20/11/2019 (1) Machine Learning in Remote Sensing | LinkedIn

1 Try Premium Free

Search for 1 Month

Machine Learning in Remote Sensing

Published on December 10, 2018

João Otavio Firigato

Python developer for Data Science, Machine 12 articles Follow
Learning and Remote Sensing

Hello, continuing our posts on AI, we will approach the use of Machine Learning in remote
sensing.

Several applications have been developed in the field of remote sensing image analysis
during the last decades. Besides well-known statistical approaches, many recent methods are
based on techniques taken from the field of machine learning. The roots of machine learning
in remote sensing date back to the 1990s. It was initially introduced as a way to automate
knowledge-base building for remote sensing. In their paper, Huang and Jensen talk about
how a knowledge-base was built using minimal input from human experts and then decision
trees were created to infer the rules from the human input for the expert system. The
generated rules were used at a study site on the Savannah River. The conclusion details how
the proposed machine-learning assisted expert system approach yielded the highest accuracy
compared to conventional methods at that time.
Messaging

https://www.linkedin.com/pulse/machine-learning-remote-sensing-joão-otavio/ 1/7
20/11/2019 (1) Machine Learning in Remote Sensing | LinkedIn

Search
After such similar developments machine learning was soon adopted as an important
1 tool by
Try Premium Free
for 1 Month
the remote sensing community. Now it is being used in all sorts of projects, from an
unsupervised satellite image scene classification to the classification of Australian native
forests.

Types of Machine Learning Algorithms

Machine learning in remote sensing comprises several different paradigms such as

classification, regression, clustering, feature extraction, dimensionality reduction and density
Like
estimation. Comment
These aspectsShare
are often interdependent, before performing a classification one
47 · 4 Comments

might extract some additional texture features and also reduce the dimensionality of the data
set with feature selection techniques. Perhaps the most commonly undertaken applications in
remote sensing are feature reduction, clustering and classification.

Feature extraction and selection

Terms such as feature extraction and feature selection are closely related to dimensionality
reduction, which refers to the mapping of the data from the original feature space into a
space of a lower dimension, without discarding any meaningful information. Furthermore,
meaningful information is defined according to the final application. Hyperspectral imaging,
for example, is characterized by high spectral resolution with up to hundreds of data
channels with narrow bands, ranging from the visible to the short-wave infrared region of
the electromagnetic spectrum. Although such data provide detailed spectral information,
theoretical and practical problems arise with increasing dimensionality of the data in the
spectral domain. The idea of the dimension is intuitive, driven by experiments in one-, two-
or three-dimensional space, and geometric concepts that are self-evident in these spaces do
not necessarily apply in higherdimensional space.

For example, normally-distributed data have a tendency to concentrate in the tails, which
seems to contradict its bell-shaped density function. For the purpose of classification, these
problems are related to the curse of dimensionality. In particular, with a limited training set,
beyond a certain limit, the classification accuracy decreases as the number of features
increases. This is paradoxical, since with a higher spectral resolution one can discriminate
more classes and have a finer description of each class, but the data complexity leads to
poorer classification. To mitigate this phenomenon, a dimensionality reduction is performed

Messaging

https://www.linkedin.com/pulse/machine-learning-remote-sensing-joão-otavio/ 2/7
20/11/2019 (1) Machine Learning in Remote Sensing | LinkedIn

Search
by feature selection and feature extraction. A feature reduction algorithms can1 be supervised
Try Premium Free
for 1 Month
or unsupervised.

Supervised transformation is in general well suited to pre-processing for the task of

classification, since the transformation improves class separation. However, its effectiveness
is correlated with how well the training set represents the whole data set. Moreover, this
transformation can be extremely time-consuming. Unsupervised methods do not focus on
class discrimination, but aim for another representation of the data in a lower-dimensional
space, satisfying some given criterion. For Principal Component Analysis (PCA), the data
are projected into a subspace that minimizes the reconstruction error in the mean square
sense.

PCA plays an important role in the processing of remote sensing images. Even though its
theoretical limitations for hyperspectral data analysis have been pointed out, in a practical
situation the results obtained using the PCA are still competitive for the purpose of
classification. The advantages of the PCA are its low complexity and the absence of
parameters. However, the PCA only considers the second-order statistic, which can limit the
effectiveness of this method.

Clustering algorithms

In general, cluster algorithms aim to indentify data (unknown) structure, such as natural
groups or clusters within the multidimensional feature space by measuring similarities
between different pixels. The pixels within a cluster or group are more similar to each other
than those pixels belonging to other clusters.

Thanks to kernel functions, the similarity can be computed easily. Moreover, it can be
adapted to the data by using an appropriate kernel. Nevertheless, the most common
similarity measurement is the Euclidean distance. However, when dealing with data that
Messaging

https://www.linkedin.com/pulse/machine-learning-remote-sensing-joão-otavio/ 3/7
20/11/2019 (1) Machine Learning in Remote Sensing | LinkedIn

Search
contains clusters of different shapes and sizes the definition of clusters and the1 selection of Try Premium Free
for 1 Month
an adequate similarity measure are critical and thus the clustering process can be difficult.
After measuring the similarity, the cluster algorithm merges the individual pixels into
groups. Of course, the same data set can result in several potential combinations of clusters
and some kind of accuracy assessment is required to evaluate the quality of the generated
clusters. A common quality measurement is the squared error measure that is also used as
the cluster criterion. Clustering techniques are also used for unsupervised classifications.
After generating groups of pixels with similar properties, the user usually assign a class label
to each cluster.

Clustering approaches, such as ISODATA and k-means are well known in remote sensing.
They usually assign each pixel to only one cluster (hard techniques), whereas fuzzy
approaches as the fuzzy c-mean results in a degree of membership for all clusters.

Other approaches as the agglomerative hierarchical clustering techniques do not rely on

specific cluster centres. In the beginning each pixel is handled as a separate cluster. During
the process neighbouring pixels (or clusters) are sequentially merged, depending on a
distance measurement. The process of clustering can be represented in a tree diagram or so-
called dendogram. In addition to agglomerative algorithms that sequentially reduce the
number of clusters, by merging similar groups (bottom-up approach), divisive techniques
procedures start with all the pixels belonging to one cluster and split it iteratively into
smaller clusters (bottom-down approach).

An alternative clustering approach is based on mixture models. This approach is based on

the assumption that each natural group is described by a different probability distribution.
Whereas the form of the distribution is similar for all groups, the parameter values are
different. In addition, a mixture can contain different models and more complex density
functions can be modelled.

Supervised classification algorithms

The major part of machine learning algortihms for remote sensing image analysis is aiming
perhaps for a supervised classification of the data.

Supervised classification is based on the idea that a user can select sample pixels in an
image that are representative of specific classes and then direct the image processing
software to use these training sites as references for the classification of all other pixels in
the image. Training sites (also known as testing sets or input classes) are selected based on
Messaging

https://www.linkedin.com/pulse/machine-learning-remote-sensing-joão-otavio/ 4/7
20/11/2019 (1) Machine Learning in Remote Sensing | LinkedIn

Search
the knowledge of the user. The user also sets the bounds for how similar other1 pixels must Try Premium Free
for 1 Month
be to group them together. These bounds are often set based on the spectral characteristics of
the training area, plus or minus a certain increment (often based on "brightness" or strength
of reflection in specific spectral bands). The user also designates the number of classes that
the image is classified into.

In supervised classification the majority of the effort if done prior to the actual classification.
Once the classification is run the output is a map with classes that are labeled and
correspond to information classes or land cover types. Supervised classification can be much
more accurate than unsupervised classification, but depends heavily on the training sites, the
skill of the individual processing the image, and the spectral distinctness of the classes. If
two or more classes are very similar to each other in terms of their spectral reflectance (e.g.,
annual-dominated grasslands vs. perennial grasslands), misclassifications will tend to be
high. Supervised classification requires close attention to development of training data. If
the training data is poor or not representative the classification results will also be poor.
Therefore supervised classification generally requires more times and money compared to
unsupervised.

Some examples of this kind of approach are: Maximum likelihood, minimum distance,
KNN, Suport Vector Machine, artificial neural network, decision tree classifier and random
forest.

After this brief introduction to AI and Machine learning, let's apply their techniques and
algorithms using python packages in the next posts...

Messaging

https://www.linkedin.com/pulse/machine-learning-remote-sensing-joão-otavio/ 5/7
20/11/2019 (1) Machine Learning in Remote Sensing | LinkedIn

Search
References: 1 Try Premium Free
for 1 Month

https://www.sciencedirect.com/science/article/pii/S1674987115000821
https://pythontips.com/2017/11/11/introduction-to-machine-learning-and-its-usage-in-
remote-sensing/ https://www.tandfonline.com/doi/full/10.1080/01431161.2018.1433343?
scroll=top&needAccess=true https://sia.planning.unc.edu/post/machine-learning-for-remote-
sensing/ http://gsp.humboldt.edu/olm_2015/courses/gsp_216_online/lesson6-
1/supervised.html

Report this

Published by

João Otavio Firigato 12 articles Follow

Python developer for Data Science, Machine Learning and Remote Sensing
Published • 1yr

#machinelearning #remotesensing #geospatialdata #unsupervisedlearning #supervisedlearning #imageclassification

#artificialinteligence #AI #ML

Reactions

+39

4 Comments

Add a comment…

Lateef Agbaje
Co-Founder at Apersoft Technologies, PhD candidate in Beihang university

Thank you …see more 7mo

Like Reply

Ahmed Mahdi Al-Jayashi

CEO at GEOMARK

Many thanks  (edited) 11mo

Like Reply 1 Like

Load more comments

João Otavio Firigato

Python developer for Data Science, Machine Learning and Remote Sensing

Follow

More from João Otavio Firigato

Cloud Computing for land cover Decision Tree Classifier: reducing Decision Tree for Satellite Image First steps to Machine Learning
classification - Part 1: How to overfitting with parameter tuning Classification
João Otavio Firigato on LinkedIn
download image from Google Eart…
João Otavio Firigato on LinkedIn João Otavio Firigato on LinkedIn
João Otavio Firigato on LinkedIn

See all 12 articles

Messaging

https://www.linkedin.com/pulse/machine-learning-remote-sensing-joão-otavio/ 6/7
20/11/2019 (1) Machine Learning in Remote Sensing | LinkedIn

1 Try Premium Free

Search for 1 Month

About Talent Solutions Community Guidelines

Careers Marketing Solutions Privacy & Terms

Advertising Sales Solutions Mobile

Small Business Safety Center

Questions? Select Language

Visit our Help Center. English (English)

Manage your account and privacy.

Go to your Settings.

LinkedIn Corporation © 2019

Messaging

https://www.linkedin.com/pulse/machine-learning-remote-sensing-joão-otavio/ 7/7