Automated Diatom Classification (Part A): Handcrafted Feature Approaches

applied sciences A icle Au oma ed Dia om Classi ica ion (Pa A): Handc a ed Fea u e App oaches Glo ia Bueno 1,*, Osca Deniz 1, Anibal Ped aza 1, Jesús Ruiz-San aqui e ia 1, Jesús Salido 1, Gab iel C is óbal 2, Ma ía Bo ego-Ramos 3and Saúl Blanco 3 1VISILAB-Uni e si y o Cas illa-La Mancha, A . Camilo José Cela s/n, 13071 Ciudad Real, Spain; Osca [email p o ec ed] (O.D.); [email p o ec ed] (A.P.); [email p o ec ed] (J.R.-S.); [email p o ec ed] (J.S.) 2Ins i u e o Op ics, Spanish Na ional Resea ch Council (CSIC), Se ano 121, 28006 Mad id, Spain; [email p o ec ed] 3The Ins i u e o he En i onmen , Uni e si y o Leon, E-24071 León, Spain; [email p o ec ed] (M.B.-R.); [email p o ec ed] (S.B.) *Co espondence: [email p o ec ed] Recei ed: 31 May 2017; Accep ed: 18 July 2017; Published: 25 July 2017 Abs ac : This pape deals wi h au oma ic axa iden i ica ion based on machine lea ning me hods. The aim is he e o e o au oma ically classi y dia oms, in e ms o pa e n ecogni ion e minology. Dia oms a e a kind o algae mic oo ganism wi h high biodi e si y a he species le el, which a e use ul o wa e quali y assessmen . The mos ele an ea u es o dia om desc ip ion and classi ica ion ha e been selec ed using an ex ensi e da ase o 80 axa wi h a minimum o 100 samples/ axon augmen ed o 300 samples/ axon. In addi ion o published mo phological, s a is ical and ex u al desc ip o s, a new ex u al desc ip o , Local Bina y Pa e ns (LBP), o cha ac e ize he dia om’s al es, and a log Gabo implemen a ion no es ed be o e o his pu pose a e in oduced in his pape . Resul s show an o e all accu acy o 98.11% using bagging decision ees and combina ions o desc ip o s. Finally, some phycological ea u es o dia oms ha a e s ill di icul o in eg a e in compu e sys ems a e discussed o u u e wo k. Keywo ds: ea u e analysis; ex u al ea u es; mo phological ea u es; au oma ic classi ica ion; handc a ed app oaches; dia oms 1. In oduc ion Dia oms a e a majo g oup o algae and a e among he mos common mic oo ganisms in ma ine and eshwa e habi a s. They a e impo an con ibu o s o he p ima y p oduc ion in aqua ic ecosys ems, placed a he bo om o he ood chain. The dia oms ha e been shown o be inc easingly impo an wo ldwide in s udies ela ed o clima e change, as well as in he de elopmen o unc ions ha allow he modeling o such change. Mo eo e , hey a e good indica o s o en i onmen al condi ions and a e commonly used in wa e quali y assessmen [1,2]. Dia om indices a e known o co ela e mo e signi ican ly wi h wa e chemical a iables, wi hin con inen al wa e s, while mac oin e eb a e- o plan -based me hods a e mo e sensi i e o changes a ec ing s uc u al pa ame e s [3]. Dia oms ha e se e al ad an ages o e o he indica o s ha make hem ideal as indica o s o wa e quali y. These ea u es a e: (a) hei abili y o sp ead o e a a ie y o habi a s; (b) hey a e ela i ely easy o sample, and such sampling has no impac on he ecosys em du ing collec ion; (c) hey ha e a quick esponse o a ia ion in en i onmen al condi ions; and (d) hey a e sensi i e o changes in en i onmen al condi ions ha may no be obse ed in o he communi ies. Appl. Sci. 2017,7, 753; doi:10.3390/app7080753 www.mdpi.com/jou nal/applsci Appl. Sci. 2017,7, 753 2 o 22 Dia om cells a e enclosed wi hin a unique silica cell wall known as a us ule made up o wo al es, which i in o each o he like a pill box. The us ules show a wide di e si y o shapes and sizes, hough mainly, hey can be di ided in o cen ic dia oms ( adial symme y, ounded) and penna e dia oms (bila e al symme y, elonga ed). The main axonomic ea u es used in dia om iden i ica ion a e ela ed o he mo phology and o namen a ion o he us ule. The p esence o aphe and he o namen a ion o he us ule, s igma a and o he ea u es a e impo an in iden i ying hese o ganisms [4]. The dia om size, in he ange o 2–2000 µ m, is sui able o obse a ion o mos species using an op ical mic oscope. Howe e , dia oms equi e specialized skills o hei classi ica ion, ha is ained dia omis s (phycologis s specializing in dia om axonomy). As s a ed in [ 5 ], mo e han 200,000 dia om species a e es ima ed o exis , al hough jus hal o hem ha e been desc ibed. Se e al in e calib a ion es s ha e shown ha he esul s o biological indices based on dia oms a e highly sensi i e o he le el o accu acy in axonomic classi ica ion [ 6 ]. The iden i ica ion ask is e y di icul due o he huge numbe o species es ima ed o exis [7]. Classical axonomic diagnosis is pe o med using key ea u es o by isual compa ison wi h ype samples o e e ence iconog aphies [ 8 ]. The con en ional app oach is o analyze hese mic oalgae using ligh mic oscopy (b igh ield, DIC, RIC, e c.). Dia om-based me ics a e calcula ed based on he ela i e abundance o di e en axa in an assemblage and he au ecological pa ame e s cha ac e izing each species. Howe e , he cu en manual analysis o images is edious, equi ing highly quali ied s a , and i is ime consuming. This is he case when dia oms a e used in he con ex o wa e quali y, as in his s udy. Acco ding o a Eu opean di ec i e, in o de o compu e an index sco e, he iden i ica ion o a minimum o 400 al es pe sample is equi ed [9]. In he case o anspa en specimens such as dia oms, b igh ield mic oscopy p esen s some di icul ies. Some de ails a e ba ely dis inguishable om he backg ound, and some o he al e na i e modali ies such as phase con as , DIC o da k ield need o be conside ed. The e a e ew species o he Ni zschia, which a e good examples o ha (see Ni zschia cos ei,Ni zschia us ulum a us ulum and Ni zschia inconspicua in Figu e 2, Images 68, 72 and 73). This is p obably due o a no well-de eloped silici ica ion p ocess. Phase con as mic oscopy is a sui able echnique o isualizing anspa en specimens. Howe e , i p oduces some a i ac s (halos and shading-o ) ha limi i s use ulness in some applica ions. Halos a e un esol ed images wi h e e se con as , and shading-o is a con as -dec easing e ec om he edge o he specimen owa ds he cen e o i [ 10 ]. DIC mic oscopy is also a popula me hod o imp o ing he con as o uns ained specimens. DIC is absen o he halo e ec ound in phase con as mic oscopy p oducing pseudo- elie images ha can be unde s ood as he de i a i e o he op ical pa h leng h de ined as p oduc o he index o e ac ion imes he specimen hickness. In da k ield mic oscopy, he specimen is illumina ed wi h a hollow cone o ligh , which is oo wide o en e he objec i e lens. Da k ield is a sui able modali y o dia om isualiza ion ha can be a subs i u e o phase con as o DIC in many cases, al hough bo h a e ou o he scope o his wo k. Scanning elec on mic oscopy is ano he sui able modali y, especially in axonomy o e ealing s uc u al de ails [ 11 ], which is also ou o he scope o his pape . On he o he hand, mos o he dia om da abases ha a e publicly a ailable use b igh ield mic oscopy. O he di icul ies a ached o mic oscopy a e: images pa ially ocused and mul iple o ien a ions ( iews). Bo h a e ela ed o he p ojec ion o 3D objec s in o 2D images. The main challenges aced by au oma ic iden i ica ion and classi ica ion me hods a e due no only o he high numbe o species o ecognize, bu o he g ea simila i ies be ween hem and e en he p esence o polymo phisms wi hin species. In some cases, analysis is simply un easible due o he huge amoun o in o ma ion and images in ol ed. Ad ances in digi al mic oscopy and image analysis sys ems o e a po en ially ad an ageous solu ion compa ed o manual me hods o coun ing and classi ica ion. Be o e going u he , i is wo h poin ing ou he di e ence be ween classi ica ion and iden i ica ion in e ms o biology and pa e n ecogni ion e minology. In biology, i may be said ha Appl. Sci. 2017,7, 753 3 o 22 he e m iden i ica ion answe s he ques ion: ‘Wha is he name o he axon in on o me?’. Howe e , classi ica ion answe s ques ions o he so : How is his axon ela ed o o he axa? Acco ding o he pa e n ecogni ion e minology, a class is a se o objec s ha wi hin a gi en con ex is ecognized as simila . Such a class has usually a unique name, he class name. The indi idual objec s wi hin a class ha e a label ha e e s o his name. Addi ionally, classi ica ion is he assignmen o a class name o an objec by e alua ing a ained classi ie o ha objec . The e o e, in his s udy, ou objec s o classes a e he dia oms, and we will classi y hem o gi e hem a label wi h hei axon name. Hence o h, a dia om axa will be e e ed o as a class. A good e iew o mo phome ic me hods o shape analysis and landma k-based analysis used in dia om esea ch is p esen ed in Pappas e al. [ 12 ]. Ano he a emp o desc ibe he mo phology and geome y o dia oms is he wo k o Klos e e al. [ 13 ], who de eloped a sys em ha allows he segmen a ion and ea u e ex ac ion om dia om con ou s, al hough i does no p o ide ex u al in o ma ion abou he us ule. As men ioned by Pappas e al. [ 12 ], wo a eas o s udy may be iden i ied wi h ega d o he me hods used in dia om esea ch, namely shape analysis and pa e n ecogni ion. We will ocus on pa e n ecogni ion o machine lea ning me hods, bu no ice ha he adi ional machine lea ning me hods a e based on a p e ious de ini ion o a se o ea u es desc ibing he objec s o be classi ied ha may o may no ha e a biologically-meaning ul in e p e a ion. Me hods o dia om de ec ion and iden i ica ion ha e been s udied in Cai ns e al., 1979 [ 14 ], Cul e house e al., 1996 [ 15 ], Pech-Pacheco and Al a ez-Bo ego, 1998 [ 16 ], Pech-Pacheco 2001 [ 17 ]. Cai ns e al. [ 15 ] ha e p oposed some dia om iden i ica ion me hods based on cohe en op ics and holog aphy. Howe e , such wo k did no ha e any impac on dia om esea ch mainly due o he ac ha he iden i ica ion sys em was oo specialized and p obably oo expensi e o be used by a dia omis communi y [ 18 ]. Cul e house e al. [ 15 ] de i ed some me hods o phy oplank on iden i ica ion based on neu al ne wo ks, bu again, hey do no p o ide a ully-au oma ic me hod. Pech-Pacheco e al. [ 16 ] ha e p oposed a hyb id op ical-digi al me hod o he iden i ica ion o i e di e en species o phy oplank on h ough he use o ope a o s in a ian o ansla ion o a ion and scale. Pappas and S oe me 2003 [ 19 ], used o m desc ip o s by Legend e polynomials and p incipal componen analysis in he iden i ica ion o he Cymbella cis ula species. An impo an a emp o au oma e dia om classi ica ion was conduc ed o he ADIAC p ojec (Au oma ic Dia om Iden i ica ion And Classi ica ion) [ 18 , 20 ]. Se e al accu acy esul s we e epo ed wi h a da abase composed o di e en numbe s o dia om axa anging om 37–55 classes. In ADIAC, 171 ea u es we e used o dia om classi ica ion. These ea u es a e in ended o desc ibe he dia om symme y, shape, geome y and ex u e by means o di e en desc ip o s, such as: ec angula i y, ci cula i y, compac ness, shape o poles, leng h, wid h, leng h-wid h a io, size, s ia densi y o ien a ion, ho izon al equency, G ay-Le el Co-occu ence Ma ix (GLCM), momen in a ian s, Gabo wa ele s, Fou ie and Scale-in a ian Fea u e T ans o m (SIFT) desc ip o s. The classi ie s ha pe o m be e a e bagging o decision ees and andom o es o p edic i e clus e ing ees, all o hem e alua ed wi h 10- old c oss- alida ion (10 c ). The bes esul s, up o 97.97% accu acy, we e ob ained wi h 38 classes using Fou ie and SIFT desc ip o s wi h andom o es . Pe o mance dec eased down o 96.17% when classi ying 55 classes wi h he same desc ip o s and classi ie [21]. New echniques based on Con olu ional Neu al Ne wo ks (CNN) ha e also been explo ed o classi y sea plank on (Kuang 2015 [ 22 ], and Dai e al., 2016 [ 23 ]). No e, howe e , ha hese images a e di e en han hose s udied in his pape since his ype o plank on a ies om phy oplank on (dia oms). In [ 22 ], a da abase o 30,000 images belonging o 121 classes was used. The esul s we e poo wi h a maximum pe o mance o 73.90%. In [ 23 ], a da abase o 30,000 images belonging o 33 classes was used. They ob ained an accu acy up o 96.3%. The o he wo k ela ed o CNN applied o dia oms is he one published by he au ho s (see he nex pape companion [ 24 ]). The wo k p esen ed he e and he me hodology ha e been compa ed o he CNN app oach [24]. Thus, mos o he e o s a e s ill ca ied ou wi h handc a ed app oaches o “hand-designed” me hods whe e a se o ixed ea u es is used. Tha is, he me hods ely on expe knowledge o Appl. Sci. 2017,7, 753 4 o 22 ex ac he mos ele an ea u es e sus CNN app oaches ha lea n ea u es om da a. Howe e , s ill, he handc a ed me hods p esen limi ed esul s as in [ 25 ], whe e 14 classes we e classi ied wi h Suppo Vec o Machine (SVM) 10 c , using 44 GLCM ea u es ha desc ibe only geome ic and mo phological p ope ies. They ob ained an accu acy o 94.7%. The e o e, he au oma ed classi ica ion o dia oms (in e ms o pa e n ecogni ion) o axon iden i ica ion emains a challenge. A p esen , he e is no sys em capable o aking in o accoun a ia ions in bo h he con ou and he ex u e in a ela i ely la ge numbe o species. One o he easons is he di icul y in acqui ing a big da ase o agged da a wi h a su icien numbe o samples pe species. The classi ica ion o dia oms is edious and labo ious, e en o an expe dia omis . In his pape , we p esen a comple e s udy o ele an ea u es o desc ibe and classi y dia oms. The main pu pose is o de ine he mos disc iminan ea u es and o make a compa ison o classi ie s based on hese ea u es e sus he CNN app oach. Fo ha , we collec ed an impo an da abase o 80 dia om axa wi h 300 samples pe axon desc ibed in Sec ion 2. The da abase was composed o an a e age o 100 dis inc dia oms pe class and augmen ed by means o compu a ional simula ions up o 300 samples pe class. In o de o ex ac he main dia om ea u es, we p opose in Sec ion 3 a segmen a ion o apply desc ip o s o he con ou and inne dia om egions. In Sec ion 4, a comple e lis o desc ip o s is p o ided. Those a e handc a ed ea u es ha desc ibe, in e ms o compu e ision, he disc iminan p ope ies o dia oms. Sec ions 5and 6desc ibe classi ica ion s a egies and some classi ie s. Expe imen al esul s a e p esen ed in Sec ion 7whe e an o e all accu acy o 98.11% is p esen ed, which imp o es p e ious ela ed wo ks. Finally, Sec ion 8concludes he pape add essing un esol ed challenging p oblems. 2. Ma e ials: Da ase P epa a ion Once ha ing collec ed he dia om samples om he i e s, he chemical ea men o he sample is ca ied ou in he labo a o y wi h hyd ogen pe oxide (120 ol.), which causes he diges ion o he o ganic ma e and allows one o ob ain suspensions o us ules and al es ee o o ganic emains. The p ocess is done a a empe a u e o 70–90 ◦ C, o accele a e he eac ion. A ew d ops o he sample a e aken and deposi ed in a ound co e slip. A e e apo a ion o he wa e , he dia om us ules emain in he co e -objec s. Then, using a syn he ic esin (Naph ax) wi h an op ical e ac i e index o 1.7, dia oms a e a ached o he glass slide o la e classi ica ion unde b igh ield mic oscopy ollowing s anda d p o ocols [9]. The 80 dia om species s udied he e we e collec ed du ing he yea s 2003 and 2015 om he Due o Basin wa e in Spain [ 26 ]. Those a e he 80 dominan axa in e ms o ela i e abundance and occu ence. A Wes bu y SP/40 B unel mic oscope and a B unel AMA 050 came a we e employed o cap u e he images a 60 × magni ica ion, wi h a nume ic ape u e o 0.85 and a physical esolu ion o 7.91 pixels/ µ m. An a e age o 100 dis inc dia om al es pe each dia om class we e hen manually c opped and labeled by an expe dia omis . The exac numbe o he c opped dia om al es is shown in Table 1. To comple e he da ase wi h up o 300 image samples pe axon, a da a augmen a ion was pe o med by means o applying o a ions o 90 ◦ , 180 ◦ , 270 ◦ and up-down and igh -le lips o he c opped images. These 6 ans o ma ions we e pe o med on he o iginal images and only i needed, o ob ain up o 300 image samples pe class. Thus, we end up wi h 24,000 images o classi y in o 80 dia om axa. Da a augmen a ion aims a inc easing he numbe o images in he da ase by ep esen ing image da a in di e en o ien a ions. Tha is, di e en copies o he same image a e made, bu om di e en pe spec i es o isual angles. Ro a ing and mi o ing he da a in di e en o ien a ions may e en ually help wi h iden i ying a simila objec in di e en o ien a ions. A mo e obus classi ie will be ob ained i he da a a e andomly o a ed in mul iple o ien a ions. In Figu e 1, a cap u e wi h a manually-selec ed dia om is shown. The elemen s ha o m he o namen a ion o he us ule in a dia om a e illus a ed in he c opped dia om. Fea u es o he s ia a e key in dia om axonomy, such as: a eola and lineolae. A eola is a pe o a ion (o po e) in he dia om Appl. Sci. 2017,7, 753 5 o 22 al e, and lineolae a e a eola elonga ed in he apical di ec ion. The lineolae densi y is calcula ed in ppm (pixels pe mic on). E e y sample was manually c opped o ensu e he bes dia om samples a oiding as much as possible nea by samples o deb is. The lis o he 80 species wi h he numbe o o iginal selec ed images is indica ed in Table 1, and some examples a e depic ed in Figu e 2. The numbe o he axon co esponding o he one lis ed in Table 1is shown in he uppe le co ne o each pic u e. The da abase can be ob ained by eques (see he con ac a h p://aquali as- e os.es/en/), and i will be publicly a ailable a he end o he p ojec . Figu e 3shows he same dia om species a di e en iews (“ al a iew” and “gi dle iew”) and sizes. Due o he deposi ion p ocess o he sample, in mos si ua ions, he dia oms appea in al a iew, al hough some imes appea in la e al iew (less han 10% o cases). Figu e 1. Dia oms obse ed by a mic oscope a 60 × magni ica ion and hei main elemen s. The o iginal image size is 903 ×614 pixels, and he selec ed dia om sample is 138 ×85 pixels. Table 1. Lis o he 80 dia om species analyzed in he cu en s udy, showing he numbe o al es pe class. 1. Achnan hes subhudsonis 123 28. Encyonema minu um 120 55. Gomphonema hombicum 64 2. Achnan hidium a omoides 129 29. Encyonema eicha d ii 152 56. Humidophila con en a 105 3. Achnan hidium ca a elense 59 30. Encyonema silesiacum 108 57. Ka aye ia cle ei a cle ei 84 4. Achnan hidium ca ena um 187 31. Encyonema en icosum 101 58. Lu icola goeppe iana 136 5. Achnan hidium d ua ii 93 32. Encyonopsis alpina 106 59. Mayamaea pe mi is 40 6. Achnan hidium eu ophilum 97 33. Encyonopsis minu a 89 60. Melosi a a ians 146 7. Achnan hidium exile 98 34. Eolimna minima 174 61. Na icula c yp o enella 136 8. Achnan hidium jackii 125 35. Eolimna hombellip ica 132 62. Na icula c yp o enelloides 107 9. Achnan hidium i ula e 305 36. Eolimna subminuscula 94 63. Na icula g ega ia 50 10. Ampho a pediculus 117 37. Epi hemia adna a 72 64. Na icula lanceola a 77 11. Aulacosei a suba c ica 113 38. Epi hemia so ex 85 65. Na icula ipunc a a 99 12. Cocconeis linea a 81 39. Epi hemia u gida 93 66. Ni zschia amphibia 124 13. Cocconeis pediculus 49 40. F agila ia a cus 93 67. Ni zschia capi ella a 123 14. Cocconeis placen ula a euglyp a 117 41. F agila ia g acilis 54 68. Ni zschia cos ei 72 15. C a icula accomoda 86 42. F agila ia pa a umpens 74 69. Ni zschia dese o um 71 16. Cyclos ephanos dubius 85 43. F agila ia pe minu a 89 70. Ni zschia dissipa a a media 81 17. Cyclo ella a omus 99 44. F agila ia umpens 49 71. Ni zschia ossilis 76 18. Cyclo ella meneghiniana 103 45. F agila ia auche iae 82 72. Ni zschia us ulum a us ulum 226 19. Cymbella excisa a angus a 79 46. Gomphonema angus a um 86 73. Ni zschia inconspicua 255 20. Cymbella excisa a excisa 241 47. Gomphonema angus i al a 55 74. Ni zschia opica 65 21. Cymbella excisi o mis a excisi o mis 142 48. Gomphonema insigni o me 90 75. Ni zschia umbona a 91 22. Cymbella pa a 177 49. Gomphonema mic opumilum 89 76. Rhoicosphenia abb e ia a 94 23. Den icula enuis 181 50. Gomphonema mic opus 117 77. Skele onema po amos 155 24. Dia oma mesodon 115 51. Gomphonema minusculum 158 78. S au osi a binodis 94 25. Dia oma monili o mis 134 52. Gomphonema minu um 93 79. S au osi a en e 87 26. Dia oma ulga is 88 53. Gomphonema pa ulum sap ophilum 52 80. Thalassiosi a pseudonana 70 27. Discos ella pseudos ellige a 82 54. Gomphonema pumilum a elegans 128 Appl. Sci. 2017,7, 753 6 o 22 Figu e 2. Examples o he 80 dia om axa classi ied in his s udy. Sample images a e s e ched o isualiza ion pu poses. Appl. Sci. 2017,7, 753 7 o 22 (a) (b) (c) Figu e 3. Di e en iews and sizes o he same species: ( a )Gomphonema insigni o me; ( b )Ni zschia ossilis; and (c)Rhoicosphenia abb e ia a. 3. Val e Segmen a ion: Bina y Th esholding The e a e se e al wo ks abou dia om de ec ion mainly ela ed o he abo e-men ioned ADIAC p ojec . I is ou o he scope o his pape o p esen all o he segmen a ion me hods; o u he de ails, he eade is e e ed o he e iews men ioned in Sec ion 1[ 12 , 13 , 18 ]. In his wo k, we p esen an au oma ic me hod used o do an ini ial quick segmen a ion, which equi ed isual supe ision a e wa d o include only he co ec segmen ed dia oms. Some o he images ha e been acqui ed wi h low con as and backg ound noise, which p oduces a poo segmen a ion in e ms o al e o e lapping wi h o he s uc u es. Tha means isual supe ision is needed o disca ding segmen a ion e o s. The al e is he mos signi ican egion o he dia oms whe e s uc u al di e ences can be dis inguished. The e o e, he segmen a ion p ocess should accu a ely ex ac such a egion o ex ac ing ele an ea u es. A p ope segmen a ion o he al e is expec ed o a ec ex u al, equen ial and s a is ical desc ip o s. This segmen a ion is done he e by means o a bina y h esholding whe e he segmen ed egion is he bina y masks whe e desc ip o s mus be compu ed. The p ocess o ob ain he bina y masks consis s o ou s eps: 1. Bina y h esholding: au oma ic segmen a ion based on O su’s h esholding. 2. Maximum a ea: calcula ion o he la ges egion (a ea). 3. Hole illing: in e io holes a e illed i p esen , using ma hema ical mo phology ope a o s. 4. Segmen a ion: he ROI is c opped wi h he coo dina es o he bounding-box o he la ges a ea (S ep 2). Due o he p esence o deb is and o e lapping wi h di e en dia om axa (o deb is), he segmen a ion was a e wa d manually checked o disca d bad segmen ed dia oms o inish co ec ion; see some examples o he segmen ed dia oms and hei masks in Figu e 4. Appl. Sci. 2017,7, 753 8 o 22 (a) (b) Figu e 4. Segmen ed dia oms and hei bina y mask: (a)Epi hemia so ex; (b)Gomphonema minu um. 4. Dia om Handc a ed Fea u e Desc ip o s An e o mus be made in ansla ing he knowledge o he dia omis s o dis inguish he di e en dia om species and desc ibe he mos ele an ea u es in e ms o compu e ision and pa e n classi ica ion. The goal is o mimic human pe cep ion and he abili y o ecognize a 3D objec om a 2D image, in his case dia oms. E en i i is some imes unclea wha ea u es a e used by he expe o dis inguish among e y simila dia om species, we p oposed and desc ibed dia om ea u es in e ms o au oma ic pa e n classi ica ion. Along he nex subsec ions, he handc a ed ea u es a e p esen ed in g oups acco ding o hei o mula ion wi h a b ie explana ion. The di e en g oups o desc ip o s a e indica ed in Table 2. A o al o 1460 desc ip o s is compu ed, and all o hem a e calcula ed uniquely in hose pixels belonging o he segmen ed bina y masks. Table 2. Lis o handc a ed ea u e desc ip o s di ided in o ca ego ies. LBP, Local Bina y Pa e n. CATEGORY HANDCRAFTED FEATURE TOTAL Mo phological A ea, eccen ici y (3 eccen ici ies) 7 ea u es Pe ime e , shape, ullness S a is ical 1s o de (his og am) 13 ea u es 2nd o de (co-occu ence ma ix) 19 ea u es dis ance =1, 3, 5 pixels di ec ion =0◦, 45◦, 90◦, 135◦ Tex u e space LBP S a .241 ea u es Momen s Hu 7 momen s Space- equency Log Gabo 4 scales (6 o ien a ions) 241 ×4 = 964 ea u es 4.1. Mo phological Desc ip o s Mo phological ea u es ela ed o us ule’s con ou and a ea a e compu ed om he bina y masks. 4.1.1. A ea This desc ip o is calcula ed as he sum o pixels in he bina y mask (B∈(0, 1)) o size MxN: A ea = N ∑ n=1 M ∑ m=1 B(m,n)(1) 4.1.2. Eccen ici y These desc ip o s e lec elonga ion in ela ion wi h he bina y mask’s cen e o mass, also called he cen oid and de ined as: (mc,nc) =   1 A ea ∑ (m,n)∈A ea m·B(m,n),1 A ea ∑ (m,n)∈A ea n·B(m,n) (2) Appl. Sci. 2017,7, 753 9 o 22 The i s Eccen ici y 1 is de ined as a quo ien o he maximum and minimum dis ance be ween he cen oid and bina y mask’s bo de (i.e., us ule’s con ou ), also called ou e and inne ci cum e ence adius. Eccen ici y1=Ou e − adius Inne − adius (3) Simila ly, Eccen ici y 2 is calcula ed as he quo ien o he semi-axes o he bes i ing ellipse o he mask, and Eccen ici y 3 is he a io o he ine ia momen s o he wo semi-axes o he bes i ing ellipse. The momen s a e de ined in Sec ion 4.4. 4.1.3. Pe ime e This desc ip o is he numbe o pixels ha belong o he dia om’s ou line (i.e., a pixel belongs o he pe ime e i i is nonze o and is connec ed wi h a leas one pixel equal o ze o). Pe ime e = N ∑ n=1 M ∑ m=1 P(n,m)(4) whe e: P(m,n) = 1i ∃B(m±1, n±1) = 1 and P(m,n) = 0 o he wise. 4.1.4. Shape This desc ip o is a measu e o he elonga ion o an objec . I is gi en by: Shape =4·π·A ea Pe ime e 2(5) 4.1.5. Fullness This desc ip o is he a io o he mask a ea o he bounding box a ea. 4.2. S a is ical Desc ip o s 4.2.1. Fi s O de S a is ical: His og am These desc ip o s, lis ed in Table 3, calcula e common s a is ics in he image his og am h(i) calcula ed on a 255-bin H . This g oup o desc ip o s is sensible o a ia ions o g ay pixel le els, bu hey igno e hei local co ela ion. Table 3. Fi s -o de s a is ical desc ip o s. Mean µ=∑H−1 i=0i·h(i) Mode i =a gmax(h(i)) Minimum min(h(i)) Maximum max(h(i)) Va iance σ=∑H−1 n=0(i−µ)2·h(i) Range max(h(i)) −min(h(i)) En opy ∑H−1 i=0h(i)·log(h(i)) 1s Qua ile µq1=∑H i=3dH/4ei·h(i) 2nd Qua ile µq2=∑3dH/4e i=2dH/4ei·h(i) 3 d Qua ile µq3=∑2dH/4e i=dH/4ei·h(i) In e qua ile Range µq3−µq1 Asymme y 1 σ3∑H−1 n=0(i−µ)3·h(i) Ku osis 1 σ4∑H−1 n=0(i−µ)4·h(i) His og am h(i), bins numbe H, loo ope a o d e. Appl. Sci. 2017,7, 753 16 o 22 94 94.5 95 95.5 96 96.5 97 97.5 98 98.5 99 All Log Gabo + S a . LBP + S a . LBP + Log Gabo Mo phological + S a . Classi ica ion (Acc.) [94%-99%] Bagging DT Figu e 8. Accu acy pe o mance combining desc ip o ypes and using 300 samples pe class. 7.3. Expe imen 3: Da ase Dimension Some ques ions a ise a his poin abou he da abase dimensionali y. How does his a ec he pe o mance? Is he classi ie able o lea n mo e wi h a highe numbe o samples? Is he classi ie able o ge he same pe o mance wi h a lowe numbe o da a, ha is wi hou he use o augmen ed da a? Thus, we es ed wi h se e al subse s o 20 dia om ypes and 40 dia om ypes wi h 300 samples pe dia om; a subse wi h he 35 classes ha had a minimum o 100 samples pe dia om; as well as wi h 80 dia om ypes wi h 2000 samples pe dia om e sus he p e ious da ase wi h 300 samples pe dia om. To ex end he da ase , a da a augmen a ion wi h o a ions e e y 2◦was pe o med. Figu e 9shows he esul s when classi ying wi h he bagging DT and 10 c . I ep esen s he minimum and maximum Acc. alue ob ained in he di e en ials. The dec ease in he numbe o classes wi h 300 samples pe class inc eased i s accu acy, hough no signi ican ly. Addi ionally, he inc ease in he numbe o samples lowe ed he e o . The use o 100 samples pe class classi ying 35 classes wi h no da a augmen a ion dec eased accu acy o 96.25% e sus 98.11% wi h 300 samples pe class. A simila dec ease happens when classi ying 20 classes, whe e an accu acy o 97.56% (100 samples/class wi hou da a augmen a ion) is ob ained e sus 98.82% (300 samples/class wi h da a augmen a ion). Resul s a e usually illus a ed wi h con usion ma ices. Each column o a con usion ma ix ep esen s he ins ances in a p edic ed class while each ow ep esen s he ins ances in an ac ual class. Since he e a e many classes in his s udy, hea maps ha e been used o display he % o co ec and inco ec classi ica ions. A hea map displays he con usion ma ix as an image whose colo in ensi ies e lec he magni ude o i s alues. In his case, g een alues indica e he % o co ec classi ica ions and pink- ed alues indica e he % o ins ances inco ec ly classi ied. These pe cen ages, as well as he ue posi i e a e (g een column) and alse nega i e a e (pink- ed column) o each class a e shown when no mo e han 20 classes a e classi ied, as well as he ue posi i e a e (g een column) and alse nega i e a e (pink- ed column). The con usion ma ix hea map o he classi ica ion o he 80 dia om classes wi h he bagging ee is shown in Figu e 10. Figu e 11 shows he con usion ma ix hea map o he classi ica ion o he dia om classes wi hou da a augmen a ion. Figu e 11 shows 20 classes chosen andomly o isualiza ion pu pose. The con usion ma ix hea map o se e al ials done wi h he 20 classes is shown in Figu e 12. The main diagonal o Figu e 12 shows how some classes ob ained 100% co ec classi ica ion. Figu e 13 shows hose dia oms species ha p oduce he majo misclassi ica ion e o s, i.e., Ni zschia cos ei,Gomphonema angus a um,Gomphonema mic opumilum,Gomphonema minusculum and Gomphonema minu um. Au oma ion me hods may be a ool o help bo h he expe and he non-expe , bu in di icul cases, i should always be he expe who makes he inal decision. We belie e ha i is he dia omis who mus conside i an au oma ic classi ica ion sys em is accep able, assuming a ce ain Appl. Sci. 2017,7, 753 17 o 22 le el o accu acy o classi y di e en axa. A ejec op ion is also possible, so ha only di icul cases a e p esen ed o he dia omis , while he easy ones a e classi ied au oma ically. No ice ha his me hodology may be applied o o he axa no included in his s udy. Mo eo e , i is cu en ly being applied o ano he 20 di e en axa no shown in his s udy, and simila esul s a e being ob ained. 94.5 95 95.5 96 96.5 97 97.5 98 98.5 99 20 Classes (100 samples) 20 Classes (300 samples) 20 Classes (2000 samples) 35 Classes (100 samples) 35 Classes (300 samples) 80 Classes (300 samples) 80 Classes (2000 samples) Classi ica ion (Acc.) [94.5%-99%] Max. Min. Figu e 9. Accu acy pe o mance wi h di e en numbe s o classes using he bagging DT classi ie . No da a augmen a ion is applied when using 100 samples pe class. Figu e 10. Con usion ma ix classi ying 80 classes wi h he bagging ee classi ie 10- old c oss- alida ion (10 c ). Appl. Sci. 2017,7, 753 18 o 22 Figu e 11. Con usion ma ix classi ying 20 classes ou o he 35 classes wi h 100 samples pe class wi h no da a augmen a ion and using he bagging ee classi ie 10 c . Figu e 12. Con . Appl. Sci. 2017,7, 753 19 o 22 Figu e 12. Con usion ma ix map classi ying 20 classes wi h he bagging ee classi ie 10 c . (a) (b) (c) (d) (e) Figu e 13. Dia oms species ha p oduce he majo misclassi ica ion e o s: ( a )Gomphonema angus a um; ( b )Gomphonema mic opumilum; ( c )Gomphonema minusculum; ( d )Gomphonema minu um; and ( e )Ni zschia cos ei. Appl. Sci. 2017,7, 753 20 o 22 8. Conclusions The main con ibu ions o his wo k a e: (a) big da abase collec ion: a big da abase o dia oms composed o 80 dia om ypes wi h an a e age o 100 dis inc dia om al es pe ype manually c opped has been collec ed; his da abase is being ex ended o 100 ypes; and (b) he s udy o he main handc a ed ea u es o classi y dia oms, p oposing an e icien me hod o classi ica ion. Thus, he s a e o he a in mo phological, s a is ical and ex u e desc ip o s has been exhaus i ely es ed o classi ying 80 dia om ypes. Some o hem, like ex u e based on LBP and log Gabo desc ip o s, had no been e alua ed be o e in his ield. The combina ion o hese ea u es p o ided an imp o emen up o 98.11% accu acy compa ed o p e ious ela ed wo ks. We concluded ha he combina ion o di e en desc ip o s ca ego ies, such as mo phological and s a is ical desc ip o s, oge he wi h space- equency ep esen a ions, speci ically log Gabo and ex u e by means o LBP, p o ided he bes classi ica ion accu acy a es. Along his esea ch, we also come ac oss se e al o he challenging issues, such as he classi ica ion o he same dia om ype du ing i s li e cycle and om di e en iews. A dia om ype may p esen di e en appea ances acco ding o i s iew wi h espec o he z- iew, and consequen ly, i s mo phological and also s a is ical desc ip o s can d as ically a y. This causes he mo phological desc ip o s o no always analyze he mos disc imina ing ea u es. This could be sol ed by ha ing mo e da a ep esen ing all possible cases o al e na i ely by using a hie a chical ee classi ica ion. Thus, i could be ecommended o simpli y he classi ica ion using mo phological e inemen in wo classes: cen ic dia oms and penna e dia oms. A e wa ds, depending on he axon, he sea ch would ocus on some o he de ails, like ex u e, s iae densi y (numbe o s ia pe mic on), lineolae densi y (pixels pe mic on), e c. I should be no ed ha some dia om axa a e almos he same, e en o expe s, which p o es he di icul y o his ask. P ecise segmen a ion is a c i ical poin o he whole classi ica ion p ocess. This is a limi a ion o handc a ed ea u e app oaches. E ec i ely, segmen a ion should be done accu a ely, al hough ou pu pose in he cu en s udy ocuses on compa ing desc ip o ’s disc iminan capaci y. The e o e, we do no pu sue pe ec bina y masks, bu sui able enough o be equally sha ed by all desc ip o s. In o de o handle he abo e-men ioned di icul ies, he au ho s sugges o explo e new classi ica ion echniques based on deep lea ning (Con olu ional Neu al Ne wo ks (CNN)) able o lea n u he om la ge da ase s and wi hou segmen a ion. In his s udy, i has been p o en ha he accu acy in adi ional me hods does no always imp o e wi h augmen ed da a (i is also coun e p oduc i e due o o ien a ion-in a ian ea u es and possible o e i ing). An accu acy up o 97.56% was ob ained classi ying 20 classes wi h no da a augmen a ion (100 samples/class) e sus 98.82% wi h da a augmen a ion (300 samples/class). Acknowledgmen s: The au ho s acknowledge inancial suppo o he Spanish Go e nmen unde he Aquali as- e os p ojec (Re . CTM2014-51907-C2-2-R-MINECO), h p://aquali as- e os.es/en/. Au ho Con ibu ions: Glo ia Bueno designed and pe o med he expe imen s. She w o e he pape , and she is he co esponding au ho . Osca Deniz designed and concei ed o he expe imen s. He also e ised he manusc ip and supe ised he esea ch. Anibal Ped aza p epa ed he o iginal and augmen ed he da abase. Jesús Ruiz-San aqui a ia pe o med pa o he da a acquisi ion and some g aphics o he igu es. Jesús Salido p epa ed he da abase o be a ailable o dissemina ion. Gab iel C is óbal supe ised he esea ch. Ma ía Bo ego-Ramos pe o med pa o he da a acquisi ion. Saúl Blanco pe o med pa o he da a acquisi ion, and he made he anno a ions o he en i e o iginal da abase. Con lic s o In e es : The e is no con lic o in e es . Re e ences 1. Ec o , L.; Rime , F. Using bioindica o s o assess i e s in Eu ope: An o e iew. In Modelling Communi y S uc u e In eshwa e Ecosys ems; Lek, S., Sca di, M., Ve donscho , P.F.M., Descy, J.-P., Pa k, Y.-S., Eds.; Sp inge : Be lin, Ge many, 2005; Chap e 1, pp. 7–19. 2. Wua, N.; Dong, X.; Liu, Y.; Wang, C.; Baa up-Pede sen, A.; Riis, T. Using i e mic oalgae as indica o s o eshwa e biomoni o ing: Re iew o published esea ch and u u e di ec ions. Ecol. Indic. 2017 ,81, 124–131. Appl. Sci. 2017,7, 753 21 o 22 3. Blanco, S.; Beca es, E.; Cauchie, H.; Ho mann, L.; Ec o , L. Compa ison o bio ic indices o wa e quali y diagnosis in he Due o Basin (Spain). A ch. Hyd obiol. Suppl. La ge Ri e s 2007,17, 267–286. 4. Round, F.E.; C aw o d, R.M.; Mann, D.G. Dia oms: Biology and Mo phology o he Gene a; Camb idge Uni e si y P ess: Camb idge, UK, 1990. 5. Mann, D. The species concep in dia oms. Phycologia 1999,38, 437–495. 6. John, D. Use o Algae o Moni o ing Ri e s III. J. Appl. Phycol. 1999,11, 596–597. 7. Hicks, Y.A.; Ma shall, D.; Rosin, P.; Ma in, R.R.; Mann, D.; D oop, S. A model o dia om shape and ex u e o analysis, syn hesis and iden i ica ion. Mach. Vis. Appl. 2006,17, 297–307. 8. Smol, J.; S oe me , E. The Dia oms: Applica ions o he En i onmen al and Ea h Sciences; Camb idge Uni e si y P ess: Camb idge, UK, 2010. 9. Eu opean S anda d, EN 14407: 2004. Wa e Quali y—Guidance S anda d o he Iden i ica ion, Enume a ion and In e p e a ion o Ben hic Dia om Samples om Running Wa e s; Technical Repo ; Eu opean Commission: B ussels, Belgium, 2004. 10. Wayne, R. Ligh and Video Mic oscopy, 2nd ed.; Else ie : Ams e dam, The Ne he lands, 2014. 11. Desikacha y, T. Elec on mic oscope s udies on dia oms. J. Mic osc. 1956,76, 9–36. 12. Pappas, J.; Kociolek, P.; S oe me , E. Quan i a i e mo phome ic me hods in dia om esea ch. No a Hedwig. Beih. 2014,143, 281–306. 13. Klos e , M.; Kaue , G.; Besz e i, B. SHERPA: An image segmen a ion and ou line ea u e ex ac ion ool o dia oms and o he objec s. BMC Bioin o m. 2014,15, 1. 14. Cai ns, J.; Dickson, K.; P y ogle, P.; Almeida, S.; Case, S.; Fou nie , J.; Fuji, H. De e mining he accu acy o cohe en op ical iden i ica ion o dia oms. J. Am. Wa e Resou . Assoc. 1979,15, 1770–1775. 15. Cul e house, P.; Simpson, R.G.; Ellis, R.; Lindley, J.; Williams, R.; Pa isini, T.; Regue a, B.; B a o, I.; Zoppoli, R.; Ea nshaw, G.; e al. Au oma ic classi ica ion o ield-collec ed dino lagella es by a i icial neu al ne wo k. Ma . Ecol. P og. Se . 1996,139, 281–287. 16. Pech-Pacheco, J.; Al a ez-Bo ego, J. Op ical-digi al sys em applied o he iden i ica ion o i e phy oplank on species. Ma . Biol. 1998,132, 357–365. 17. Pech-Pacheco, J.; C is obal, G.; Al a ez-Bo ego, J.; Cohen, L. Au oma ic sys em o phy oplank onic algae iden i ica ion. Limne ica 2001,20, 143–158. 18. Du Bu , H.; Baye , M. Se ies in Machine Pe cep ion and A i icial In elligence. In Au oma ic Dia om Iden i ica ion; Wo ld Scien i ic Publishing Co.: Singapo e, 2002. 19. Pappas, J.; S oe me , E. Legend e shape desc ip o s and shape g oup de e mina ion o specimens in he Cymbella cis ula species complex. Phycologia 2003,42, 90–97. 20. Du Bu , H.; Baye , M.; D oop, S.; Head, R.; Juggins, S.; Fische , S.; Bunke, H.; Wilkinson, M.; Roe dink, J.; Pech-Pacheco, J.; e al. Dia om iden i ica ion: A double challenge called ADIAC. In P oceedings o he In e na ional Con e ence on Image Analysis and P ocessing, Venice, I aly, 27–29 Sep embe 1999; pp. 734–739. 21. Dimi o ski, I.; Koce , D.; Losko ska, S.; Dze oski, S. Hie a chical classi ica ion o dia om images using ensembles o p edic i e clus e ing ees. Ecol. In o m. 2012,7, 19–29. 22. Kuang, Y. Deep Neu al Ne wo k o Deep Sea Plank on Classi ica ion; Technical Repo ; S an o d Uni e si y: S an o d, CA, USA, 2015. 23. Dai, J.; Yu, Z.; Zheng, H.; Zheng, B.; Wang, N. A Hyb id Con olu ional Neu al Ne wo k o Plank on Classi ica ion. In Lec u e No es in Compu e Science—Compu e Vision, ACCV 2016 Wo kshops; Chen, C.-S., Lu, J., Ma, K.-K., Eds.; Sp inge In e na ional Publishe : Ams e dam, The Ne he lands, 2017; Volume 10118, pp. 102–114. 24. Ped aza, A.; Deniz, O.; Bueno, G.; C is obal, G.; Bo ego-Ramos, M.; Blanco, S. Au oma ed Dia om Classi ica ion (Pa B): A deep lea ning app oach. Appl. Sci. 2017,7, 1–25. 25. Lai, Q.T.; Lee, K.C.; Tang, A.H.; Wong, K.K.; So, H.K.; Tsia, K.K. High- h oughpu ime-s e ch imaging low cy ome y o mul i-class classi ica ion o phy oplank on. Op . Exp ess 2016,24, 28170–28184. 26. Blanco, S.; Béca es, E.; He nández, N.; Ec o , L. E aluación de la calidad del agua en los íos de la cuenca del Due o median e índices dia omológicos. Publ. Téc. CEDEX Ing. Ci . 2008,148, 139–143. 27. Ha alick, R.; Shanmugam, K.; Dins ein, I. Tex u al ea u es o image classi ica ion. IEEE T ans. Sys . Man Cybe n. 1973,SMC-3, 610–621. 28. Wang, L.; He, D. Tex u e classi ica ion using ex u e spec um. Pa e n Recogni . 1990,23, 905–910. Appl. Sci. 2017,7, 753 22 o 22 29. Ojala, T.; Pie ikainen, M.; Ha wood, D. Pe o mance e alua ion o ex u e measu es wi h classi ica ion based on Kullback disc imina ion o dis ibu ions. In P oceedings o he 12 h In e na ional Con e ence on Pa e n Recogni ion-Con e ence A: Compu e Vision Image P ocessing (IAPR), Je usalem, Is ael, 9–13 Oc obe 1994; Volume 1, pp. 582–585. 30. Na a, R.; C is obal, G.; Escalan e-Rami ez, B. A comp ehensi e s udy o ex u e analysis based on local bina y pa e ns. P oc. SPIE 2012,8436, 84360E–84372E. 31. Sahu, H.; Bhanodia, P.An Analysis o Tex u e Classi ica ion: Local Bina y Pa e ns. J. Glob. Res. Compu . Sci. 2013,4, 17–20. 32. Ojala, T.; Pie ikäinen, M.; Maenpaa, T. Mul i esolu ion g ay-scale and o a ion in a ian ex u e classi ica ion wi h local bina y pa e ns. IEEE T ans. Pa e n Anal. Mach. In ell. 2002,24, 971–987. 33. Na a, R.; Escalan e-Ramí ez, B.; C is óbal, G. Tex u e Image Re ie al Based on Log-Gabo Fea u es. In P og ess in Pa e n Recogni ion, Image Analysis, Compu e Vision, and Applica ions; Al a ez, L., Mejail, M., Gomez, L., Jacobo, J., Eds.; Lec u e No es in Compu e Science; Sp inge : Be lin/Heidelbe g, Ge many, 2012; Volume 7441, pp. 414–421. 34. Hu, M.K. Visual Pa e n Recogni ion by Momen In a ian s. IRE T ans. In . Theo y 1962,IT-8, 179–187. 35. Fische , S.; S oubek, F.; Pe ine , L.; Redondo, R.; C is óbal, G. Sel In e ible Gabo Wa ele s. In . J. Compu . Vis. 2007,75, 231–246. 36. Guyon, I.; Elissee , A. An in oduc ion o a iable and ea u e selec ion. J. Mach. Lea n. Res. 2003 , 3, 1157–1182. 37. Duda, R.O.; Ha , P.E.; S o k, D.G. Pa e n Classi ica ion, 2nd ed.; Wiley: New Yo k, NY, USA, 2001. 38. Mi a, P.; Mu hy, C.; Pal, S.K. Unsupe ised ea u e selec ion using ea u e simila i y. IEEE T ans. Pa e n Anal. Mach. In ell. 2002,24, 301–312. 39. Pudil, P.; No o ico a, J.; Ki le , J. Floa ing sea ch me hods in ea u e selec ion. Pa e n Recogni . Le . 1994 , 15, 1119–1125. 40. Alpaydin, E. In oduc ion o Machine Lea ning, 2nd ed.; The MIT P ess: Camb idge, MA, USA, 2010. 41. B eiman, L. Bagging p edic o s. Mach. Lea n. 1996,24, 123–140. c 2017 by he au ho s. Licensee MDPI, Basel, Swi ze land. This a icle is an open access a icle dis ibu ed unde he e ms and condi ions o he C ea i e Commons A ibu ion (CC BY) license (h p://c ea i ecommons.o g/licenses/by/4.0/).