Object Oriented Classification of Very High Resolution ImageryPrajwal .M., Pruthviraj .U.Department of Applied Mechanicsand Hydraulics National Institute of Technology Karnataka-Surathkal*Communication address- [email protected]
com Abstract- The application of unmanned aerial vehicles(UAV) has become more diverse and commercialized. They are used from monitoringand surveillance, precision farming, search and rescue, also majorly insurveying and mapping. When it comes to surveying and mapping the datacollected by the UAV must be survey grade in for it to be useful ininfrastructure projects. Once the data is accurate it needs to be postprocessed and all the topographical features on the ground must be identified.This process is currently manually done. To automate this process, we have usedObject- Oriented classification technique to classify tree canopies on veryhigh resolution imagery obtained by an UAV. Classification was done byassigning relative weights to spectral, size, shape and texture and the optimumrelative weights were found out.
This was followed by few raster and vectorfilters and operations. The classification results showed an accuracy level of88%. Future to validate the findings the same classification parameters wereused on other similar data sets which showed an accuracy of 77%. Hence usingthis technique, the amount of man hours spent on post processing and digitizingUAV imagery can be significantly reduced by at least 50%. Keywords: UAV,Object-Oriented Classification, Tree canopy I. INTRODUCTIONSurveying forms an integral component of civil engineering projectsas it lends and shapes the planning and designing process from a very nascentstage. Hence, getting accurate and reliable representation of the land with allthe necessary parameters forms the cornerstone of any successfully executedproject. The road network in India is scheduled to undergo massive changes withthe expansion of roads happening at about 13km per day.
The value of roads and bridges infrastructure in India is projected to grow at a Compound Annual Growth Rate (CAGR) of 17.4 per centover financial year 2012-2017. In addition to this the Government aimsto provide additional incentives for timely completion of projects.
The current practices in land, engineering and cadastral surveyinvolve a combination of GPS and Total Station tools. Although this is a timeand tested practice it also comes with its inherent drawbacks, some of whichare listed below accuracy is wholly dependent on human measurements, thequality of work is limited and confined by the current standards of tolerancewhich are quickly becoming outdated and insufficient, time consuming and laborintensive process, completeness and the richness of data is an area open toconstant improvement and refinement. Despite the push from both financial andofficial sources the road construction industry still faces delays due toineffective planning that further has repercussions down the value chain.
Someof which are delay in resource allocation and procurement, clearance delays forland acquisition (raghuram et al), decrease in levels of accuracy and quality ofmapping practices.Remote sensing and GIS techniques are both increasingly valued asuseful tools for providing large- scale basic information on landscapecharacteristics. They are used for habitat and species mapping (Rotenberry),biodiversity determination, land change detection (Ramachandra), monitoring ofconservation areas, and the development of GIS layers. In many cases, remotesensing data can partially replace the often time consuming and expensiveground surveys. Also change detection of the earth’s surface can beinvestigated due to the availability oflong-term data.Remote sensing through satellites offers a cost efficient means forsurvey(Mumby).Satellites have a high spatial resolution of upto 25 centimeters with temporalresolution of 1day.
But they are not precise enough to achieve survey levelaccuracies of less than 10 centimeters which is required by the Infrastructure sector. The alternatives for this is through remote sensing is use ofmanned aerial vehicles, LIDAR and Unmanned aerial vehicles. Manned aerialvehicles can cover large areas of land with short duration of time but the drawbacks are the spatial resolution is low, it is expensive and not aneconomically feasible solution for small areas. LIDAR comes in both terrestrialand aerial platforms (Lefsky). Terrestrial LIDAR has two types namely static and mobile. Though all thesetypes of LIDAR give fairly accurate survey grade data it is expensive.
LIDARdata takes high amount of post processing time. Manned system is required foraerial LIDAR platform as the payload is too heavy for an Unmannedaerial system.Unmanned aerial systems have been used by themilitary for surveillance activities from post-World War era.
Recently in thepast decade they have been used for commercial applications. When used as atool for mapping purposes they negateall the draw backs of Manned systems and LIDAR as mentioned above. They arecost effective, easy to operate and give survey grade accuracies. They areeffective for both long and short areas. In the face of increasing demands,shortening timelines and the need for increased efficiency, there are a numberof case studies to show that Unmanned aerial system based Photogrammetry (Westoby)can achieve results nearly identical to traditional method of Manual Survey,while simultaneously reducing cost and saving time. Hence it can be seen thataccelerating the planning process in any possible way by using technology toreplace time consuming manual labour is an attractive part of the solution.
Post processing and interpreting the data from unmanned aerialsystem is a time consuming process. This involves vectoring each and everyfeature of the existing data set. Classical classification techniques which arepixel based methods do not precisely work as the clients require the data to bein vector format.
Hence object based classification techniques are adopted. Theresults of these techniques vary which respect to spatial resolution of thedata. The parameters used to classify satellite imagery do not hold good toclassify very high resolution imagery obtained from unmanned aerial systems.II.
STUDY AREA AND DATA Figure1: Study AreaThe study area is chosen for this study is a 590 meter road stretch.The center of the road stretch has a latitude and longitude of 12.92550N and 77.54680 respectively.
It lies in Banashankari 3rd stage of Bangaloredistrict, Karnataka State, India. It has at least 100+ trees withbuildings and roads. Sl.No Data Purpose 1 UAV imagery Object oriented classification 2 Google Earth Overview 3 Field data Validation 4 SurveyorsCAD data Validation Table 1: Data and purpose?. METHODOLOGYFigure 2:MethodologyOn reaching the sitebefore the operation starts weneed to assess the onsite conditionsand find out parameterssuch as wind speed, maximum topographic variation, Lighting conditions, Height of obstructions such as trees and buildings. After on site conditions are assessed the flight planningis done.
At this stage based on terrain and the GSD requiredthe flight altitude and overlap areset. Mission planner opensource software is used. We adopteda flight altitudeof 110m above ground level andfrontal overlap of 70% and sides overlap of 70%.. The estimated Ground sampledistance (GSD) was 3.86cmand the numbers ofphotos to be captured are112 images.
Figure 3:Flight PlanningGround control points are necessary to bring insurvey grade accuracies of 2 cm in to themaps created. To enable this we used markers which are surveyed through DGPS instrument. The latitude,longitude and altitude are determined to cm level accuracies. The UAV used for the survey was fixed wing SkylarkOsprey UAV.
It has a wing span of 1.5 meters. It can carry a payload of 3kgs.The camera on board is Sony Power shot SX230hs. It has a maximum endurance of1hour.
The flight duration was 8 minutes.Geo-tagging is done to tag the flight log into theEXIF data in photo of the data set. The geo-tagged images are then uploadedinto photogrammetric software ie, Pix4D. The images align themselves in thesequence of capture using the geo- tagged by taking into to latitude,longitude and altitude. Figure 5: Skylark Osprey UAVFigure 4: DGPSUsing GCP editorthe center of each marker on the ground is manually marked and tagged withrespective coordinates from the DGPS readings. The whole map is then reoptimized.
Post this process the point cloud isgenerated followed by the orthomosaic and the digital surface models. The orthomosaicproduced has a GSD of 4.24cms and is true colour representation i.e.
RGB bands.Imagine objectiveextension in ERDAS Imagine software was used for object oriented classificationof very high resolution UAV obtained imagery. In object oriented classificationthe various training sites are selected and initial training of theclassification algorithm is done through Multi Bayesian network classification (He). Withthe multi bayesian network classifier, there is only one network with differentoutput nodes for each class. It is a directed acylic graph where each node correspondsto random variables of interest. It may be trained empirically or have adistribution assigned to each input manually. The uncertainity of theinterdependence of variables is represented locally with the help ofconditional probability table Pr(xi | ?i).
It is associated with each node xi,where ?i is the parent set of xi. The classifier produces a vector object layerwhere each object has a probability attribute between zero and one representingthe probability the object is part of the feature being detected.Segmentation is away of partitioning raster images into segments based on pixel values and locations. Pixels that are spatiallyconnected and have similar values are grouped in a single segment. Thisoperator performs segmentation on the raster image specified by the Inputvariable parameter.
The result is a thematic image where pixels values represent class IDs ofcontiguous raster objects. The Pixel Probability Layer (Margono )input is used to compute the pixel probability zonalmean of each segment and that zonal mean are used as the value of the segmentsPixel Probability attribute.Pixel to segmentRatio Specifies average number of pixels that each output segment will contain.This is an average and that the variability in the segments sizes will be determined by the other parameters, primarily theSize parameter. We specified it to be 500.
Lambda Schedulesegmentation (Xin )was used. It adopts a bottom up merging algorithm and itconsiders spectral content as well as the segment’s texture, size, and shapefor merging decisions. Relative weights are given to spectral, texture, sizeand shape. Higher values produce more spectrally homogeneous segments and lowervalues produce less spectrally homogeneous surfaces. Spectral 0.75 Texture 0.25 Size 0.
75 Shape 0.75 Spectral definesthe weight to given to the spectral component of the segments. This is measuredas the mean of the DN values of the pixels in the segments.
Texture defines theweight to be given to the texture component of the segments. This is measured as the standard deviation ofthe DN values of the pixels in the segments. Size defines the weight to be given to the size component of thesegments. This is measured as the number of pixels in the segment. Shapedefines the weight to be given to the shape component of the segments. This isa measurement of the boundary complexity of the segment i.
e. compactness andsmoothness.For the segmentedobject files probability and size filters are applied. Raster Probability (GOVEDARICA )Filter removesall raster objects whose zonal probability mean is less than the specifiedMinimum Probability. We have adopted a probability size of 0.
55. Size filteroperator filters out raster objects thatare too small, too large, or both. It allows you to restrict the pool of raster objects to those of an appropriate size. Filtering out objectscan also improve efficiency of the model, since fewer objects are processed inlater stages of the mode.
In this case we have not defined the maximum size.The constraint on the minimum size is 20pixels.Then the remainingsegmented objects are converted into vector format from rasterformat. There are two popular ways of converting raster to vector , they arePolygon Trace and Line Trace. In thisstudy we have used Polygon Trace (Xu) is a raster to vector converter thattraces the outlines of raster objects and converts them to vector objects. Vector operationssuch as smoothing and circularity are used to refine the results.
The biggerthe factor, the smoother the polygon object becomes. We have adopted asmoothness factor of 0.8. Circularitydefines how close the object is to a circle. First a center point is computedby averaging the coordinates of all points in the polygon. The distances from each point on the polygon to the center point iscalculated. Then the standard deviation of the distances is computed. The standard deviation is subtracted from 1.
0. A perfect circle willhave a circularity of 1.0. Some other shapes such a squares, rectangles andregular polygons will also have very high circularity. The minimum value is 0for a circle and increases as non- circularity increases.
The adoptedcircularity factor was 0.8.The result will contain a vector file which willgive the automated tree count. Futurethe same parameters are used withdifferent data sets to produce tree counts and they are validated and compared.
IV. RESULTS AND DISCUSSIONSUsing the parameters defined in the methodology we arrived at a tree count of 59 trees. The actual number of treedenoted from the CAD report was 67 trees. There is a difference of 8 trees andthe accuracy of the tree count was 88%. The error in under counting of tree wasdue to the fact that trees which are closely spaced are detected as one tree asshown in fig 8. Hence this current set of parameters has limitation if trees are less closely space.
Figure 4: OrthomosaicFigure 5: Grayscaledenoting pixel probabilityFigure 6: Result of Lambda schedule segmentationFigure7: Final result of the vectors shape files after cleaning Figure 8: Errors in classification of trees ?. CONCLUSIONFrom the study itis evident that Imagine objective extension can successfully classify very highresolution UAV imagery. The major observation while using the same set ofparameters to classify trees is that the resolution of the data set must be thesame. If resolution is greater or lesser than 5 times the original data set,the parameters defined will not give satisfactory results. This can becorrected by changing the pixel to segment ratio during training stage of theclassifier, but more research has to be done in this regard.This classificationfails when the trees are closely packedtogether similar to the likes of tropical rain forests. This isthe reason for wrongly classified 8 trees in the original data set.
Hence thisclassification can be used only in dry and deciduous forests. The size filterneeds to be adjusted with respect to the average size of the tree canopy in thestudy area. While applying size filter one must bear in mind the resolution ofthe data set as it is defined withrespect to number of pixels.The other majorobservation is shade factor.
Unlike satellite data where shade exists onlyoutside the circumference of the tree canopy, here shade exists outside as wellas within the canopy of the tree. Therefore, if the shaded region is trained asa separate training site it will lead to excessive segmentation within the tree canopy.ACKNOWLEDGEMENTWe would like tothank Skylark Drones for providing the data and the infrastructure forconducting this study.ABOUT THE PROJECTA project wasproposed to build an expressway between Mumbai and Nagpur. The total length ofthe expressway was 280kms. Prior to the start of construction of the expresswaya DPR had to be prepared.A survey of theexisting features both manmade and natural features had to be carriedout.Traditional method that is by the use of total station was not adopted dueto lack of time and manpower.
Instead aerial photogrammetry was carried usingUAV’s for the entire length of 280km and a width of 200meters. The total timetaken was 28 days this was including the constraints of bad weather,unfavorable topography and protests from local farmers.Once all the datawas back to the office, 3D reconstruction of the terrain was carried out.
Identifying and indexing the trees in the corridor was one of the tasks. Inorder to achieve high accuracy it used to be done manually on anorthophotograph. But it is a time consuming task.To automate thiswith the help of skylark drones the study on object oriented classification ofvery high resolution imagery was carried out on a pilot and various test plotsto classify and convert trees into vectors files in one step. After completionof the study the same methodology was adopted in the project along with a quickmanual screening for quality control purpose.