Research on oil-palm using remote sensing
A variety of research has been done using
remote sensing for oil-palm plantation management.
Tan, Kanniah, & Cracknell (2013) used
remote sensing data to determine the age of oil-palm tree, one of the
significant factor which influences the fruit bunch production. Several
techniques were discussed in this study and it concludes that texture
measurement using Grey-level co-occurrence matrix and fraction of shadow are
the most significant for studying the age of oil-palm trees.
Srestasathiern & Rakwatin (2014) used high
resolution satellite images to detect oil-palm plants by computing vegetation
index. Peaks were detected using a 2D semi-variogram technique for the image
texture characterization. The main contribution of this research is the use of feature
selection process to select vegetation index which can best distinguish between
oil-palm and background. This method
claims a detection accuracy of 90%.
Harnessing the ability of higher spatial
resolution of airborne imagery Shafri et al., (2011) used a multiple step approach to count oil-palm plants from the data
with 1m spatial resolution. This semi-automated approach includes spectral
analysis for discriminating oil-palm and non-oil-palm regions followed by
texture analysis, edge enhancement, image segmentation, morphological analysis
and blob analysis giving an accuracy of 95%. However determining the exact
threshold during the segmentation stage is very crucial to the end results in
Some research efforts have been applied for
detecting diseased oil-palms plants. Santoso et al. (2011) used high
resolution Quick bird imagery to map Basel stem rot disease in oil-palm trees.
In this study image segmentation using red band was utilized to discriminate
oil-palm and non-oil-palm zones and then the performance of different
vegetation indices was investigated to differentiate between healthy and
infected oil-palm trees. It concludes that performance of vegetation indices
was acceptable in the late development stage of disease when oil-palm shows
severe symptoms however it was only about 35% in areas with low infection rates
and it varied in different fields depending upon the age of palm and infection
Zulhaidi et al. (2009) used
airborne hyper spectral data to detect diseased oil-palm plants by
investigating different vegetation indices and red edge techniques. This study
shows that the highest accuracy achieved was 84% using Lagrangian Interpolation
technique (red edge technique) and 80% using Modified simple Ratio (vegetation
Using airborne hyper-spectral imaging system Jusoff (2009) produced
a thematic map of oil-palm plantation in Malaysia. Using the appropriate band
combinations, supervised classification was performed to classify image based
on the reflectance using Spectral Angle Mapper (SAM) algorithm. Ground verification
data showed that the thematic map was able to distinguish between healthy,
stressed and dead plants with an accuracy of 93%.
Early detection of diseased oil-palm plants
using remote sensing image analysis by relying only on spectral data is challenging.
Some of the reasons include that the oil-palm plant canopy does not provide
good spectrum of disease symptoms and it also requires stable sunlight and long
duration to capture significant spectral signatures (Chong, Kanniah, Pohl, & Tan, 2017). Thus
there is need of high spatial accuracy as well which could be achieved by using
on oil-palm using Machine Learning
Machine learning techniques have also been
used by some researchers for oil-palm monitoring studies. To detect palm trees Malek et al. (2014) used Scale-invariant Feature Transform (SIFT) to extract
key points from UAV images with spatial resolution of 3.5cm. Key points were
then analyzed using a pre trained extreme learning machine (ELM) classifier. Using
an active contour method based on level sets, key points were merged together
and to distinguish oil-palm from other vegetation local binary patterns were
used for textural classification of regions. The proposed approach gave an
accuracy of 89-96% on the test data. The study concludes that to screen
distinctive key points, SIFT is a useful approach but as a feature vector it is
less competitive for RGB images.
Manandhar, Hoegner, & Stilla (2016) used the
concept of circular autocorrelation of the polar shape matrix (CAPS) representation
of an image as the shape feature to detect palm tree. To standardize and reduce
dimension of feature, linear support vector machine was used. In the final
stage local maximum detection algorithm was used on the spatial distribution of
standardized feature for detecting palm trees in aerial images. The performance
of the method is stated as 84% on test images. The research concludes that
feature extraction efficiency of CAPS is low as its extraction is a non-linear
Miserque Castillo et al. (2016) used sliding window to generate candidate windows from the images
obtained through UAV to count oil-palm trees in aerial images. A model was
trained using a dataset of 2950 samples having 450 positive and 2500 negative
samples. Local binary patterns (LBP) was used to model the texture of images and
a logistic regression was used to classify the image. The classifier gives a
detection rate of 95.34%.
on oil-palm using Deep Learning
In recent year deep learning has been used for
different applications in the domain of precision agriculture. Irrespective of
the traditional remote sensing techniques deep learning has proven to give
higher accuracy in research related to oil-palm monitoring. Cheang et al. (2017) used
convolutional neural networks for counting palm trees using Digital Globe
satellite imagery. The study claims to achieve tree count accuracy of 94.5%.
This study concludes that satellite imagery gives poor results as compared to
UAV images because of lower resolution, cloud cover and availability issues.
The use of high resolution data from UAVs and
airborne platforms has recently increased as it provides higher spatial
resolution which is very significant for precision farming. Li, Fu, Yu, & Cracknell (2016) discussed a comparison of different approached to count oil-palm trees
in high resolution Aerial Images. A sliding window approach was used to predict
labels of all the samples from an image dataset as oil-palm and background.
Convolutional Neural Network gave an accuracy of 96% which was higher than
other two traditional approaches (template matching and local maximum filter)
used in the research.
As discussed above most of the research
done for oil-palm plantation management using deep learning deals with
detecting or counting the oil-palm trees in an image but none of the research effort
has been applied to use convolutional neural networks for detecting diseased