Landcover Classification

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 3

Landcover Classification

Rahul Delwar
Computer Science

Indian Institute of Technology Bombay


(of Affiliation)
Mumbai, India
rahuldelwar7@gmail.com

213050071

Abstract:- In recent years the area of Image classification There are different types of CNN 1d,2d, and 3d. In Subject
using remote sensing is in rapid growth as there is need of to HIS the different models are used to capture different
method which can detect pixel distinction with good accuracy features.
and less noise When it comes to image classification which
contains data of high wavelength the Deep learning models
B. Attention Mechanisms
present either perform badly with lot of noise or require lots
of parameters.. The attention mechanism was built based on the attention
mechanism our brain uses when watching some content
I. INTRODUCTION e.g. it focuses on some part of the data and ignores the rest
Landcover Classification:- Our land is covered with lots of of the data not completely but doesn’t give attention as
different objects (e.g. grass, lake, buildings ) and to classify much to rest. By using this mechanism the ML models
this using current technologies brings up a new challenge as have incorporated a feature called attention mechanism in
there are lots of spectral bands which are very complex. And which we assign the attention weights as how much the
as there is advancement in remote sensing technology there model should focus on a certain part of the data
are lots of types of images available few of which are basic
RGB image multispectral and hyperspectral images. So, for
0.1 0.5 0.01 0.3 0.8 0.6 0.2 0.07 Attention weight
this, we will be using Hyperspectral images to distinguish
between different types of land covers such as grass, lake, land The low value means less attention high value means high
etc. The current method uses image sensing with deep learning attention.
methods mostly CNN to classify the Hyperspectral images but
they have their disadvantages such as a 2d CNN is not very We have incorporated attention mechanisms in few of the
efficient when it comes to classifying hyperspectral images presented models such as octave CNN
and 3d CNN contains lots of parameters which makes itself
very complex to implement. Lot more models have been
developed with their own flaws and advantages like the Dual C. Feedback Connections
path network which is developed using the mixture of residual As in the brain, there are lots of feedback connections
network and dense convolution network. Spectral-spatial the same kind of structure is incorporated into neural
residual network i=which uses identity mapping to form a networks. Feedback connections are used mostly in
flow between other 3D conovultion layers. DPRN Deep CNN to refine the visual data as its passes through the
pyramidal residual networks, Spaectral-spatial capsule and layers lots of times. Feedback connections dramatically
few other models. increase the detail extracting capability in certain
Hyperspectral Images:- Hyperspectral images are images models.
that contain more spectral bands than a normal image and
more wavelength than the normal human eye can perceive.
This will help in classifying more accurately. As the III. PROPOSED MODELS FOR CLASSIFICATION OF HSI
Hyperspectral image Contains more information the
Overview of challenges
processing of HIS is overhead for the present ml models. Our
main concern is the complexity of current models for HSI
classification either they are too complex to implement or the 1. Presently we have lots of models for remote
output data received from them are not upto the standard. sensing Images such as 1d CNN, 2d CNN, 3D
CNN.
II. BACKGROUND 2. And all have had their advantages as 2d CNN is not
complex to implement it can only explore the
A. CNN
spatial features of the data.
CNN is the most popular Deep learning models present 3. 1D on the other hand can only explore the spectral
today for the data involving images. In the initial years of features.
AI, a model was introduced which was the first one to make
a way forward with the visual data and the at the core of the
model of Convolution Neural Networks engine.

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE


4. 3d CNN can analyze both spatial and spectral 2. While on the other hand 3D CNN becomes very
features but requires a lot more parameters and is complex. And requires lot more parameters which inevitably
complex to implement. takes up lots of memory.
5. Using 3d CNN to extract spatial and spectral Both models however efficient loses accuracy when it to
features results in too many parameters. Because classifying Hyperspectral images
the Hyperspectral images are very dense in nature.
So the model is developed named HybridSN which is the
A. HyperloopNet
combination of the complementary features of both 2d and
Most of the ML today is forward in nature means the flow 3D CNN , the main challenge was that to to introduce high
of data occurs only in the forward direction. Lots of poetries interclass variability and similarity simultaneously then
to come up with the idea of multi-direction feedback loop Principle component analysis is then applied over
mechanisms and Hyperloopnet is one of those but more Hyperspectral images data.
efficient. Key idea :- is to combine the features of both 2d and 3d CNN
Our HSI classification model named Hyperloopnet models to tackle the decripencies faced by the models on their
contains dense feedback loop connections through all own.
convolution layers. And the model is connected in parallel
form to manage information flow between all layers. The Given table compares the accuracy on Houstan 18 dataset
input from the same layer in the different iteration is the
function of the output previously computed. Class name Hyperloopnet OctaveCNN HybridSN
In Hyperloop net we used concatenation operators
with the addition operators. Healthy 94.02% 96.34% 94.65%
The valuation of the model is done on 4 grass
Hyperspectral Image datasets which are available online and water 100% 99.90% 100%
are Benchmarked to use. Unpaved 100% 100% 100%
parking lots
Key idea of the model is that the connection are between all cars 94.65% 95.34% 90.51%
blocks both in feedforward and feedback manner and on all
levels of the network. Which makes the model less complex Overall 76.64% 73.93% 72.15%
as there is reduction in parameters. accuracy
Average 87.91% 87.14% 85.08%
B. Multiscale Octave 3D CNN accuracy
kappa 65.91% 71.08% 68.08%
HSI have too many redundancies which results in too
many parameters if 3d CNN is Used. So we proposed a And we can see that hyperloop net overall is more accurate
new model name octave CNN which decomposes than the HybridSN and octave CNN it also achived the
feature maps output by CNN into different spatial highest kappa
frequency. Conclusion and open challenges
The method developed yet may have gotten some
progress in the field of HIS classification, but the research in In this paper we tried to compare 3 different models for
the area of spatial attention related to Hyperspectral Images Hyperspectral Image classification. The model which are
seem to be not evoloving far more than it started. already in use have some discrepancy as they produce lots of
To improve the feature maps a novel attention noise in the output data
method binary attentive module is used. The binary attention There is still needed a advancement in data with 3d
module helps in improving the performance and the main octave CNN as data enhancement can get rid of the problem
work is to capture the discriminative features within spectral of limit labeled Hyperspectral Image pixel samples.
and spatial images. On the other hand the hybrid SN model has combined the
2d and 3D CNN models to take the features of both and use
We have used different kernel in the octave CNN to capture it to integrate the spatio -spectral and spectral in different
multitype features which in turn also reduces the spatial convolution formations such as 3d and 2d. the hybrid SN
reduandancy And the introduced method is way more turned out to be efficient than both the CNN models. For the
efficient than the normal CNN for visual data. small training data hybridSN also have given fitter
performance.
Key idea:- The core power of the given hyperloopnet model is that it
can perform efficiently even when the parameters are very
C. Hybrid SN few in numbers compared to yet available different models.
The motivation to develop the hybridSN model came up There is a future scope of improvement in hyperloopnet
by analyzing both 2d and 3d CNN’s . In both the models there which is the absence of attention mechanism in the model.
are drawback that holds them to be the better model for HSI
classification.
ACKNOWLEDGEMENT
1. 2DCNN alone cannot extract a good distinction Feature
map from the spectral dimension. Prof. Biplab Banerjee (Hyperspectral Images)
Prof. Om Damani
REFERENCES
Roy, S.K., Krishna, G., Dubey, S.R., Chaudhuri, B.B., 2019. Hybrids: Exploring 3-d-2-
d cnn feat
Xu, Q., Xiao, Y., Wang, D., Luo, B., 2020. Csa-mso3dcnn: Multiscale octave 3d cnn
with
channel and spatial attention for hyperspectral image classification. Remote Sens.
12, 188.
Xu, X., Li, W., Ren, Q., Du, Q., Gao, L., Zhang, B., 2017. Multisource remote sensing
data
classification based on a convolutional neural network. IEEE Trans. Geosci.
Remote
Sens. 56, 937–94

You might also like