Internatonal Conference on Appled Scence and Engneerng Innovaton (ASEI 205) Desgn and Implementaton of Novel Agrcultural Remote Sensng Image Classfcaton Framework through Deep Neural Network and Mult- Feature Analyss YOUZHI ZHANG Remote Sensng Technque Center of Helongjang Academy of Agrculture Scences, Helongjang 50086, Chna Keyword: Agrcultural Remote Sensng; Image Classfcaton; Deep Neural Network; Feature Selecton. Abstract: Wth the rapd and burstng development of computer scence and sensor technology, effcent remote sensng (RS) mage classfcaton algorthm s ur-gently needed. There are plenty of applcatons of re-mote sensng mage processng technques. In ths paper, we propose a new agrcultural remote sensng mage classfcaton and recognton method based on sparse auto-encoder deep neural network. Usng an unsuper-vsed learnng algorthm features a large number of small peces of sparse auto-encoder learnng from some deep unlabeled mages have already completed the tranng neural networks, and then learn features. The experment and smulaton prove the correctness of our model compared wth other methods. INTRODUCTION. Background Survey Wth the rapd and burstng development of computer scence and sensor technology, effcent remote sensng (RS) mage classfcaton algorthm s urgently needed. There are plenty of applcatons of remote sensng mage processng technques whch could be categorzed as the followng parts. () RS mage segmentaton. In [], Yuan et al. conducted research on feature analyss. They present segmentaton solutons where representatve features are ether known or unknown. They also show that feature dmensons can be greatly reduced va subspace projecton. (2) RS mage super-resoluton. In [2], Alfred conducted research on ths topc. Ths presentaton gves an overvew of some essental steps. Class separablty s accounted for by means of controllng the balance tuned by a smoothness parameter λ between the pror and the lkelhood terms n the posteror energy functon. (3) RS mage change detecton. In [3], Badr s group conducted research on ths ssue, they ponted out that gbbs the proposed technque uses fuzzy markov random feld model of spatal gray level propertes of multspectral mage dfference. Change detecton problem s solved usng the maxmum a posteror probablty estmaton prncple. (4) RS mage regstraton. In [4] L et al. ntroduces novel feature selecton technques to handle the problem. In addton, each key feature descrptor s refned to overcome dfferent remote mage of the gradent between the strength and drecton. More applcatons such as de-haze and denosng technques are also belongng to the usage of RS applcatons. Remote sensng mage classfcaton s all lke yuan accordng to the feature of the mage s dvded nto several categores. Tradtonal mage classfcaton methods are manly supervsed and unsupervsed classfcaton, but the tradtonal unsupervsed remote sensng mage classfcaton method s accordng to the data set of potental smlarty clusterng and approprate measures, sometmes get better classfcaton results, and supervsed classfcaton requres a lot of tranng data set to desgn the classfer, f the tranng data set s not enough to estmate classfcaton parameters selected, often get better classfcaton effect. More related research could be found n the lteratures..2 Overvew of Our Work 205. The authors - Publshed by Atlants Press 025
Color remote sensng mage classfcaton based on SVM effect depends on the characterstcs, s used to color the characterstcs of remote sensng mage classfcaton has a lot of, the mage color, texture feature s the most commonly used two knds of feature vector. Hgh resoluton color remote sensng mage contans abundant nformaton, color s more smlar to some of the area, but ts texture feature s large, such as grass and forest land and some regonal texture s smlar, but the color s dfferent, such as roads and bare land. And the color and texture are only part of the remote sensng mage characterstcs, the author n the process of classfcaton, f you only use a sngle color or texture features to represent the mage nformaton, t does not fully descrbe the content of the mage contans. To deal wth ths hardshp, we propose the novel agrcultural remote sensng mage classfcaton framework through deep neural network and mult-feature analyss. 2 OUR PROPOSED METHODOLOGY 2. Overvew of Deep Learnng Snce the code s a knd of unsupervsed neural network learnng algorthm, t makes the output value of the sample s equal to the nput values. If the Numbers of hdden layer neurons s far less than neural network nput layer and output layer, thus forcng the codng of neural network to learn the nput data compresson, sad assumes that the neural network nput data s completely random data, then, to learn these random data compresson s very dffcult; But f mpled has a certan relatonshp between the nput data of the specfc structure, such as certan nput characterstcs are related to each other, so, the encodng algorthm can fnd the correlaton between the related data, and to reconstruct the nput data n the output layer. On the contrary, f the neural network hdden layer neurons number s more, or wth the nput and output layer s when, also can add a sparse sex of hdden layer neurons lmt, so the encodng neural network can stll fnd out the correlaton between the nput data. Assumng that the actvaton of neurons functon as the sgmod functon, when the output of the neuron s close to that the neurons s actve; Conversely, when the output of the neuron s close to zero that the neurons n the nhbtory state. Sparse sex of neural network, therefore, lmt refers to most of the tme lmt neurons n the nhbtory state. 2.2 The Deep Structure Analyss Fgure.The structure of tradtonal neural network (TNN) Bleachng s to reduce the phase correlaton between the nput mage pxels, as many algorthms for preprocessng step. ZCA bleachng n ths artcle, there are two man applcatons: the unmarked sample data preprocessng and learn the characterstcs of the vsualzaton. For a color remote sensng mages, the arbtrary a pxel color nformaton, and can be through the use of R, G, B value for combnaton sad, therefore, by extractng each pxel of R, G, B value, can get the color nformaton of mage features. Color features should be defned clear, easy to extract, smple calculaton, rela- 026
tve to other features, color features s very stable. The conducton functon could be expressed as the formula. S l ( l+ ) ( l) l = j j + j= z w x b () In the encodng neural network, stll use BP algorthm and make no label sample nput and output the results equal to target. Therefore, the unlabeled sample can be expressed as: 2,,..., m x x x,..., x, m (2) { } Through usng ( 2 ) j ( ) a x to denote the frst set of sample nput cases frst j a hdden layer neurons of the output value. The average j neuron actvaton value s: m ( 2 ρ ) j aj ( x ) = (3) m = Bleachng s to reduce the phase correlaton between the nput mage pxels, as many algorthms for preprocessng step. To jon regularzaton weghts n the regresson model cost functon attenuaton after tem, a new cost functon s convex functon, the exstence of mnmum feature learnng algorthm snce sparse codng s mnmum cost functon used when some of the teratve algorthm, such as batch gradent descent and Newton's method, LBFGS get the global optmal soluton. Fnally, wat for after the completon of the teratve algorthm and got the traned classfer. If the classfer output label s consstent wth the test sample, show that classfcaton results correctly, on the other hand, the classfcaton result error; The statstcal label test set classfcaton correct sample amount dvded by the total number of samples, get on the accuracy of mage classfcaton label test set. Therefore, the objectve functon could be revsed to be the formula 4. m 2 J( wb, ) = y hwb, ( x) + m = 2 nl sl sl+ 2 s2 l ( l) ( w ) pq + b KL ρ ρ 2 l= q= p= j= ( j ) (4) In the formula 3, t s composed of three parts and, among them, the frst part s the mean square error, the second part s the regularzaton tem, last part s the penalty term relatve entropy s a common method of measurng dfference between two dstrbutons. The depth of the neural network weght vector and the offset of the gradent descent drecton vector can be expressed as: ( ) T l ( l+ ) ( l) ( l) w = σ ( σ ) + lw m (5) m ( l) ( l+ ) b = σ m = As s shown n fgure 2, 400 hdden neurons after tranng after the correspondng feature n the vsual results, s dvded nto 20 columns show 20 lnes, each character says of the hdden unts learned n dfferent poston and drecton of the mage edge detecton. 027
Fgure 2.The learned feature through deep neural network (NDD) and overall selecton procedure 2.3 Mult-feature Analyss In remote sensng mage classfcaton, mage texture feature s also a knd of commonly used feature vector. Commonly used texture characterstcs are manly texture features, gabor rpple and gray level co-occurrence matrx form. Of gray level co-occurrence matrx through the study of spatal gray level characterstcs of mage texture nformaton, t can not only reflect the dstrbuton of the brghtness of the mage features, but also can reflect the same or close to the brghtness of the pxel locaton between dstrbuton characterstcs, and gradually become an mportant method used for analyzng the characterstc of mage texture. The defnton of gray level co-occurrence matrx s expressed as: Pδθ, j Pj = (6) N N2 P = j= δθ, j The SVM algorthm based on texture feature must be used for mage or has obvous phenomenon of texture regons. For those who have no obvous phenomenon of texture area or mage, and the extracton of texture feature not only correctly descrbe the underlyng characterstcs of the mage, and can lead to ncorrect results. The color and texture features are combned together, therefore we can get a 7 dvson feature vector as follows: f = [ RG,, BW,, W2, W3, W4] (7) Comprehensve features usng the SVM classfcaton algorthm combnng the color features and texture features of mages, t s not only applcable to all knds of don't change color between mages of the obvous, but also for all knds of color nformaton between don't close to the mage, due to ts combnaton of texture nformaton, s stll able to mage for better classfcaton. For a natural mages, mage wthn the scope of local statstcal characterstcs smlar to those of other parts, so there's no need to let all the connectons between nput layer and hdden layer neurons, n order to reduce the neural network model parameters caused by the hgh resoluton mage of too much problem, convoluton s appled n operaton, so as to realze local to connect to the Internet, the basc dea s based on local connected network vsual cortex of the bran neurons only response to the stmulaton of some local area, only allow local connecton neural network hdden layer neurons connected to the part of the nput layer neurons. And for mage texture feature s not obvous, because the combnaton of the color of the mage nformaton, stll can for better classfcaton, compared wth the former two knds of algorthm whch more wdely applcable. The correspondng steps could be vsually descrbed as the followng fgure 3. 028
Fgure 3.The descrpton of our proposed method 029
3 EXPERIMENT AND SIMULATION 3. Set-up of the Experment The smulaton envronment s ntalzed as the follows. Sx physcal machnes equpped wth 4 TB hard dsk and 6 GB of RAM, and the smulaton software s nstalled on Wndows Wn8 platform and Intel core 2 quad core 3.0 GHz and 6 GB of RAM. Comprehensve features usng the SVM classfcaton algorthm combnng the color features and texture features of mages, t s not only applcable to all knds of don't change color between mages of the obvous, but also for all knds of color nformaton between don't close to the mage, due to ts combnaton of texture nformaton, s stll able to mage for better classfcaton. And for mage texture feature s not obvous, because the combnaton of the color of the mage nformaton, stll can for better classfcaton, compared wth the former two knds of algorthm whch more wdely applcable. 3.2 Smulaton Result The followng fgures shows the result of classfcaton usng our method and other relate algorthms. () (2) (3) (4) (5) (6) (7) (8) (9) Fgure 4.The expermental result of the research 030
Fgure 5.The statstcal descrpton of the smulaton 4 CONCLUSION AND SUMMARY Wth remote sensng mage classfcaton and recognton of agrcultural problems and t s hard to dstngush between smlar objects, ths paper proposes a new agrcultural remote sensng mage classfcaton and recognton method based on sparse auto-encoder deep neural network. Usng an unsupervsed learnng algorthm features a large number of small peces of sparse auto-encoder learnng from some deep unlabeled mages have already completed the tranng neural networks, and then learn features, features can be extracted from the massve mages and wndng and gatherng. The expermental results show that the new method of mage classfcaton can be more effectvely dstngush between forest fre and ts object, smlar to that of the red flag, red leaves, than the tradtonal neural network, etc. In the future, we plan to refer to more related work to deal wth the ssue. The references we wll combne are prmary n the lteratures [5-8]. 5 REFERENCE [] Yuan, Jangye, DeLang Wang, and Rongxng L. "Remote sensng mage segmentaton by combnng spectral and texture features." Geoscence and Remote Sensng, IEEE Transactons on 52, no. (204): 6-24. [2] Sten, Alfred. "Advanced remote sensng mage analyss wth super resoluton mappng." In South Afrcan Symposum on Numercal and Appled Mathematcs. 205. [3] Subudh, Badr Narayan, Francesca Bovolo, Ashsh Ghosh, and Lorenzo Bruzzone. "Spatocontextual fuzzy clusterng wth markov random feld model for change detecton n remotely sensed mages." Optcs & Laser Technology 57 (204). [4] Fatyga, Mrek, et al. "A comparson of three Deformable Image Regstraton Algorthms n 4DCT usng conventonal contour based methods and voxel-by-voxel comparson methods." Name: Fronters n Oncology 5 (205): 7. [5] Mendoza, Nusvel Acosta, et al. "A Nectar of Frequent Approxmate Subgraph Mnng for Image Classfcaton Un nectar sobre la mnería de subgrafos frecuentes aproxmados en clasfcacón de mágenes." Revsta Cubana de Cencas Informátcas 9. (205). [6] H. Wang and J. Wang, An effectve mage representaton method usng kernel classfcaton, n Tools wth Artfcal Intellgence (ICTAI), 204 IEEE 26th Internatonal Conference on, Nov 204, pp. 853 858. [7] Ge, ZongYuan, et al. "Modellng Local Deep Convolutonal Neural Network Features to Improve Fne-Graned Image Classfcaton." arxv preprnt arxv:502.07802 (205). [8] Tran, Duc Toan, et al. "An approach for combnng multple descrptors for mage classfcaton." Seventh Internatonal Conference on Machne Vson (ICMV 204). Internatonal Socety for Optcs and Photoncs, 205. 03