Supervised and unsupervised classification techniques for satellite imagery in Porto Alegre, Brazil
Supervised and unsupervised classification techniques for satelliteimagery in Porto Alegre, BrazilGeisa Bugs1, Christian Martin Reinhold1, and Kathryn Clagett11 ISEGI, Universidade Nova de LisboaLisboa, Portugalgeisa80@yahoo.com.br; email@example.com; firstname.lastname@example.orgAbstractThe learning used in a satellite image classification may result in vastly different final landcover maps. The goal of this study is to compare supervised and unsupervised classificationmethods and to analyze the resulting differences in regards to the area of land in each classdepending on the method as well as the accuracy of each map. This comparison will result inan informative understanding of the fundamental differences in the two methods as well as asuggestion of which method may be most appropriate given the conditions surrounding thisstudy of the Porto Alegre region in Brazil.Keywords: Supervised Classification, Unsupervised Classification, and Remote Sensing.1 IntroductionHaving accurate and reliable land class maps are important for a number of applicationsincluding planning, environmental management, and land use change analysis, among others.Satellite imagery is vital to creating these maps, since it is extraordinarily time consuming andnot always plausible to acquire such data from fieldwork.This study seeks to generate a land cover map of the region surrounding Porto Alegre in RioGrande do Sul, the southernmost state in Brazil. The images used in this analysis weregenerated by the CBERS2 (China Brazil Earth Resources Satellite) in November of 2006; acolor composite of the study area can be seen in Figure 1 with the city of Porto Alegre near tothe top center of the image.Although these images have 20-meter spatial resolution and the spatial unit of analysis will bea pixel, the final product will be in vector format with a minimum mapping unit of 1 hectare.Owing to the fact that our goal is vector format, our classification will be a hard classificationrather than a fuzzy one. This final mapping unit size was determined by looking at examplesof the classes we wished to extract from the image and determining a size at which the mapwould be more user-friendly (i.e. not too salt and peppered) while still representative of theland cover classes we wanted to differentiate.Instead of simply creating a single map, this study will compare two methods for producingsuch a map: supervised and unsupervised classification. This comparison will be based uponvisual analysis, the statistical difference in land class area in each version, as well as the errormatrix of each result. The ultimate goal of this study is to create two land class maps of thePorto Alegre region and compare and contrast the advantage and disadvantages of eachmethod and each result.
Figure 1: color composite of the study area in southern Brazil (from CBERS2, using green, red, and near infrared bands).2 Study area and datasetThis section will cover not only the physical study area on the ground but also the study area inregards to the images used in processing.2.1 Porto Alegre regionWith a population in the city of about 1.4 million, Porto Alegre is the capital of Rio Grande doSul, a state in the south of Brazil. The city is centered around 30°01’59’’ south latitude and51°13’48’’ west longitude and has an area of approximately 500 km2. We chose to look notjust at the city itself but also its surroundings due to the fact that this provides an interestingexample for remote sensing because of the diversity of landscape this area represents; whilethe city is a thriving metropolis, this image also shows many smaller farm regions and hamletsin the countryside.2.2 CBERS2 imageAs discussed before, this image is from the China Brazil Earth Resources Satellite and, morespecifically, from the CCD (charge-coupled device) sensor. The CCD sensor is multi-spectral,providing images at a 20m by 20m resolution with five bands: band 1 is a blue band from .45-.52mµ; band 2 a green band from .52-.59 mµ; band 3 a red band from .63-.69 mµ; band 4 anear infrared band from .76-.89 mµ; and band 5 a panchromatic band from .51-.73µ. Thetemporal resolution of this sensor is 26 days. Because we had access to a number of highresolution images from November that could be used both in gathering the training sites andaccuracy assessment, we chose an image from November, in this case November of 2006which was taken at 11:13 am Brazilian time; this image’s ID number is
CBERS_2_CCD_1XS_2006_1123_157_134_L2. The study area is also 113 kilometers oneach side and comes pre-projected in the South American Datum 1969 for UTM zone 21S.Lastly, the image come geometrically corrected using cubic convolution, where each pixel’sdigital number is actually an average of its surrounding pixels.3 MethodologyThis section will review the classification nomenclature used, the feature selection process,preprocessing, classification method, and post-processing in turn. Due to the landscape of ourstudy area and the spatial unit of analysis we opted out of any image stratification orsegmentation for this study.3.1 Classification nomenclatureSince it can be useful to use an established classification system in order be able to morewidely apply and compare the results of a land class map, we attempted to use the USGS LandUse/Land Cover Classification System for our classification purposes. However, we alsowanted to be able to represent the specific landscape features of this area well, so weadditionally turned to existing land class maps of various parts of the area represented in theimage. Combining these systems, we were able to come up with eight classes that representthe land cover of the Porto Alegre region well; these are presented in Table 1. To ensure thatour classes account for all possible types of landscape, we used the Land Cover ClassificationSystem software to double-check that the classes we created represented every possible part ofthe landscape. Table 1: Classification nomenclature and description. Class Description 1 Urban Built-up and mixed urban land use (residential, commercial, industrial, etc) 2 Fields Pasture, grassland, <20% sparse vegetation, rangeland 3 Natural Forest Areas with >60% crown cover vegetation 4 Regrowth Forest Areas with 20-60% crown cover vegetation 5 Agriculture Cropland, orchards, groves 6 Water Lakes, rivers, canals, dams (>30% water) 7 Wetland Areas with <30% water 8 Beach Sandy areas3.2 Pre-processingPreprocessing encompasses any geometric or radiometric correction that you need to do beforeprocessing your image. In our case, as mentioned, the image came geometrically corrected byINPE (Instituto Nacional de Pesquisas Espaciais), the organization that provides the satelliteimages, through the cubic convolution method. Likewise, radiometric correction was deemedunnecessary for this image; radiometric correction is only a necessity when doingmultitemporal analysis, quantitative analysis, or when the study area has significanttopography that might lead to the over or underestimation of digital numbers. Since none ofthese conditions were a part of our analysis, as with geometric correction, we did not have toworry about radiometrically correcting the image.
3.3 Feature selectionAs mentioned, we had four bands to work with for the purposes: blue, green, red, and nearinfrared. Our first attempts were to use all four bands in our analysis. However, due totechnical problems, we were ultimately not able to use the blue band in the analysis. Thislimited us to the other three bands and we used all three of them in our analysis. Because ofthe tendency for urban areas and certain types of agriculture to have similar digital numbers,we decided to also employ a texture analysis to one of the bands in an attempt to retrieve betterresults; this type of analysis has been shown to greatly improve the outcome of a land classanalysis (Lu and Weng, 2007). To create this texture, the ‘filter’ tool in ArcGIS was used, at a‘low’ setting on the near infrared band. This image was then included in the analysis in anattempt to specifically better differentiate urban and agricultural land.3.4 Classification methodThe goal of our study is not simply to create a single classification map, but also to comparetwo learning methods—supervised and unsupervised—by undertaking both types of analysison the image.3.4.1 Supervised classificationThe main difference between a supervised and unsupervised classification is the use (or non-use, in the case of the latter) of training sites. For the supervised classification our goal was tocollect at least thirty training sites per class since it is suggested that you need at least 10ntraining sites per class where n is the number of features to be used in the analysis (Jensen,1996). These sites were collected using a combination of sources: personal knowledge of thearea, existing land cover maps for the area, Google Earth, and the aerial photographs that wehad access to. As a general rule we tried to have our training sites be as homogeneous aspossible in the class they represented while also providing a variety of different types of landwithin that class throughout the whole image. Our training sites also were always larger thanone pixel in area. Both the homogeneity and variety of the training sites as well as their sizewere chosen in hope of obtaining as accurate a mean vector for this class as possible. Withoutgetting into the results too much, our first attempts at classifying proved less than ideal so thetraining sites used were revisited, altered, and added to in an effort to yield better results.Supervised classifications using as few as 240 sites and as many as over 550 were run to figureout which training site file was the most appropriate.For the actual classification, the first step was to use the training sites we had found to createsignatures in ArcGIS. Maximum likelihood was used as the classification algorithm and withthis we input the signatures we had created. Since this algorithm works with probabilities,treating every point as having an equal likelihood of belonging to any class, we had theopportunity to include weights to help the classification; however, we opted to let the classifierrun without any additional weights. The maximum likelihood classifier is advantageousbecause it works with probabilities so that high correlation in bands with classes is lessproblematic and because it classifies every pixel in an image, which suits our purpose.3.4.2 Unsupervised classificationIn the unsupervised classification, we began the analysis with running the images through theISOcluster tool. This tool creates a specified number of spectral classes from the images thatare input. In our case, aiming to have 8 final classes, we asked the ISOcluster tool to createthirty-six spectral classes. The result of this tool is a signature file like created with the
training samples that could be input again into the maximum likelihood tool. From this wewent through and reclassed these thirty-six spectral classes to informational classes by visuallyanalyzing the types of land each class represented. Because we were able to assign everyspectral class to an information class, we did not find it necessary to do any cluster bustingwhere we would further break down a spectral class into spectral sub-classes. Doing this wewere able to satisfactorily generate an unsupervised land cover map.3.5 Post-processingBecause of our goal to ultimately create a vector map with a minimum mapping unit of onehectare, we needed to generalize our image. The first step in doing this was to run severalmajority filters on the raster maps for both the supervised and unsupervised classifications,first with four neighbors and later with eight. This was necessary because in order to convertto vector and eliminate the polygons with less than one hectare area we could not have toomany polygons—after our initial majority filter runs, the vector conversion resulted in overone million polygons for the unsupervised classification, something that the computer did nothave enough memory to process.Ultimately, when the majority filter had been run an adequate number of times, the vectorconversion of both the supervised and unsupervised images resulted in about 300,000polygons. In order to generalize the polygons with an area less than one hectare, the eliminatefunction was used in ArcGIS. The eliminate function has two options, either to eliminate tothe adjacent polygon with whom it shares the longest edge or with the adjacent polygon thathas the largest area, we made use of the second option. This tool was used first to eliminatepolygons with an area of less that .1 hectare, then .25, then .5, and finally one hectare. Inbetween each phase the areas were recalculated. The logic behind this incremented approachwas that after the first elimination, perhaps two smaller polygons would be merged resulting inan area larger than 1 hectare, whereas if you immediately eliminate all polygons with an arealess than one hectare, you will skip ahead to a much grosser generalization. Using majorityfilter and polygon elimination in ArcGIS allows for us to create our final product, twoclassification maps for land cover in the Porto Alegre region, one for supervised classificationand another for unsupervised classification, both of which are in vector format and have a finalminimum mapping unit of one hectare.4 ResultsThe results of the two mapping classifications are presented in Figures 2 and 3, both in theirraster form before generalization and in their final vector form. Clearly there are some majordifferences in the two maps. Interestingly, these differences seem less apparent with the rasterimages—generalization seems to heighten the variation between the two. The exactsignificance of the differences between the two maps will be addressed in the discussionsection.The accuracy assessments of the final vector-based maps for the supervised and unsupervisedclassifications are presented in Tables 2 and 3. The accuracy assessment was done bymanually choosing approximately fifty pixels per class and assigning them the class theyrepresent in ‘real life’—these pixels were chosen to be as spread out and as representative ofthe class as possible while still being clearly one class. The overall accuracy for the supervisedclassification was 76% while the overall accuracy for the unsupervised classification wasaround 48%. Working off the idea that a good accuracy would be about 85%, it is clear that
Figure 1: Initial raster (left) and final vector map (right) for supervised classification after generalizing and eliminating polygons less than one hectare.
Figure 2: Initial raster (left) and final vector map (right) for unsupervised classification after generalizing and eliminating polygons less than one hectare.
5 DiscussionThe discussion of our results will be broken down by our overall satisfaction and generalconsiderations for this study before exploring a comparison of the two methods and ultimatelylooking at ways that our work could have been improved.5.1 Overall evaluationAs the accuracy assessments show, both the supervised and unsupervised maps havesignificant problems in representing what is actually happening on the ground. Interestingly,with both methods, wetlands, fields, and agriculture had the lowest accuracy rate for bothproducer’s and user’s accuracy. Problems in defining these categories likely come from theclose spectral relationship between pixels in these vegetated classes. Alternately oradditionally, the confusion between these classes may also stem from the incorrect placementof training sites in the case of the supervised classification or the inappropriate assigning ofspectral classes in the unsupervised classification. Furthermore, these images are taken inNovember, at the end of the rainiest part of the year in this region. This means that vegetationwould be thriving in most every type of land and this may lead to incorrectly assigning trainingsites or problems in differentiating spectral classes of variant types of land class.On the other end of the figurative spectrum, both methods were (not surprisingly) quite good atdistinguishing water and beach from the other classes. These two classes lie at either end ofthe literal spectrum and therefore are less likely to be confused with other classes. In themiddle ground between wetland/fields/agriculture and beach/water are urban, natural forest,and regrowth forest. The supervised classification actually distinguished the two types offorest quite well while the results were not as good for the unsupervised. For the urban class,there was a different situation altogether; in the supervised classification the producer’saccuracy was very good, meaning that those pixels on the ground that were urban were classedas urban while the user’s accuracy was quite low, while in the unsupervised classification, theuser’s accuracy was slightly better, meaning that those pixels classes as urban were in facturban and the producer’s accuracy was lower. Overall, it can be concluded that some of theclasses, those that tend to have a very unique range of digital numbers (beach and water), werewell classed while those whose spectral range is harder to distinguish had poorer results,leading to generally mixed results for both final maps.Since both of our maps were generalized several times, using equally the majority filter andwith the vector-based eliminate function, it is informative to compare the results from beforethe generalization took place to those from after. You might assume, since these methods ofgeneralization reduce the number of distinct pixels for the sake of a larger minimum mappingunit, that the error matrices from before the generalization would be higher than those fromafter and, in fact, this is the case. While the entire accuracy matrices for the pre-generalizationimages are not included here, the overall accuracy for the supervised classification is 77% forbefore generalization (as compared with 76% for after) and almost 57% before forunsupervised (as compared with 48% after). The overall accuracy of the supervised methodremains fairly constant while the overall accuracy of the unsupervised classification changessignificantly.Even though overall supervised classification apparently doesn’t change that much aftergeneralization for the supervised method, Table 4 shows that in fact there is significant changecategory to category from before generalization to after. Interestingly, a class likes wetland,
which has fairly bad overall accuracy, changes very little from before to after generalization.Alternately, urban, which, relatively speaking, has higher accuracies than most of the otherclasses, changes significantly. Likewise, beach and water, the other two categories with highergeneral accuracy, have significant change from before and after generalization. Cleary, thegeneralization of the images significantly alters the accuracy of different categories, betteringsome and worsening others, which may not be apparent from only looking at the change inoverall accuracy, which, for supervised classification, was only 1%. Table 4: Difference in accuracies from before generalization to after. Natural Regrowth Urban Fields Forest Forest Agriculture Water Wetland Beach Difference in - - Supervised 23.5% 5.8% -16.9% -1.9% -2.7% 11.5% 2.2% 31.4%Producers Difference in -Accuracy Unsupervised 7.6% -6.7% -18.5% -19.7% -7.6% 32.8% 0.0% -5.6% Difference in Supervised -25.3% -5.8% 12.0% 1.9% -5.2% 6.1% -6.9% 31.4%Users Difference in - -Accuracy Unsupervised -22.0% -4.0% -20.4% -6.0% -4.0% 10.2% -10.0% 17.0%5.2 Comparison of supervised and unsupervised methodsA simple visual comparison of the results of the two methods for classification (Figures 2 and3) reveal large differences in the result created by each image. These differences arereinforced by Figure 4 that presents the percent area of each land class with each method in thestudy area. For the urban, water, and beach categories, the percent area in each map is fairlysimilar. The greatest differences, which are also visually apparent, are with agriculture andwetland. Out of these, wetland has by far the largest difference, with vastly more wetland inthe supervised classification than in the unsupervised. Interestingly, wetland has fairly lowaccuracy across the board, suggesting that despite the two methods providing such differentresults, in fact neither is particularly correct—possibly the unsupervised classification isunderestimating the amount of wetland while the supervised classification is overestimating it.By far the most important determinant of which methodology is preferable is the accuracyassessment. While neither is ideal, the accuracy for the supervised classification is far superiorto that of the unsupervised. This fact is not surprising considering that with the supervisedclassification training sites are used that have been hand-picked to determine the classes.Overall, this study seems to confirm our suspicions that the supervised has better results thanthe unsupervised. Ultimately, while it is true that the unsupervised method may save timemoney in a project, in this case the results were only approximately 50% reliable; suggestingthat the time saved may not be worth the sacrifice.
40% 35% 30% 25% Super % 20% Unsuper % 15% 10% 5% 0% re r st s n nd st ch te ld ba re re tu a tla wa fie be fo fo ur ul we ric th l ra ow ag tu na gr re class Figure 3: Comparison of areas of different classes of land for supervised and unsupervised classification.5.3 ImprovementsWhile decently satisfactory results were achieved with the supervised classification, it isalways important to consider ways in which these results could be improved upon.First off, there seems to be considerable mixing in the various vegetated classes. While itcould be that more, and more precise, training sites may be beneficial to distinguishing theseclasses, our attempts at using additional training sites did not prove particularly productive.Perhaps the best solution would to include different additional classes by which to differentiatetypes of land. For example, in the supervised classification, hilltops just to the southeast of thecity of Porto Alegre are classes as wetland when in fact they are rocky fields. Due to theirmakeup, the digital numbers for these fields may be lower than that of most of the other fieldsthat were primarily grassy. This means that perhaps the classifiers were not able to groupthese fields properly. Creating two different categories for different types of fields may helpthe results. Likewise, for the unsupervised classification, perhaps starting with more than 36clusters would help differentiate this mixing in vegetation types.Likewise, it could help to generate objects first, and then apply a decision tree model to thoseobjects to generate a class for each object. In fact, our initial hope was to create objects forthis assessment, however, technical difficulties made this impossible. Creating objects wouldlikely reduce our need to generalize in order to obtain our desired minimum mapping unit andtherefore perhaps prevent the loss of accuracy that we discovered as we generalized outoutputs for the final map.Furthermore, while we ran a four-neighbor generalization four times and then an eight-neighbor generalization five times before eliminating the small polygons, it may be worthwhileto explore the accuracy at different phases along this generalization. It could be that while thegeneralizations apparently hurt the overall accuracy after the steps we took, that the accuracymight get better after just a few generalizations since this may get rid of stray outlier pixels.
Lastly, we were unable to make use of the blue band for an unknown reason. This means thatwe were missing an important feature in our analysis. Working more to correct whateverproblems the software was having with this dataset may return better results.6 ConclusionsOur goal in this study was to create land cover maps for the Porto Alegre region of Brazilusing both supervised and unsupervised methods to explore the differences between the twoanalyses and which method more accurately represents the reality of the land cover. Theresults show significant visual differences in the results of the two classification methodswhich very dissimilar accuracies. Ultimately, the supervised method, where training sites arecreated for every desired class, proved the more accurate although both suffer from theinability to properly distinguish some vegetative classes. Notably, in an attempt to provide amap with a minimum unit of 1 hectare, significant generalization was done to each map whichseemingly sacrificed some of the accuracy of either map. While the supervised classificationmethod produced better results, they were still not ideal, and could possibly be improved bythe introduction of objects, less generalization, additional classes, and/or additional bands.Unsupervised methods may save time in gathering training sites, but the results reveal that thisthrift may severely jeopardize the accuracy of the map. Additionally, to produce a truly usefulmap, more time and effort need to be exerted in either case to refine and improve the results.ReferencesLu, D. and Q. Weng, 2007, A survey of image classification methods and techniques forimproving classification performance. International Journal of Remote Sensing, 28(5), pp.823-870.Jensen, J., 1996, Thematic Information Extraction: Image Classification, In IntroductoryDigital Image Processing: A Remote Sensing Perspective, Jensen, J. (Ed.), pp. 197-252, NewJersey: Prentice Hall.