2.2 Basic Image Classification Categories
2.2.2 Object-based Approaches
The seasonal effects on tree species classification in an urban environment were studied in [50]. Pixel-based classifiers were reported to have difficulties in dealing with the spectral variations in tree crowns and point to [66] for further consideration. Due to the high variability, the study was conducted object-based, using Definiens eCognition [67]. The study area covered approximately 49 hectares and two data sets were available. Both data sets were collected by an AISA hyperspectral sensor, the first one in July with a 2 m spatial resolution and the other one in October at a resolution of 1 m per pixel. An additional LIDAR data set was also used. A nearest neighbor classification was used for the seven species based on image objects. The overall accuracy was 57 % and 56 % respectively. The summer image was reported to provide better accuracy. Furthermore, the classification accuracy was reported to achieve almost 90 % for two classes but decreases to around 50 % for six species. In [68], it was stated that semi-automatic or automatic single tree based tree species classification can save considerable amounts of money and time in forest
inventory and forest monitoring tasks. The separability of five common tree species in central Europe using CIR aerial images was studied. In addition, textural mea- sures for the description of the tree crowns were considered. Mean and standard deviation of the image bands were calculated and a ML classifier was used. Pixel- based classification achieves an overall accuracy of 38.6 %, which was stated to be insufficient for practical applications. Object-based classification, using Definiens eCognition for tree crown delineation, yielded a classification accuracy of 62.1 %. It was noted that the delineation was insufficient as a single tree crown was split up into several segments. Furthermore, spruce was inseparable from Douglas fir and larch was inseparable from beech based on their data set. It was also observed that, in the texture filtered images the edges of tree crowns were represented differently to the rest of the crown and therefore these images can help in delineating tree crowns, especially for beech and fir. To support automatic classification, the mean and the contrast texture bands were the most promising and calculating the difference of these bands might increase separability even more. For the manually delineated crowns an overall accuracy of 91 % was observed.
Definiens eCognition was also evaluated for object-based fuzzy analysis of multi resolution data in [69]. The hierarchical network created by eCognition was stated to be capable of helping with the detection of objects at different scales. Only fuzzy concepts exist for land cover and land use in reality, as there was no exact threshold between low and high vegetation for example. Furthermore, it was reported that automatic segmentation results cannot be expected to be fully convincing for human interpreters. Fuzzy classification was explained to be a very powerful soft classifier besides neural networks and probabilistic approaches. It yields an n-dimensional tuple of membership degrees describing the extent of class agreement of the consid- ered object to the n available classes. A reliability map was introduced, which was important for manual post-processing and finalize that the method does not replace manual interactions, but reduces the amount significantly. However, no information on the achieved accuracies was given in the paper.
Object-based classification of remote sensing data for change detection using four land-use classes was studied in [70]. In the experiments, data from an optical air- borne digital camera with a resolution of 0.5m and four spectral bands (blue, green, red and near-infrared) was used. The data was resampled to a resolution of 2m per pixel. Five land-use classes were detected. In addition to the spectral bands, a
texture operator based on a co-occurrence matrix with a 5x5 pixel window and the normalized differential vegetation index (NDVI) were used. Supported by results for the variance, the best discrimination was reported in the blue band whereas in the near-infrared band all land-use classes had similar distributions. In a first step pixel-based classification was performed and the results were used as additional input to the object-based approach, which used existing objects from a geographic information system (GIS) database. All the objects in the test area were used as training data. An overall classification rate was not given as the focus was on the detection of change. From the 8.6 % of the objects that were marked as changes 45 % were reported to be real changes, 31 % were possible changes and 23 % were classified incorrectly. LIDAR data was reported to improve classification accuracy.
The spectral properties of the crowns of old growth trees in conifer stands were analyzed in [71]. The stands contained three distinct species and a supervised ML classifier was used. The tree crowns were delineated manually to avoid confusion re- sulting from errors in automated tree detection. A compact airborne spectrographic imager (CASI) with eight and 10 bands in the visible and near-infrared spectral region was used to record the data with a ground resolution of approximately 0.7 m per pixel. Spectral signatures were calculated for each crown and the sunlit area of each crown and class signatures were calculated from the mean and the covariance of the single valued tree vectors. The main difference between species was the over- all intensity of the signature. A high variability within species and a large overlap between species was reported. The signatures calculated from the sunlit parts of the crowns produced better classification results. However species classification was poor and additional internal subclasses were introduced to reduce confusion. Shadowed trees were reported to have low classification accuracy however, the introduction of separate shaded classes was encouraged to improve the classification rates of the sunlit trees. The highest average accuracy for the sunlit crowns was 79 %. When all trees were included the accuracy decreased to 55 %. It was concluded, that species classification in old growth stands based on spectral properties was difficult.
Texture features obtained from the panchromatic band were used to classify conif- erous and broad-leafed in [72]. Multispectral bands were used to classify six forest types using a nearest neighbor approach. QuickBird images with a resolution of 0.6 m panchromatic and 2.4 m multispectral were used. Texture features were cal- culated according to [73]. The highest average accuracy achieved was 73 % for the
fusion of multispectral and panchromatic data.
Another object-based classification on satellite data using DT statistical analy- sis was given in [74]. High spatial resolution sensors with a resolution below 4m per pixel were more appropriate for forest stand-level parameterization. Therefore, Ikonos-2 satellite data with a resolution of 4x4 m per pixel in the multispectral bands and 1x1 m resolution in the panchromatic band was used. For multiresolution seg- mentation eCognition was used as described in [75]. Three species were separated using a DT approach, where all the available pure stands were used as training data. Only four of the 87 available features were used for the DT. One important band was reported to be the ratio of the NIR band calculated as the NIR-value of the image object divided by the sum of all spectral layer mean values. This feature was stated to have the effect of reducing within-class variation. The overall accuracy was 100 %, however, the authors state that the small sample size and the absence of an independent test set have undoubtedly contributed to that overly optimistic result.
Based on hyperspectral imagery, tree species identification in an urban environ- ment was studied in [76]. Two AISA data sets with a 2 m and a 1 m resolution were acquired, whereof only the first 15 and 20 spectral bands respectively were used. In addition a QuickBird satellite image with a resolution of 0.6 m panchromatic and 2.4 m multispectral and a LIDAR data set with a spatial resolution of 1 m were used. Object segmentation was performed with Definiens eCognition using the shadow- less LIDAR elevation and intensity layers. Eight species were used in the study. A class hierarchy together with nearest neighbor rules and simple membership func- tions were used for classification. The highest overall accuracy of 93 % was achieved using the 20 bands of 1 m resolution hyperspectral data in combination with the LIDAR data. The fusion with the LIDAR data was stated to significantly improve classification accuracy and the authors noted that using shadow-less LIDAR data for segmentation was perhaps the most important factor.
Another study performed with Definiens eCognition and a nearest neighbor ap- proach was [77]. Objects were used to avoid the salt- and pepper effect. Eight tree species were discriminated based on SPOT satellite data and overall accuracies between 65.8 % and 77.7 % were achieved.
A tool for seeking significant features for optimal class separation in object-based classification was described in [22]. Bhattacharyya distance [78] was used as a mea-
sure for the separability of two object classes. The optimal threshold to separate two object classes was calculated from a Gaussian probability mixture model. An aerial image with a ground resolution of 1.25 m was used for the classification, which was performed using Definiens eCognition. Altogether, 74 features were used including spectral, shape and texture features. Six land-cover classes were discriminated with an overall accuracy of 95 %. In a second test case with a 0.6 m ground resolution QuickBird satellite images, an overall accuracy of 92 % was achieved.
Spectral information and the textures of high-resolution images and a one against one (OAO) SVM with a linear kernel as first method and a NDVI based method were used in [79]. The texture measures were calculated from the GLCM. In particular, the following measures were computed: mean, standard deviation, range, angular second moment, contrast, correlation, entropy and inverse difference moment. The training and testing databases contained two species, whereof the first was repre- sented by 18 tree samples and the second by 19 trees. For the delineation of the image objects a method based on a seeded region growing approach was used. In the first step the seed points were found by estimating tree tops in the Gaussian-filtered DSM. In the second step, the tree tops were used as seeds to obtain the tree crown borders by a region growing approach based on geometric criteria of the trees. The feature vectors were calculated in different color spaces and a classification rate of 100% was reported for the texture features computed on the value component of the HSV color space.
Another study of forest species classification at the individual tree level was per- formed in [80]. It focused more on the segmentation of individual tree crowns than on classification. Details of the algorithms were not given, but it was stated that the accuracy of the classification depends on the result of crown segmentation.
A locally adaptive classification strategy to classify five tree species in heteroge- neous low mountain range was described in [52]. Inventory data from state forests was used to find suitable parameters for the ML classifier. This classifier was stated to be a robust and well-established method, which could be included in their already established workflow. It was also stated to be the best of all supervised classifica- tion methods, provided that the requirements were met, which means that the data was actually distributed according to the used model (e.g. normal distribution). In a large-scale classification, an overall accuracy rate of 82 % was reported. Using the locally adaptive algorithm, the accuracy was improved and achieved 87 %. The
required spectral bands were the visual bands, a NIR band and preferably also a SWIR band.