Identifying Land Patterns from Satellite Images using Deep Learning
Somnath Rakshit and Soumyadeep Debnath
{somnath52,soumyadebnath13}@gmail.com
Department of Computer Science and Engineering, Jalpaiguri Government Engineering College, Jalpaiguri, India
Objectives
In our proposed work, we aim to achieve the following:
• Analyze the dataset containing satellite imagery of the
Amazon rainforests, released by Planet and SCCON.
• Use the VGG16 model to classify the images into one
of the 17 pre-defined classes with high accuracy.
• Use this system to label satellite imagery in an
automated manner so as to track the changing land
patterns of Amazon.
Introduction
Deforestation in the Amazon rainforests have increased con-
siderably in the past few decades. This has affected its biodi-
versity and climate adversely. Using satellite imagery, it has
been possible to track the changes taking place within a re-
gion. However, due to the presence of huge amount of data,
it requires considerable manual resources for proper labelling.
In our proposed work, we develop an automated way to label
satellite images with their corresponding class of land cover
using the VGG16 model and achieve high accuracy.
(a) Roads + Primary (b) Habitation + Partly cloudy
(c) Agriculture + Roads + Primary (d) Agriculture + Pasture +
Primary + Partly cloudy
Figure: Sample chips and their labels
Dataset and Preprocessing
The dataset for this work has been derived from Planet’s full-
frame analytic scene products using its 4-band satellites in sun-
synchronous orbit (SSO) and International Space Station (ISS)
orbit. It has 17 type of labels for each of the chips. They are:
• agriculture
• artisinal_mine
• bare_ground
• blooming
• blow_down
• clear
• cloudy
• cultivation
• habitation
• haze
• partly_cloudy
• primary
• road
• selective_
logging
• conventional_
mine
• slash_burn
• water
The obtained dataset contains the images of various dimen-
sions. Thus, we resize all images to a standard size, in this
case, 128x128 pixels. This is also an important step as it helps
in speeding up the training. In this dataset, we have used
40479 images for training and 40669 images for testing. Each
image may be classified into multiple classes. We analyzed the
dataset as below:
(a) Distribution of training labels
(b) Correlation matrix
Figure: Data Analysis
Methodology
In our proposed work, we use the VGG16 model to classify im-
ages into various classes. Here, we use 20% of the training data
for validation while training after each epoch. The architecture
of our model is as below:
Table: Architecture of VGG16 model
Layer (type) Output Shape Param
input_1 (InputLayer) (None, 128, 128, 3) 0
batch_normalization_1 (None, 128, 128, 3) 12
vgg16 (Model) (None, 4, 4, 512) 14714688
flatten_1 (Flatten) (None, 8192) 0
dense_1 (Dense) (None, 17) 139281
Here, we use the Adam optimizer to minimize the loss, which
is measured by binary cross-entropy, with a learning rate of
10−4
. We use a batch size of 128 and train this for 15 epochs.
By this time, the training loss converges. Using an NVIDIA
Tesla K80 GPU, this takes around one hour to train.
Figure: Plot of Training Loss vs Epoch
Results
In our experiment, we obtain a training loss of 6.88%, training
accuracy of 97.35% and testing accuracy of 96.71%.
Also, we obtain an F-beta score of 92.69%. The F-beta score
is a weighted harmonic mean of the precision and recall. An
F-beta score reaches its best value at 1 and worst score at 0.
Conclusion
In this work, we have demonstrated a way to classify satel-
lite imagery in an automated manner using deep learning with
the help of the VGG16 model. We have achieved a high ac-
curacy with this consuming one hour while training with an
NVIDIA Tesla K80 GPU. This model can be successfully ap-
plied to track the changing land pattern in the rainforests of
Amazon. This data about the location of deforestation and hu-
man encroachment on forests can help governments and local
stakeholders respond more quickly and effectively.
Future Scope
The following additions may be made to our work to improve
its robustness.
• Using a larger neural network.
• Increased preprocessing of the dataset.
• Performing data augmentation to make the system more
robust.
References
[1] Wen Yang, Xiaoshuang Yin, and Gui-Song Xia.
Learning high-level features for satellite image classification with
limited labeled samples.
IEEE Transactions on Geoscience and Remote Sensing,
53(8):4472–4482, 2015.
[2] Emmanuel Maggiori, Yuliya Tarabalka, Guillaume Charpiat, and
Pierre Alliez.
Convolutional neural networks for large-scale remote-sensing image
classification.
IEEE Transactions on Geoscience and Remote Sensing,
55(2):645–657, 2017.
Acknowledgements

Identifying Land Patterns from Satellite Images using Deep Learning

  • 1.
    Identifying Land Patternsfrom Satellite Images using Deep Learning Somnath Rakshit and Soumyadeep Debnath {somnath52,soumyadebnath13}@gmail.com Department of Computer Science and Engineering, Jalpaiguri Government Engineering College, Jalpaiguri, India Objectives In our proposed work, we aim to achieve the following: • Analyze the dataset containing satellite imagery of the Amazon rainforests, released by Planet and SCCON. • Use the VGG16 model to classify the images into one of the 17 pre-defined classes with high accuracy. • Use this system to label satellite imagery in an automated manner so as to track the changing land patterns of Amazon. Introduction Deforestation in the Amazon rainforests have increased con- siderably in the past few decades. This has affected its biodi- versity and climate adversely. Using satellite imagery, it has been possible to track the changes taking place within a re- gion. However, due to the presence of huge amount of data, it requires considerable manual resources for proper labelling. In our proposed work, we develop an automated way to label satellite images with their corresponding class of land cover using the VGG16 model and achieve high accuracy. (a) Roads + Primary (b) Habitation + Partly cloudy (c) Agriculture + Roads + Primary (d) Agriculture + Pasture + Primary + Partly cloudy Figure: Sample chips and their labels Dataset and Preprocessing The dataset for this work has been derived from Planet’s full- frame analytic scene products using its 4-band satellites in sun- synchronous orbit (SSO) and International Space Station (ISS) orbit. It has 17 type of labels for each of the chips. They are: • agriculture • artisinal_mine • bare_ground • blooming • blow_down • clear • cloudy • cultivation • habitation • haze • partly_cloudy • primary • road • selective_ logging • conventional_ mine • slash_burn • water The obtained dataset contains the images of various dimen- sions. Thus, we resize all images to a standard size, in this case, 128x128 pixels. This is also an important step as it helps in speeding up the training. In this dataset, we have used 40479 images for training and 40669 images for testing. Each image may be classified into multiple classes. We analyzed the dataset as below: (a) Distribution of training labels (b) Correlation matrix Figure: Data Analysis Methodology In our proposed work, we use the VGG16 model to classify im- ages into various classes. Here, we use 20% of the training data for validation while training after each epoch. The architecture of our model is as below: Table: Architecture of VGG16 model Layer (type) Output Shape Param input_1 (InputLayer) (None, 128, 128, 3) 0 batch_normalization_1 (None, 128, 128, 3) 12 vgg16 (Model) (None, 4, 4, 512) 14714688 flatten_1 (Flatten) (None, 8192) 0 dense_1 (Dense) (None, 17) 139281 Here, we use the Adam optimizer to minimize the loss, which is measured by binary cross-entropy, with a learning rate of 10−4 . We use a batch size of 128 and train this for 15 epochs. By this time, the training loss converges. Using an NVIDIA Tesla K80 GPU, this takes around one hour to train. Figure: Plot of Training Loss vs Epoch Results In our experiment, we obtain a training loss of 6.88%, training accuracy of 97.35% and testing accuracy of 96.71%. Also, we obtain an F-beta score of 92.69%. The F-beta score is a weighted harmonic mean of the precision and recall. An F-beta score reaches its best value at 1 and worst score at 0. Conclusion In this work, we have demonstrated a way to classify satel- lite imagery in an automated manner using deep learning with the help of the VGG16 model. We have achieved a high ac- curacy with this consuming one hour while training with an NVIDIA Tesla K80 GPU. This model can be successfully ap- plied to track the changing land pattern in the rainforests of Amazon. This data about the location of deforestation and hu- man encroachment on forests can help governments and local stakeholders respond more quickly and effectively. Future Scope The following additions may be made to our work to improve its robustness. • Using a larger neural network. • Increased preprocessing of the dataset. • Performing data augmentation to make the system more robust. References [1] Wen Yang, Xiaoshuang Yin, and Gui-Song Xia. Learning high-level features for satellite image classification with limited labeled samples. IEEE Transactions on Geoscience and Remote Sensing, 53(8):4472–4482, 2015. [2] Emmanuel Maggiori, Yuliya Tarabalka, Guillaume Charpiat, and Pierre Alliez. Convolutional neural networks for large-scale remote-sensing image classification. IEEE Transactions on Geoscience and Remote Sensing, 55(2):645–657, 2017. Acknowledgements