PROCEEDINGS, Kenya Geothermal Conference 2011 Kenyatta International Conference Centre Nairobi, November 21-22, 2011
OBJECT-ORIENTED IMAGE CLASSIFICATION OF INDIVIDUAL TREES USING ERDAS IMAGINE OBJECTIVE: CASE STUDY OF WANJOHI AREA, LAKE NAIVASHA BASIN, KENYA Lucy Chepkosgei Chepkochei Geothermal Development Company P.o Box 17700-20100 Nakuru Email:
[email protected]
ABSTRACT Object-Oriented Image Classification method is a useful and promising method of classifying objects from high resolution satellite images. The method segments the image pixel into objects and utilizes the texture and contexture information of the object rather than only using spectral information relied upon by traditional methods. This paper, using high resolution multispectral satellite imagery from WorldView-2, sought to explore ways to extract accurate trees of varying crown sizes. IMAGINE Objective tools from ERDAS IMAGINE software were used to define individual trees model parameters by employing different feature detection and extraction techniques. These enable geospatial data layers to be created and maintained through the use of remotely sensed imagery. The results show that IMAGINE Objective provides a high accuracy function for tree extraction especially when one is dealing with cluster of individual tree crowns. By using cue parameters like color, tone, orientation, texture, etc. the spectral differences between tree and others features were able to be detected. Based on training samples, the trees were quantitatively extracted by means of probability of Bayesian Network on Single Feature Probability (SFP) function. In conclusion, Object-oriented analysis proves a successful method of identifying and extracting individual trees of varying crown sizes. A high accuracy is achieved compared to other pixel-based classification techniques. Keywords: Object-Oriented Image Classification, IMAGINE Objective, Multi-spectral satellite imagery
INTRODUCTION Object-Oriented classification techniques based on image segmentation are gaining interest as methods for producing output maps directly storable into Geographical Information System (GIS) databases [Geneletti, and Gorte, 2003]. This approach considers not only identification of land cover on a pixel level, but also organization of such pixels into groups (segments) that correspond to real world objects. It involves partitioning image into meaningful objects called segments. The basic processing units of object-oriented image analysis are segments, so-called image objects, and not the single pixels [Benz, 2004]. Object oriented image analysis approach combines spectral information and spatial information. The approach segments the pixel into objects according to the color/tone, texture etc. of the image and classifies by treating each object as a whole. Utilizing characteristics information like shape, size, orientation, shadow etc. of an object in addition to using spectral information; object oriented image analysis becomes a powerful image analysis approach. The basic theory of object oriented approach is the fuzzy theory, in the case of
the overlapping area in the feature space, pixels in the overlapping areas will not be classified only into one information class, which is not correct in the real world, but are given different membership to one (with the value 1) or more than one (with the value between 0 to 1) information class [Yan G. 2003]. IMAGINE Objective in ERDAS 2010 version and versions released later is one of the software solutions to object-oriented classification and feature extraction available in the market. Compared to other tools, it is relatively new and less explored by users. It is less complex than some of the other tools, which limits its possibilities while making it easier to use. IMAGINE Objective tool employs feature models which work on objects produced by image segmentation and various other pixel-based algorithms which, after being vectorised, can be processed by geometric and textural parameters [Lack and Bleisch, 2010].With object-oriented analysis it is possible to get better results from remote sensing information. That information may be immediately integrated in the
Chepkochei
GIS allowing the direct realization of vectorial maps [Barrile and Bilotta, 2008]. STUDY AREA The study area is Wanjohi Area in the Lake Naivasha basin. It covers an area of approximately 366 square kilometres. Wanjohi is at latitude of 0.33 (0° 20' 0 S) and a longitude of 36.52 (36° 31' 0 E) situated at about 200 kilometres south west (about 230°) of the approximate centre of Kenya and about 110 kilometres north (about 340°) of the capital city, Nairobi. According to 2009 census, 100 square kilometre area around Wanjohi has an approximate population of 385000 (385 persons per square kilometre) and an average elevation of about 2240 meters above sea level. From whole Wanjohi area, a study area was chosen. The study area covers about 12 square kilometres of which two subsets A and B (see figure 1 below) both of area of 0.25 square kilometres were chosen for the purpose of this study. Subset A varies from subset B; subset A covers a lowland area which represent rich agricultural area and Subset B covers hilly area representing a hilly cropland terrain. Wanjohi area has several tree variation of tree inter-distance, surrounding elements that make and realistic benchmarking identification and classification.
species with a crown size and it a challenging case for tree
In the study area, trees are planted along the roads, along the fences, in the croplands and homesteads and have a crown diameter ranging from 1 to 25 m. Depending on the location; surfaces surrounding
Figure. 1. Map of Kenya showing the study area
the trees are grassland, cropland or bare ground. As for the distance between trees, the area contains isolated trees, tree pairs and clusters of interlocked trees. For this study we treat individual trees as isolated areas of tree crown with a diameter less than 25 m and cluster individual trees as those areas exceeding this value. DATA AND METHODOLOGY Data The remote sensed data used is from the World View-2 from DIGITALGLOBE taken on 10th December 2010. WorldView-2 is the first commercial high-resolution satellite to provide 8 spectral sensors in the visible to near-infrared range [ DIGITALGLOBE, 2010]. The spectral, spatial and temporal characteristics of World View-2 are as shown in table 1 below.
Spectral, spatial and temporal characteristics Number of Bands Resolution Panchromatic optics onboard Slew time Swath width Collection capacity
Quantities 8 bands 2m 46 cm at nadir
300 km in 9 seconds 16.4 km at nadir Average 550,000 km2/day Average revisit 11 days. Table. 1. Spectral, spatial and temporal characteristics of World View-2
Chepkochei
Figure. 2. The 8 spectral bands and panchromatic band of World View-2 (Source: DIGITALGLOBE) In World View -2, each sensor is narrowly focused on a particular range of the electromagnetic spectrum that is sensitive to some particular feature on the ground, or a property of the atmosphere. Together they are designed to improve the segmentation and classification of land and aquatic features beyond any other space-based remote sensing platform [DIGITALGLOBE, 2010]. The 8 spectral bands of World View-2 offer a unique perspective of the data within each scene. Unlike traditional multispectral imagery, World View - 2 captures the coastal, yellow, red edge and NIR 2 wavelengths as illustrated in Figure 2 above. METHODOLOGY Pixel cue classifiers based on Bayesian network Among the functions available in the software to perform classification are; Normalized Difference Vegetation Index (NDVI), Single Feature Probability (SFP), shadow and texture. In this study, SFP and shadow function was used to extract the shadows for the purpose of shadow association. The NDVI, SFP and texture were tested to find out the most suitable for the extraction of individual trees.
Classification of individual trees using SFP IMAGINE Objective was used to define individual trees model employing different feature detection and extraction methods from the satellite imagery. A „feature model‟ formed the basis for the extraction which consists of seven sequenced „process nodes‟. Constructing a feature model is more linear and intuitive. In addition, the support for supervised training and evidential learning of the classifier itself means that the feature models are more transportable to other images once built. There is a given set of algorithms within these nodes that can be freely arranged in terms of order and settings. Especially for the „operator nodes‟ the chosen order of algorithms is essential. IMAGINE Objective also uses cues in the field of image interpretation; these are attributes used by human beings for manual image interpretation. They are two distinct categories: pixel level cues and object level cues. Pixel level cues (e.g. color/tone, texture, and site/situation) must be dimensionless, i.e. measurable as pixel values. Object level cues (e.g. shape, size, orientation, shadow and background) measure properties of one dimensional line or two dimensional polygons and can also measure relationships of an object‟s distributional patterns within a feature class or feature objects spatial association with other feature. The definition of training areas for individual trees as well as for background pixels is of central importance to the outcome. Training
Chepkochei
areas had to be chosen carefully not include any background pixel. During the training phase, pixels that are representative of the individual trees were submitted to compute pixel cue metrics to train the pixel classifier. These training pixels were identified with training polygons in the imagery and during the automated extraction phase candidate pixels from imagery are submitted to the pixel classifier for query to measure how closely they resemble the training pixels. The output at this stage is the pixel probability layer (figure 3b below) in which each pixel‟s value represents the probability that it is the object of interest i.e. the trees. This layer becomes the starting point for several operators that can perform in the raster domain, convert to the vector domain, and further operate in the vector domain. These operators translate the information in the pixel probability layer into vector objects or vector object attributes that are used in all downstream processing. The final output layer gives tree objects as shown in figure 3h below. The following procedure describes the developed feature extraction model for individual trees stepby-step; Raster Pixel Processor (RPP) For this pixel based classification, the SFP (single feature probability) was chosen which uses a Bayesian-classifier (statistic classification). For a machine learning algorithm to be suitable for this process it must be able to handle both continuous and discrete variables, learn from training samples, and return a query metric for candidate pixels indicating a goodness of match to the training samples; and in this case Bayesian Network was used as shown in figure 3b below. At this stage the following are worth noting; The definition of training areas for individual trees as well as for background pixels is of central importance to the outcome. Training areas had to be chosen carefully not to include any background pixel. During the training phase, pixels that are representative of the individual trees were submitted to compute pixel cue metrics to train the pixel classifier. These training pixels were identified with training polygons in the imagery and during the automated extraction phase candidate pixels from imagery are submitted to the pixel classifier for query to measure how closely they resemble the training pixels. The cue algorithms quantify human visual attributes by computing cue metrics for the pixel level cues are non-dimensional per-pixel calculations. The cues that fall into this category are: color/tone, texture and site/situation. In an image processing context these can be extended to include other image operations performed at the pixel level. This allows for the use of pixel cues that can go beyond the human‟s visual ability by
processing information that is visually undetectable, such as vegetative indices, spectral transforms etc. The output at this stage is the pixel probability layer (figure 3b below) in which each pixel value represents the probability that it is the object of interest i.e. the trees. This layer becomes the starting point for several operators that can perform in the raster domain, convert to the vector domain, and further operate in the vector domain. These operators translate the information in the pixel probability layer into vector objects or vector object attributes that are used in all downstream processing. Raster Object Creators (ROC) In this step the function ‘threshold/clump’ was applied which performed a threshold on a pixel probability layer (figure 3b) keeping only pixels with a probability greater than or equal to the threshold value. It converts the remaining pixels to binary (0, 1), then performs a contiguity operation (clump) on the binary values of 1. It then converts the pixel probability layer into a raster object layer that contains pixels that are grouped as raster objects as shown in figure 3c below. A raster object i.e. collection of pixels that touch each other using a 4 neighbor rule and are discontinuous to the other raster objects, share a common class ID and also have an attribute table. This is the starting probability value for the object. Raster Object Operators (ROO) Probability filter‟ and „size filter‟ allow keeping pixel objects with high probability and a certain amount of pixels only. Size filter filtered out raster objects that are too small or too large thus allowing one to restrict the set of raster objects to those of an appropriate size of individual trees. Filtering out objects improved efficiency of the model, since fewer objects are processed in later stages of the model. The output was a new raster object layer as shown in figure 3d below. Raster to Vector Conversion (RVC) Output from the step 1 to step 2 contained pixels that were grouped as raster objects which had associated probability metrics. With „polygon trace‟ raster objects were automatically vectorised converting objects from the raster domain to the vector domain as shown in figure 3e below. It takes as input the ROO and converts each raster object into a vector object as polygon then produces a vector object layer. The following steps were applied on vectorised objects; Vector Object Operators (VOO) This processed the geometric features of the vector objects and stored the probability value for each feature of each object in the attribute table. In this step the vector objects were „smoothened‟ in shape which accelerates later processing. An example of
Chepkochei
the subset objects are as shown in figure 3f below. The trees vector polygons objects have the probability attribute associated with the raster object.
polygon divided by the average distance and standard deviation of the distances is computed and the standard deviation is subtracted from 1.0. If the result is less than zero, it is set to 0.0
Vector Object Processor (VOP) This performed classification on the tree vector objects from VOO above which involved specifying circularity and area cues. This was used by object classifier to measure shape and size property of the tree objects and uses the cues to assign a probability to each object in the group of tree vector objects. It is worth noting that; During the training phase of objects using circularity and area cues, the vector objects that are representative of the trees were submitted to compute object cue metrics to train the object classifier. These training objects were identified with training polygons in the image.
Gaussian distribution was used to describe the expected outcomes of an object cue metric. This is the frequency distribution of many natural phenomena e.g. tree objects which can be graphed as a bell-shaped curve. It can be expressed parametrically as a minimum, maximum, mean, and standard deviation. Object cue metric outcomes that are near the mean have the higher probability.
One difference between pixel training and object training is that with object training the data will usually be sparse, i.e. not enough examples to learn from. For this reason object training was specified with distribution parameters such as minimum, maximum, mean and standard deviation. The automated extraction phase candidate objects from vector object layer were submitted to the object classifier for query to measure how closely they are similar to the training objects (or fit the parametric distributions given). Of course, objects processed during the extraction phase must also pass through the object cue metrics calculation. The query results for an object were recorded as an object attribute to be used in downstream processes. The cue algorithms quantify human visual characteristic by computing cue metrics and the object level cues are computed on the tree polygons. The object cues used are shape (circularity) and size (area). These cue metrics were then used to populate the attributes of the objects on which they are computed. The object cue metrics result used are as shown in figure 3g and are as follows;
Figure. 3a. Original subset of the image
Figure. 3b. SFP using Bayesian network
Figure. 3c. Threshold and clump applied
Area It computed the area of each polygon shape in the input vector file. The output tree vector object has an area attribute containing the result of this metric. Figure. 3d. size and probability filter applied Circularity It measured how close the tree crown object is to a circle. Even though the results of circularity range from 0 to 1, it cannot be considered a probabilistic metric. The minimum value is 0 for a circle and increases as non-circularity increases. A perfect circle would have a circularity of 1.0. The result was computed as follows: Centre point was computed by averaging the coordinates of all points in a polygon with distance from each point on the
Figure. 3e. Polygon trace to convert to vector
Chepkochei
RESULTS
Figure. 3f. Smoothening in shape
Extraction & classification of individual trees In the study area, nearly 90 % of the total trees could be identified and extracted with this method. This means that more than 10 % of the individual trees were not confirmed by the extraction results as shown in figure 4 and 5 below. A manual comparison with the satellite image showed that these were false positive extraction – trees that can be hardly detected visually on the image. There was a spatial pattern concerning the quality of results which could be narrowed to vegetation areas where the feature extraction had more difficulties detecting individual trees.
Figure. 3g: Area and circularity cue metric
Figure 3h: final individual tree objects Figure 3a to 3h: Step-by-step object oriented tree objects extraction using IMAGINE objective tool Shadow – tree association The shadows were extracted so as to associate it with the trees. This was to determine if shadow as object cue metric measures the spatial association a tree object has with individual tree vector object. The association features, shadows was generated by an entirely different feature model which was designed to find shadow objects. The parameter elevation angle 1250 was used expressing the angle from the individual tree vector objects to the association vector object (shadow). A parameter of 1m showing the distance from the tree vector object to the association vector object (shadow) was used for the objects shadow were not adjacent to the primary object (tree). Automatic shadow angle computation by the software was also tested in which the vector object classifier was expected to invoke (without the association of shadow metric) to get the initial vector object probabilities. Then the neighbourhoods of the most highly probable objects are examined to find any shadow vector objects in close proximity. From these objects an estimation of the angle from tree object to shadow object was expected to be computed but it was realized that the software uses 3590 by default as reported in the IMAGINE session log.
Figure. 4. Results of the extracted trees on subset A image
Figure. 5. Results of the extracted trees on subset B image
Chepkochei
Shadow –tree association The shadow association with trees did not improve extraction of trees. The problem comes if one has tree objects of too fine granularity. So if one has tree crown objects for example that touch and form a continuous canopy, only the objects at one side of the overall canopy will have an associated shadow. The other objects won‟t. So if one is using shadow association to increase the likelihood of the object being “off ground” most of the objects will actually fall in probability. So before one try to use the shadow association cue metric one really need to make sure that one have agglomerated the objects together into canopy objects rather than crown (or other smaller) objects. Legend shadow Tree
subset B image
a major challenge. For tree object extraction, a polygon based approach shows the best results. Additionally, the models proof to be robust to be applied to satellite imagery in different example areas (i.e. subsets). In most cases only the training set has to be adjusted. The current software is good in transferring models because of no need of saving absolute file path references together with the models which is a valuable advantage. Tests showed that training set of individual trees as samples allow transferability of feature model to another images and/or other regions with no or only minor adjustments which makes this approach very useful for change detection where a lot of different images of different times will have to be processed. The object-based feature extraction of the individual trees by IMAGINE Objective offered open, modifiable and extensible feature models. This flexibility means that the classifier could be fine-tuned to the specific circumstances of the image being analyzed if initial results are not adequate.
Integration with ERDAS IMAGINE 2010 provided a full suite of vector editing tools for further cleanRed: Layer_6up and editing. This integration provided an end-toGreen: Layer_5end feature extraction process within one integrated Figure. 6. Shadow and trees integrated for shadow association Blue: Layer_2package. This can be very useful in utilizing ancillary layers (data fusion) – e.g. slope, aspect, Best-fit model parameters LIDAR, texture, along with the capability inherent with an object based approach to employ shape It was found out that large cluster of tree crowns metrics, proximity, association, leading to have maximum of 10.2 meters in diameter (about increased feature extraction accuracy and leverages 325m square), small cluster of tree crowns have holdings of all remotely sensed imagery, including diameter of about 6.1 m (115 m sq.) and minimum panchromatic, multispectral, hyperspectral, SAR, tree crown of 2m diameter of individual single LIDAR, etc. trees (16m sq.). It was noted that few isolated individual tree as compared to clustered individual IMAGINE objective could extract tree objects trees. Most tree objects had an average circularity which are attributed based on the measures used to of about 0.76 with the maximum being 0.97 and the identify them, including a final probability of being minimum being 0.47. Shape and size outcomes that tree class, enabling quicker validation of final showed near the mean had the higher probability results and analysis of problem areas. This can be while the outcomes farther from the mean had the valuable in providing the ability to prototype and lower probabilities. test a new feature model on user-selected view windows (even if the training evidence lies outside the area to be processed) for rapid fine-tuning and Accuracy assessment applying the finished feature model to the whole An overall accuracy of about 88.33% was achieved data set( or to different data sets). Finally it was indicating that the probability that a randomly also observed that IMAGINE Objective deploys selected point on the map was correctly mapped. floating licenses as standard, ensuring that The tree object accuracy was at 100% and what is maximum usage of the software is made across an not tree object was found to be 87% meaning if one institution‟s network. selects a random point on the map then the probability of correctly mapped tree object i.e. has the same value as the reference data is 100% and CONCLUSIONS what is not tree is 87%. The probability that a The overall objective was to explore possibilities of randomly selected reference data point is correctly the ERDAS IMAGINE Objective for object mapped was found to be 77% for tree objects and oriented image classification of tree crowns. World 100% for what was not tree. View-2image was used to demonstrate a set of steps that IMAGINE objective tool uses to accomplish the task. We made use of Bayesian DISCUSSION network classifier that uses single feature The developed feature extraction models detected probability function. the tree objects in the image. The results show that the similarity of the grass and crop plantations was RGB
Chepkochei
By using visual interpretation techniques like colour, tone, orientation, texture etc. the tree objects could be detected. Those are cue parameters that enabled us to answer to the first question of spectral differences between tree and others features. The trees were quantitatively extracted by means of probabilistic approach of Bayesian Network. Single feature probability (SFP) outperformed NDVI and texture model. With this step we came to answer the second question of the best function of IMAGINE objective suitable for tree extraction. Of the Bayesian network model parameters, it resulted to a successful tree object map. We calibrated the extraction model using probabilistic approach in IMAGINE objective tool. This was answer to the question raised in the beginning of the model parameters for tree extraction on World View-2 image. At the end of the results of tree extraction, we assessed the accuracy of the Bayesian network model using independent digitized dataset. The overall accuracy was fairly good (88.3%). None was classified as tree while no-tree were found in the assessment, and very few trees could not be depicted by the models (23%) due to their irregularities vis-a-vis to the model parameter. And therefore an in-depth parameterisation for such single spot need to be further developed in IMAGINE objective. Nevertheless IMAGINE objective was judged suitable for the tree extraction especially when one is dealing with cluster of individual tree. It is also important to notice challenges about the object oriented classification, that identification and classification of trees requires in-depth understanding of the factors affecting the interaction between electromagnetic radiation and vegetation in a particular environment, selecting appropriate spatial and spectral resolution as well as suitable processing techniques for extracting spectral information of the trees. RECOMMENDATIONS Further research needs to be done on World View – 2 images for possibilities of classification of other objects like soils, geological objects and biomass studies among other applications. The suitability of IMAGINE Objective in identification and extraction of these objects is also recommended for further research. The NDVI function does not work for individual trees. It is recommended that it is better applied for identification of tree species. Similarly with shadows association. However, it should be further explored how the extracted shadows can be used to mask affected pixels to allow the optimum classification of other neighboring features.
REFERENCES Benz, U.C (2004)., Multi-resolution objectoriented fuzzy analysis of remote sensing data for GIS- ready information. Isprs Journal of Photogrammetry and Remote Sensing, 58(3-4): p. 239-258. Daida, J.M., J.D. Hommes, and e. al., (1996)., Algorithm Discovery Using the Genetic Programming Paradigm: Extracting LowContrast Curvilinear Features from SAR Images of Arctic Ice. DIGITALGLOBE, (2010)., 8 recognized researches Challenge Book, find out what 8Band imagery can do for you. DIGITALGLOBE, (2010)., Feature Classification. DIGITALGLOBE, (2010)., The Benefits of the 8 Spectral Bands of WorldView-2. Ellis-Jones, (2007)., Naivasha-Malewa payments for Environmental / Watershed services. ERDAS (2009)., Automating Feature Extraction with IMAGINE Objective. ERDAS (2009). IMAGINE Objective, User's Guide,. Geneletti, D. and B.G.H. Gorte, (2003). A method for object-oriented land covers classification combining Landsat TM data and aerial photographs. International Journal of Remote Sensing,. 24(6): p. 1273-1286. Grignetti, A., F. Giannetti, and D. Coaloa, (2011)., Identification of death trees in forest areas through object and pixel oriented approaches applied to IKONOS images. Rivista Italiana Di Telerilevamento,. 43(1): p. 3-17. Hassan and David, (1988)., Object-Oriented feature extraction method for Image data Compaction. IEEE Control Systems. Heene, et al.,( 2000)., Computer Vision techniques for Remote Sensing. Department of Telecommunication and Information Processing, University Ghent,. Huurneman, G.C., R. Gens, and L. Broekema, (1996)., Thematic information extraction in a neural network classification of multi sensor data including microwave phase information. In: ISPRS (1996). Vol. XXXI, Part B2. Vienna, pp. 170- 175, Jellema, H.W., (1997)., Application of neural network techniques to digital remote sensing images for image fusion and image classification.
Chepkochei
Enschede: ITC. 75. Lack N. and S.Bleisch, (2010)., Object-based change detection for a cultural-historical survey of the Landscape - From cow trails to walking paths. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXVIII-4/C7.
Ruiz;, L.A., A. Fdez-Sarría;, and J.A. Recio, Texture feature extraction for classification of remote sensing data using wavelet decomposition: a comparative study. UNEP, (2003)., Aerial Survey of the destruction of Aberdare range forest..
Lehrbass, B. and J.F. Wang, (2010)., Techniques for object-based classification of urban tree cover from high-resolution multispectral imagery. Canadian Journal of Remote Sensing 36: p. S287-S297.
V. Barrile and G. Bilotta, (2008)., An application of remote sensing: Object-oriented analysis of Satellite data. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences,. XXXVII.
Lucieer, V.L., (2008,)., Object-oriented classification of sidescan sonar data for mapping benthic marine habitats. International Journal of Remote Sensing, 29(3): p. 905-921.
Wang, K. and S.W. Huang, (2010)., Using fast adaptive neural network classifier for mutual fund performance evaluation. Expert Systems with Applications, 37(8): p. 6007-6011.
Mladinich, C.S., (2010). An Evaluation of ObjectOriented Image Analysis Techniques to Identify Motorized Vehicle Effects in Semi-arid to Arid Ecosystems of the American West. Giscience & Remote Sensing,. 47(1): p. 53-77.
Yan, G., (2003)., Pixel based and object oriented Image analysis for coal fire research. ITC: Enschede. p. 82.