metla.fi
b Dept, of Forest Science, University of Helsinki, FI00014, Helsinki, Finland - arto.haara@metla.fi
KEY WORDS: Change detection, aerial photograph, segmentation, A>NN, rectification
ABSTRACT:
The increased need for timely forest information is leading to the continuous updating of stand databases. In continuous updating,
stand attributes are estimated in the field following a forest operation (cutting or silvicultural treatment) and stored in databases. To
determine the changes caused by forest operations and forest damage, a semi-automatic method was developed based on bi-temporal
aerial photographs.
The field data consisted of 2 362 forest stands, from which the changes between years 2001 and 2004 were collected from different
databases. Stands were divided into three classes according to the type of change. The No-change class (1 890) included stands with
no changes other than growth. The Moderate-change class (373) included stands with changes such as thinning, partly operated
stand and improvement of young stand. The Considerable-change class (99) included stands with major changes such as clear
cutting and severe storm damage. The data were randomly divided into training and test data. The aerial photographs were acquired
for the years 2001 and 2004 with almost the same image specifications and the photographs were temporally registrated. As change
detection is sensitive to location errors, locational adjustments were made at the stand and segment levels. Linear stepwise
discriminant analysis and the non-linear ^-nearest neighbour (&-NN) method were tested in classification.
The classification results at the stand level were found to be better than at the segment level. Compared to previous studies, the
results of this study demonstrate remarkable improvement in the classification accuracy of moderate changes. The results showed
that change detection substantially improved when the registration at the stand level was used, especially in the detection of thinned
stands.
1. INTRODUCTION
Stand attributes in Finland have been traditionally gathered by
periodical field inventories with inventory cycles of 10-15
years. The increased need for timely forest information is
leading to the continuous updating of stand databases. In
continuous updating, a stand database is kept up-to-date
computationally using statistical growth models. After a forest
operation (i.e. cutting or silvicultural treatment), stand attributes
are estimated in the field.
A forest operation can be reported at the time of the work, but
forest damage, for example, must be determined by some other
method. There are also errors that should be controlled in the
databases. Medium-resolution satellite images (e.g. Landsat TM
images) have been successfully used for detecting considerable
changes such as clear cuttings, removals of hold-over trees, soil
preparations or drastic damage. The detection of moderate
changes such as thinnings, preparatory cuttings or slight
damage, has been difficult (Holmgren & Thuresson, 1998;
Wilson & Sader, 2002; Heikkonen & Varjo, 2004). The reason
for this is that typical thinnings, where about 20-40% of the
basal area is removed, cause only subtle changes in reflectance
(Olsson, 1994).
The reflectance captured by a single pixel in medium-resolution
images is the average of the reflectance from an area of more
than 100 m 2 on the ground. However, with high-resolution
remote sensing materials, the disappearance of individual trees
can be detected. For example, using data from airborne laser
scanning (ALS), Yu et al. (2004) found 61 of 83 harvested
trees. However, ALS is very expensive compared to aerial
photography.
Operational high-resolution applications for change detection of
vegetation cover are based on the visual interpretation of aerial
photographs, although more automatic methods have also been
proposed. Hudak and Wessman (1998) investigated a transition
from grassland to shrubland using historical aerial photographs.
The images were classified using variograms that characterised
the image texture. Changes were determined by post
classification comparison. Kadmon and Harari-Kremer (1999)
also studied vegetation dynamics using pixel-level spectral
classification and then averaging the results to larger cells and
differencing the cell values. In both studies, good results using
automatic change detection were achieved; but compared to the
context of the present study, the time-intervals were much wider
and the changes more radical. In the study of Saksa et al.
(2003), clear cuttings were detected using three approaches:
pixel-by-pixel differencing and segmentation, pixel block-level
differencing and thresholding, and presegmentation and
unsupervised classification. Each method was found to be
suitable for operational use. Hyppanen (1999) applied image
differencing to bi-temporal aerial photographs. In that study,
considerable changes were detected while moderate changes
were not. Consequently, the problem of how to detect moderate