• No results found

Vol 7, No 10 (2017)

N/A
N/A
Protected

Academic year: 2020

Share "Vol 7, No 10 (2017)"

Copied!
7
0
0

Loading.... (view fulltext now)

Full text

(1)

Research Article

a

October

2017

Computer Science and Software Engineering

ISSN: 2277-128X (Volume-7, Issue-10)

Deep Learning Based High-Resolution Remote Sensing

Image Classification

Sumit Kaur*,

Research Scholar, Gurukashi University, TalwandiSaboo, India

Dr. R.K Bansal

Dean Research, Gurukashi University, TalwandiSaboo, India

Abstract- Deep learning is an emerging research area in machine learning and pattern recognition field which has been presented with the goal of drawing Machine Learning nearer to one of its unique objectives, Artificial Intelligence. It tries to mimic the human brain, which is capable of processing and learning from the complex input data and solving different kinds of complicated tasks well. Deep learning (DL) basically based on a set of supervised andunsupervised algorithms that attempt to model higher level abstractions in data and make it self-learning for hierarchical representation for classification. In the recent years, it has attracted much attention due to its state-of-the-art performance in diverse areas like object perception, speech recognition, computer vision, collaborative filtering and natural language processing. This paper will present a survey on different deep learning techniques for remote sensing image classification.

Keywords: remote sensing images, deep learning, image classification

I. INTRODUCTION

The swift advancement in remote sensing image processing area, make the availability of remote data in bulk, in different spatial and spectral resolutions and in dynamic ranges. In recent years, there have been great advances in remote sensing image processing, for both low-level tasks, such as denoising or segmentation, and high level ones, such as classification. Remote sensing image classification is a obligatory step for remote sensing applications like forest management, disaster warning and assessment, military target recognition, thematic mapping, environment monitoring, urban planning. There are a plethora of image classification algorithms have been developed with strong conceptual foundation based on pixels spectral and spatial characteristics like; conventional classifiers, neural networks, machine learning algorithms, genetic algorithms and artificial intelligence. Labeling each pixel of an image to a particular class according to its spectral properties becomes incrementally more challenging as the level of abstraction increases, going from pixel to pixel which is the reason of occurrence of mixed pixels in and image. Classifying mixed pixels is the hardest part to process an remote sensed image because in some cased as high intra class variability then couples with low inter class distance, , a problem that grows ever more as finer classifications are sought, which result the wrong labeling of the pixels. For this problem increasing the resolution is not the solution, theencoding spectral, textural, and geometrical properties, become mostly ineffective. More complex features and descriptors are necessary to capture the semantics of the scene[1], [2], [3]. To improve the results of classification which gives birth to deep learning approach?

Deep learning comes from the concept of human brain having multiple types of representation with simpler features at the lower levels and high-level abstractions built on top of that. Humans arrange their ideas and concepts hierarchically. Humans first learn simple concepts and then compose them to represent more abstract ones. The human brain is like deep neural network, consisting of many layers of neurons which act as feature detectors, detecting more abstract features as the levels go up. This way of representing information in a more abstract way is easier to generalize for the machines.

II. DEEP LEARNING

(2)

ISSN(E): 2277-128X, ISSN(P): 2277-6451, pp. 22-28

face verification in wild conditions. A key commitment of this work is to straightforwardly learn social visual features,which demonstrate character similitudes, from crude pixels of face sets with a cross breed profound system [8]. The profound ConvNets in our model copy the essential visual cortex to mutually separate nearby social visual elements from two face pictures contrasted and the educated channel sets.Li Deng and John C[5]. Platt survey presents that deep learning systems have dramatically improved the accuracy of speech recognition, and various deep architectures and learning methods have been developed with distinct strengths [7],[16] and weaknesses in recent years. Deep learning is typically applied to computer vision, speech recognition, and NLP. These are non-linear classification problems [10] where the inputs are highly hierarchal in nature. In 2011, Google Brain project, created a neural network trained with deep learning algorithms, which recognized high level concepts, like cats, after watching just YouTube videos and without being told what a "cat" is. Facebook [6] is creating solutions using deep learning expertise to better identify faces and objects in the photos and videos uploaded to Facebook each day[13,[14],[15]. Another example of deep learning in action is voice recognition like Google Now and Apple’s Siri. According to Google, the voice error rate in the new version of Android stands at 25% lower than previous versions of the software after adding insights from deep learning. Deep learning hypothesis [9] demonstrates that deep nets have two distinctive exponential points of interest over exemplary learning calculations that don't utilize circulated representations. Both of these favorable circumstances emerge from the force of organization and rely on upon the hidden information producing dispersion having a fitting componential structure. To start with, learning dispersed representations empower speculation to new blends of the estimations of educated elements past those seen amid preparing (for instance, 2n mixes are conceivable with n twofold elements). Second, making layers out of representation [8] in a deep net brings the potential for another exponential preferred standpoint (exponential in the profundity). Deep architectures [12] are composed of multiple levels of non-linear operations, such as in neural nets with many hidden layers or in complicated propositional formulae re-using many sub-formulae. Searching the parameter space of deep architectures is a difficult task, but learning algorithms such as those for Deep Belief Networks have recently been proposed to tackle this problem with notable success, beating the state-of-the-art in certain areas [11].

The early works for scene classification area unit principally supported handcrafted options. These strategies principally specialise in employing a right smart quantity of engineering skills anddomain experience to style numerous human-engineering options, admire colour, texture shape, abstraction and spectral data, or their combination that area unit the first characteristic of a scene image and thence carry helpful data used for scene classification. Here, we tend to concisely review many most representative handcrafted options, together with colour histograms , texture descriptors , GIST , scale invariant feature rework (SIFT) , and bar chart of directed gradients (HOG) Among all handcrafted options, the world colour bar chart feature is sort of the best, nevertheless a good visual feature normally utilized in image retrieval and scene classification. A serious advantage of colour histograms, except for their ease to cipher, is that they're invariant to translation and rotation concerning the viewing axis. However, colour histograms aren't ready to convey the abstraction information, therefore it's terribly tough to tell apart the photographs with similar colours however different colour distribution. Besides, colour bar chart feature is additionally sensitive to tiny illumination changes and quantisation errors.Most of these progressive approachestypicallyconsidersupervised learning to getsensible feature representations. Particularly in 2006, a breakthrough in deep feature learning was created by Hinton and Salakhutdinov. Since then, the aim of researchers has been to interchange hand-engineered options with trainable multilayer networks and a quantity of deep learning models have shown spectacular feature illustration capability for a goodvary of applications as well as remote sensing image scene classification. On the one hand, as compared with ancient handcrafted optionsthose needa substantialquantity of engineering ability and domain experience, deep learning optionsarea unitmechanically learned from informationemploying ageneral learning procedure via deep-architecture neural networks. This is often the key advantage of deep learning strategies. On the opposite hand, compared with aforesaidunsupervised feature learning strategies that area unittypically shallow-structured models (e.g., thin coding), deep learning models that area unit composed of multiple process layers will learn a lot of powerful feature representations of knowledge with multiple levels of abstraction.

(3)

ISSN(E): 2277-128X, ISSN(P): 2277-6451, pp. 22-28

Figure 1.1 Following is the Deep learning process for remote sensing image processing

Associate in Nursing SAE consists of multiple layers of autoencoders within which the outputs of every layer are wired to the inputs of the ordered layer. To coachAssociate in Nursing SAE, a possiblemanner is to use greedy layer-wise coachingtheme. Specifically, one ought toinitial train the primary layer on raw computer fileto get parameters Associate in Nursingd transfer the data into an intermediate vector consisting of activations of the hidden units. Then, this method is recurrent for ensuant layers by victimisation the output of every layer as input for the next layer. This technique trains the parameters of every layer separatelywhereastemperature reduction parameters for the rest of the model. to gethigher results, when layer-wise coaching is completed, fine-tuning is performed to tune the parameters of all layers at identical time with a smaller learning rate. Compared to one autoencoder as mentioned in previous segment, the feature illustration power of SAE may beconsiderablyreinforced. This could be simply explained: with the composition of multiple autoencoder that every transforms the illustration at one level (starting with the raw input) into an illustration at a better, slightly a lot of abstract level, we are able to learn terribly powerful representations. This has been proved in literatures [13].CNNs are designed to methodinformation that are available inthe shape of multiple arrays, maybe a multi-spectral image composed of multiple 2nd arrays containing constituent intensities within the multiple band channels. Beginning with the spectacular success of AlexNet ,several representative CNN models together with Overfeat, VGGNet , GoogLeNet , SPPNet and ResNet areprojectedwithin the literature. There exist four key ideas behind CNNs that benefit of the properties of natural signals, namely, native connections, shared weights, pooling, and therefore the use of the many layers.

(4)

ISSN(E): 2277-128X, ISSN(P): 2277-6451, pp. 22-28

Table1.1: Summary of deep learning techniques Paper Application

Area

Deep Learning Algorithms used

Tools Merits

[9] Security  DBN with Dropout

Cuckoo sandbox

The New domain is tried as malware detection

[8] [10] [11] [12] [13] [14]

Image Processing

 CNN,

 Stacked Denoising autoencoders Autoencoder, Deep SVM and GMM,

 SVR with Linear Rectifier Units,

 Deep NuralNetwork with SVM, SDA and SVM

CAFFE, LibSim

Classification is easier for uncertain data and reduced the errors, Performance is equal to the performance of DBN, DAE are helpful for learning of higher level representations, Easier classification of the uncertain data, better accuracy than normal SVM, Simple for implementation,

Preprocessing is not required. [15] Regression

Analysis

 Deep SVM - Experimented on 10 regression datasets. Performance and accuracy is improved

[16] [17] Finance  Deep SVM with Fuzzy,

 Deep Neural Network with Genetic

Algorithm

Keras Accuracy and performance is

improved, Ant Colony Optimization and bees algorithm. Better accuracy

[18] [19] [20]

Big Data:

 Traffic Network Analysis

 Image processing

 Education

 SAE with Logistic regression,

 SAE and SVM

RapidMiner tool, neural network toolbox and Matlab

Better results than SVM

[21] Music Genere Classification

 Deep Feed forward network

 LRU

- Used rectifier linear units as the activation function.Used preprocessing of data to reduce the dimension.

[22] Handwritten character recognition

 DBM Matlab Best RBM variants in terms of the classification errors and it is time consuming and the tuning of parameters is very difficult for the users.

[23] Network Analysis  Restricted Boltzmann Machines

Matlab RBM is used to do link prediction and node classification of linked data that is a network data like Facebook network.

[24] Sentiment Classification

 Deep network using RBM

- Used RBM with unsupervised learning

[25] Credit Card Analysis  Deep Neural Network and SVM

(5)

ISSN(E): 2277-128X, ISSN(P): 2277-6451, pp. 22-28

[26] [27] [28]

Big Data  Tensor Auto Encoder,

 Deep SVM,

 Deep Neural Network

MATLAB 2012a

Takes more time to train the parameters, Accuracy is increased, new visualization method is demonstrated.

[29] Time Series Forecasting

 DBN with SVR - Provided best training construction result and most accurate prediction results.

[30] Multimedia  DCNN LIBSVM Solve the structure output problem.

III. CONCLUSION

In spite of the fact that Deep Networks are proposed in 1996 their utilization has picked up fame from 2006. In view of progressions in handling abilities of machine and parallel preparing structure, execution of Deep Networks and Profound Learning Algorithms appears to be conceivable. Profound Learning calculations are observed to be exceptionally helpful in the range of Picture Processing, Signal Processing and Natural Language Processing. Be that as it may, now there is a need to apply these ideas in new zones. With the headway of Big Data new difficulties are rising and Deep Learning is demonstrating valuable in taking care of these difficulties. From the table it is apparent that new application territory challenges are settled adequately with the assistance of Deep Learning calculations. It will be extremely fascinating to test these calculations in still more new ranges to get precise outcomes and forecast.

REFERENCES

[1] M. Pesaresi and A. Gerhardinger, “Improved textural built-up presence index for automatic recognition of human settlements in arid regions with scattered vegetation,” IEEE Journal of Selected Topics on

EarthObservation and Remote Sensing, vol. 4, no. 1, pp. 16–26, March 2011.

[2] I.A. Rizvi and B.K. Mohan, “Object-based image analysis of highresolution satellite images using modified cloud basis function neural network and probabilistic relaxation labeling process,” IEEE Transactions on

Geoscience and Remote Sensing, vol. 49, no. 12, pp. 4815–4820, December 2011.

[3] R. Gaetano, G. Masi, G. Poggi, L. Verdoliva, and G. Scarpa, “Markercontrolled watershed-based segmentation of multiresolution remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing, vol.53, no. 6, pp. 2987–3004, June 2015.

[4] Sun, Yi, Xiaogang Wang, and Xiaoou Tang. "Hybrid deep learning for face verification."Proceedings of the IEEE International Conference on Computer Vision. 2013.

[5] LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. "Deep learning." Nature 521.7553 (2015): 436-444

[6] ALISON RUSHTON, A. University of Birmingham,

[7] Edgbaston,Birmingham, UK : Formative assessment: a key to deep learning? , 2007.

[8] Attias, Hagai, et al. "Speech denoising and dereverberation using probabilistic models." Advances in neural information processing systems (2001): 758-764.

[9] Coates, A., Lee, H., and Ng,A. Y. An analysis of singlelayer networks in unsupervised feature learning. In AISTATS 14,2011.

[10] LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. "Deep learning." Nature 521.7553 (2015): 436-444. [11] He, Kaiming, et al. "Deep residual learning for image recognition." Proceedings of the IEEE Conference on

Computer Vision and Pattern Recognition. 2016.

[12] Dr P.Kiran Sree, et al. “Investigating an Artificial Immune System to Strengthen the Protein Structure Prediction and Protein Coding Region Identification using Cellular Automata Classifier”, International Journal of Bioinformatics Research and Applications ,Vol 5,Number 6,pp 647-662, ISSN : 1744-5493. (2009) (Inderscience Journals , UK )

(6)

ISSN(E): 2277-128X, ISSN(P): 2277-6451, pp. 22-28

[14] Dr P.Kiran Sree, et al. “ Fast Multiple Attractor Cellular Automata with Modified Clonal Classifier for Coding Region Prediction in Human Genome”, Journal of Bioinformatics and Intelligent Control, Vol. 3, 2014, pp 1- 6. DOI:10.1166/jbic.2014.1077 (American Scientific Publications, USA)

[15] Dr P.Kiran Sree, et al. “A Fast Multiple Attractor Cellular Automata with Modified Clonal Classifier Promoter Region Prediction”, in Eukaryotes.Journal of Bioinformatics and Intelligent Control, Vol. 3, 1–6,

[16] Ciresan, D. C., Meier, U., Gambardella, L. M., and Schmidhuber, J. Deep big simple neural nets excel on handwritten digit recognition. CoRR, 2010.Kien Nguyen, Clinton Fookes, SridhaSridharan , “Improving Deep Convolutional Neural Networks WithUnsupervised Feature Learning,”ICIC2015, 978-1-4799-8339-1/15,2015 [17] Omid E. Davidt,Nathan S. Netanyahu , “DeepSign: Deep Learning for Automatic Malware Detection”,IEEE

,2015, 978-1-4799-1959-8/15

[18] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie,YoshuaBengio, Pierre-Antoine Manzagol , “Stacked DenoisingAutoencoders: Learning Useful Representations in a Deep Network with a Local Denoising Criterion”, Journal of Machine Learning Research 11 (2010) 3371-3408

[19] Sangwook Kim ,Minho Lee,Jixiang Shen , ”A novel deep learning by combining discriminative model withgenerative model, ” IEEE, 2015

[20] KentaroKuwata and RyosukeShibasaki , ”Estimating crop yields with deep learning and remotely sensed data,”IGARSS , 2015

[21] M.A. Wiering, M. Schutten, A. Millea, A. Meijster, and L.R.B. Schomaker , “Deep Learning using LinearSupport Vector Machines,” International Conference on Machine Learning 2013: Challenges inRepresentation Learning Workshop. Atlanta, Georgia, USA, 2013

[22] Francis Quintal Lauzon, “An introduction to deep learning,”The 11th International Conference on InformationSciences, Signal Processing and their Applications: Special Sessions

[23] M.A. Wiering, M. Schutten, A. Millea, A. Meijster, and L.R.B. Schomaker ,“Deep Support Vector Machines forRegression Problems”

[24] Yue Deng, Zhiquan Ren, Youyong Kong, Feng Bao and Qionghai Dai, “A Hierarchical Fused Fuzzy DeepNeural Network for data classification,”IEEE,2016

[25] Luna M. Zhang, “Genetic Deep Neural Networks Using Different Activation Functions for Financial DataMining,” IEEE International Conference on Big Data (Big Data),2015

[26] YishengLv, YanjieDuan, Wenwen Kang, Zhengxi Li, and Fei-Yue Wang, Fellow , “Traffic Flow Prediction WithBig Data:A Deep Learning Approach,” IEEE Transactions On Intelligent Transportation Systems, VOL. 16,NO. 2, APRIL 2015

[27] Yao Ju, Jun Guo, Shuchun Liu , “A Deep Learning Method Combined Sparse Autoencoder with SVM,” IEEE,2015

[28] Bo Guo∗, Rui Zhang, Guang Xu, Chuangming Shi and Li Yang , “Predicting Students performance in educationdata mining,” IEEE-ISET 2015

[29] Arjun Raj Rajanna, KameliaAryafar, Ali Shokoufandeh and Raymond Ptucha, “Deep Neural Networks: A CaseStudy for Music Genere Classification,” IEEE 14th International Conference on Machine Learning andApplications,2015

[30] XiaoweiGuo , Haiying Huang , Jason Zhang , “Comparison of Different Variants of Restricted BoltzmannMachines,” 2nd International Conference on Information Technology and Electronic Commerce(ICITEC),2014

[31] Kang Li, Jing Gao, SuxinGuo, Nan Du, Xiaoyi Li and Aidong Zhang , “LRBM: A Restricted BoltzmannMachine based Approach for Representation Learning on Linked Data,” IEEE International Conference onData Mining 2014

[32] Shusen Zhou a,n, QingcaiChen b, XiaolongWang b , “Active deep learning method for semi-supervisedsentiment classification,” Elsevier, Neurocomputing, 2013

[33] AyahikoNiimi , “Deep Learning for Credit Card Data Analysis,” World Congress on Internet Security (WorldCIS-2015) 2015

[34] Qingchen Zhang, Laurence T. Yang, and ZhikuiChen , “Deep Computation Model for Unsupervised featurelearning on Big Data,” IEEE TRANSACTIONS ON SERVICES COMPUTING, VOL. 9, NO. 1, JANUARY/FEBRUARY 2016

(7)

ISSN(E): 2277-128X, ISSN(P): 2277-6451, pp. 22-28

[36] B.M. Wilamowaski, Bo Wu., JanuszKorniak , “Big data and deep learning,” 20th Jubilee IEEE International Conference on Intelligent Engineering Systems , June 30-July 2, 2016

[37] XuehengQiu, Le Zhang, Ye Ren and P. N. Suganthan , “Ensemble Deep Learning for Regression and TimeSeries Forecasting,” IEEE 978-1-4799-4512-2, 2014

Figure

Figure 1.1 Following is the Deep learning process for remote sensing image processing

References

Related documents

5 Zmenu DN kódu respektívne jej hodnoty prevediete pomocou tla č idiel teploty.. 6 Po zmene hodnoty DN kódu stla č te tla č idlo FIX na

The entrepreneurial competence indicators for business department students include a total of 10 major competences, which are entrepreneurial spirit and entrepreneurial

designed DTSA has been optimized in the array environment in HFSS using unit cell model, as shown in Figure 6(a), for the desired specifications mentioned previously.. The

What challenges does climate change present to smallholder coffee producers in each country, both in terms of impacts to the quality and reliability of coffee production, but also

Forward-looking statements included in the MD&A include, but are not limited to, statements related to the Corporation’s forecast gross consolidated capital expenditures for

The main approached aspects were the following ones: sheep and goat livestock evolution in the analyzed period, their distribution by macro and micro- regions, number of sheep

As the functions of miRNA can be inferred based on its gene targets, we extracted the known human gene targets from CLASH dataset [ 45 ], miRTarBase [ 46 ] and DIANA-TarBase [ 47 ]