1.

Introduction

Aerial photography technology has undergone sophisticated development in recent years. The use of aerial photography to classify forest types, perform land cover change studies, and monitor disaster areas for the forestry industry subsequently becomes a crucial source of forest inventory-related information.^1–6 The forest land cover situation in Taiwan is remarkably complex, and mixed forest types are markedly prevalent. The artificial stereoscopic interpretation method is generally adopted in the use of aerial photographs to determine forest types in Taiwan; however, this method creates interpretation results that are often inconsistent. Image classification is another method that can be employed to determine forest types. This method generally examines differences in image pixel values in which the results are used for subsequent analyses. When high-resolution images are used for analyses, they provide more detailed spatial information. However, conventional pixel value-based classification methods create excessively large data volumes during calculations and ignore the implication of adjacent pixels, which results in errors.^7,8When performing image classifications, the use of spatial correlation can effectively elevate classification result accuracy.^7–10 In high-resolution images, because terrain features may consist of several pixels, the characteristics of objects forming the terrain features can be identified more easily through manual interpretation. Therefore, many studies performed high-resolution image classifications using object-based image analysis (OBIA).^11–13 OBIA primarily uses image segmentation techniques to divide highly homogenous adjacent pixels of an image into spatially correlated objects.¹¹ Next, the unequally sized image objects are separated into different categories. The image segmentation techniques are thus viewed as a type of classification technique that integrates spatial and spectral information.^7,8,11–13 In addition, shapes of image blocks obtained from the image segmentation process can be used to identify terrain features, and the interpretation process is similar to a manual-like interpretation process.⁷ All image blocks feature their own spectra and shapes.¹⁴ Similarly, knowledge-based systems can be used to interpret and classify the terrain features of image blocks. By adopting hierarchical logic and adding appropriate spatial characteristic-related information, classification result accuracy can be improved^1,12,13,15 and combined with some statistical models, such as classification and regression tree (CART)^16–18 and random forest.¹³

This study investigated the feasibility of using high-resolution digital images in forest land cover classification. The feasibility test procedure is as follows: first, we selected ADS-40 digital airborne multispectral photographs as the study materials. Next, we combined OBIA techniques with a knowledge-based classification method. Finally, we performed forest land cover classification using a stratified classification method.

2.

Research Area and Materials

2.1.

Overview of the Research Area

The research site is located at the Da-Chia stream working cycle, Central Taiwan (longitude 120°43′ to 121°26′; latitude 24°03′ to 24°28′). This area is under the jurisdiction of the Dongshih Forest District Office (of the Forestry Bureau) and faces the Liwuhsi Business Area to the East, the Luanta Business Area to the southwest, the Puli Business Area to the south, and the Nanchuang, Taipingshan, Nan’ao, and Ta’anhsi Right Bank Business Areas to the north (Fig. 1). The forest land cover for this site primarily includes coniferous tree, broadleaf tree, bamboo forest, fruit tree, farmland, grassland, landslide area, wetland, roads, and bare grounds.

Fig. 1

Image and map of the research area.

2.3.

Ground Truth Data and Supporting Images

Ground truth data used by the Dajia River Business Area (i.e., the sample area) are mainly comprised of land cover type mapping (prepared by the Taiwan Aerial Survey Office) as well as national land cover mapping, natural vegetation mapping, and the Third Forest Resource Survey mapping (compiled by the National Land Surveying and Mapping Center). Such reference data were then combined with aerial photographs to determine terrain object and type, and the ground verification was carried out between September 16, 2011, and September 17, 2011 (Fig. 2).

Fig. 2

Ground verification photos for each land cover type with coordinates (WGS 84: $x$, $y$).

In addition, the digital elevation model data for the area of Taiwan were collected by the Aerial Survey Office, Forestry Bureau. Analysis and digitization of the 3-D aerial photos resulted in a resolution of $20\mathrm{m}\times 20\mathrm{m}$ and files with a scale of $1/5000$.

3.

Research Method

3.3.

Image Classification

3.3.1.

Classification rules

The classification method first selected terrains that possessed unique features and were highly distinguishable. Next, terrain type was divided into different subcategories in a hierarchical format, which effectively reduced terrain type-related confusions. Subsequently, logical classification rules were set by referring to expert interpretation knowledge, and a logical procedure was established by making logical inferences. This simplified the complex classification process and allowed intricate terrain features to be effectively classified.

The stratified classification method was divided into four levels: level 1 was comprised of shaded and nonshaded areas; level 2 consisted of vegetation-covered and nonvegetation-covered areas; level 3 incorporated high and low biomass areas; and level 4 involved terrain feature-related classifications. The stratified classification diagram is shown in Fig. 3.

Fig. 3

Stratified classification rules.

3.3.2.

Knowledge-based classification

This study employed a knowledge-based classification and constructed a stratified classification model using known classification knowledge. Next, the CART algorithm was used to assess training data of each terrain type from which classification rules were established. Concurrently, supporting mapping data were used to create the classification criteria. Classification methods for levels 1 to 4 were shown in Table 3.

Table 3

Classification methods for different classification level.

Classification level	Classification algorithm
Level 1	Brightness with bimodal histogram splitting method
Level 2	CART algorithm
Level 3	CART algorithm
Level 4	CART algorithm

Level 1 classification

In level 1 classification, the researchers initially performed shadow detection procedures to divide images into the shadow and non-shadow image areas. Shadow detection procedures were conducted using brightness. This study defined brightness as the mean of the R, G, B, and NIR wavebands [Eq. (1) in Table 2].

The difference between the shadow and the nonshadow areas was enhanced using the brightness of the image before dividing the image into the respective shadow and nonshadow areas. Because the brightness values for the shadow and nonshadow areas presented bimodal distribution, the brightness thresholds of these areas can be determined using the bimodal histogram splitting method. The minimum or trough value between two brightness peaks in the histogram can be used as the threshold value to differentiate between the shadow and the nonshadow areas.

Levels 2 to 4 classification

After level 1 classification, the image was divided into the respective shadow and nonshadow areas, and the nonshadow areas were used for the classification of levels 2 to 4. The classification of levels 2 to 4 were performed using CART algorithm. First, by selecting the training data of each terrain type, classification variables or wave band image-related data were extracted. Next, the CART algorithm proposed by Breiman et al.²⁴ was used to establish classification rules. The CART is a technique that can be used to make classifications and predictions in the field of data mining-related research. By analyzing original data, appropriate classification rules are established and used to make predictions for data with an unknown outcome; the process also involves the use of single input variable functions to construct dichotomous decision trees. Using cross validation, this study randomly distributed data to $n$ groups. A total of $n-1$ groups were then used to build a model in which the remaining group was selected to verify the model; the verification process continued until all groups had been chosen as a model verifier. The maximum difference in risk was selected as the postpruning rule to create tree structures with minimum risk values. Next, changes in the misclassification rate were investigated for all tree structure sizes, and models were corrected to identify the optimal model size. The classification threshold values provided by decision trees can be used to set up classification rules. In addition, supporting data or classification conditions may be added according to the literature or expert experience to enhance image classification accuracy.

3.3.3.

Classification accuracy assessment

Image classification accuracy assessment refers to the process in which the image results obtained from the image classification process are compared with current ground truth data to determine the classification accuracy. In this study, 633 sample location data were randomly selected and paired with aerial photographs for verification. The verification indicators that measured the accuracy of the classification results consisted of producer’s accuracy, omission error, user’s accuracy, commission error (CE), overall accuracy (OA), and the kappa index.

4.

Results

4.1.

Image Segmentation

The selection of a favorable segmentation scale facilitates the extraction of terrain features. Images that underwent image segmentation using the segmentation scale 200 generated a higher number of image blocks, whereas those that underwent image segmentation using the segmentation scale 400 produced a lower number of image blocks because its parameter values were more tolerant of image pixel value changes (Table 4).

Table 4

Comparison between the number of segmented blocks.

Segmentation scale	Number of segmented blocks
${\text{Segment}}_{200}$	3,381,578
${\text{Segment}}_{400}$	1,117,689

Figure 3 shows the segmented blocks of forest lands obtained using segmentation scales 200 and 400. According to the figure, the segmentation scale 200 produced more intricate images in which each canopy crown was segmented into multiple blocks. Conversely, the segmentation scale 400 produced more complete images in which each crown was segmented into one block; however, many crown clusters were not segmented. Figure 4 also compared the segmentation results of coniferous and broadleaf forests; the segmentation scale 200 oversegmented the broadleaf forests but favorably segmented the coniferous forests. Such a result may be caused by coniferous trees featuring smaller crowns, which enabled the smaller segmentation scale to display superior results. By contrast, the segmentation scale 400 produced more favorable broadleaf forest segmentation results but failed to segment a number of crowns in coniferous forests.

Fig. 4

Segmentation results for a coniferous forest obtained using segmentation scales (a) 200 and (b) 400; segmentation results for a broadleaf forest obtained using segmentation scales (c) 200 and (d) 400.

Concerning segmentation images of landslide areas and wetlands, the segmentation scale 400 produced more complete object blocks than the segmentation scale 200. However, the former showed slight boundary confusions at the boundary between the landslide areas and wet lands, whereas the latter showed no boundary confusions (Fig. 5).

Fig. 5

Boundary detection results of landslide areas and wetlands produced using segmentation scales (a and b) 200 and (c and d) 400.

Both segmentation scales 200 and 400 showed favorable terrain boundary detection results for roads, buildings, and farmlands in which the latter produced more complete blocks (Fig. 6). Regarding the segmentation results for farmlands, the segmentation scale 200 was able to accurately detect terrain boundary between vegetation and nonvegetation farmlands but oversegmented farmlands that grew the same crops. Conversely, the segmentation scale 400 showed more complete farmland segmentation blocks (Fig. 6). The completeness of the segmentation blocks improved the accuracy of the image classification results and prevented fragmented pixels of images from other image categories from affecting the classification accuracy.⁷

Fig. 6

Image segmentation results for nonforest areas obtained using segmentation scales (a) 200 and (b) 400.

4.3.

Classification Accuracy Assessment

In this study, images were classified using a hierarchical classification method comprised of four classification levels. The OA and the kappa coefficients for each level and classification are shown in Table 5, whereas the classification results for levels 1, 2, and 3 are shown in Fig. 7.

Table 5

Classification results of different level classifications.

Level	Classification	Overall accuracy (%)	Kappa coefficient
1	Shaded and nonshaded areas	98.50	0.9608
2	Vegetation-covered and nonvegetation-covered areas	96.67	0.9333
3	Woody and herbaceous species areas	95.00	0.8854
4	Herbaceous species areas	93.40	0.8639
4	Woody species areas	64.29	0.5229
4	Nonvegetation-covered areas	85.09	0.8125
4	Terrain types	78.20	0.7597

Fig. 7

Classification results for levels 1, 2, and 3.

The classification results for vegetation-covered and nonvegetation-covered areas were combined in Fig. 8. A total of 633 inspection areas were selected for a classification accuracy assessment; the results showed an OA of 78.20% and an overall kappa coefficient of 0.7597. The confusion matrix is shown in Table 6.

Fig. 8

(a) Original image and (b) image classification results.

Table 6

Confusion matrix of the final classification results.

Type	Coniferous tree	Broadleaf tree	Bamboo forest	Fruit tree	Farmland	Grassland	Landslide area	Wetland	Road	Building	Bare ground	Sum	User’s accuracy (%)	Commission error rate (%)
Coniferous tree	31	30	0	0	0	0	0	0	0	0	0	61	50.82	49.18
Broadleaf tree	7	56	2	0	0	0	0	0	0	0	0	65	86.15	13.85
Bamboo forest	14	30	17	3	0	0	0	0	0	0	0	64	26.56	73.44
Fruit tree	0	4	0	58	0	0	0	0	0	0	0	62	93.55	6.45
Farmland	0	0	0	0	59	6	0	0	0	0	0	65	90.77	9.23
Grassland	0	0	0	0	1	40	0	0	0	0	0	41	97.56	2.44
Landslide area	0	0	0	0	0	0	55	3	1	0	7	66	83.33	16.67
Wetland	0	0	0	0	0	0	4	39	1	0	1	45	86.67	13.33
Road	0	0	0	0	0	0	1	1	49	3	6	60	81.67	18.33
Building	0	0	0	0	0	0	0	0	2	41	2	45	91.11	8.89
Bare ground	0	0	0	0	0	0	3	0	4	2	50	59	84.75	15.25
Sum	52	120	19	61	60	46	63	43	57	46	66	633
Producer’s accuracy (%)	59.62	46.67	89.47	95.08	98.33	86.96	87.30	90.70	85.96	89.13	75.76
Omission rate (%)	40.38	53.33	10.53	4.92	1.67	13.04	12.70	9.30	14.04	10.87	24.24

Overall accuracy: 78.20%; overall kappa coefficient: 0.7597.

5.

Discussion

Among vegetation types, the classification results for farmlands, fruit trees, and grasslands were superior. Their omission rates were 1.67%, 4.92%, and 13.04%, respectively, whereas their CE rates were 9.23%, 6.45%, and 2.44%, respectively (Table 6). The use of support mapping and stratified classification allowed the aforementioned terrain types to achieve superior classification accuracy. However, farmlands and grassland with herbaceous species showed poorer classification results when the general classification model was used because they featured similar spectral reflectance values. The use of geographical data as supporting data reduces classification errors, showing that the use of such data is a key to improving knowledge-based classification accuracy.^25,26 Therefore, this study formulated classification rules by combining empirical knowledge and logic with support mapping using the knowledge-based system in which the classification range for farmlands was controlled.

The classification results for coniferous trees, broadleaf trees, and bamboo forests were inferior. Their omission rates were 40.38%, 53.33%, and 10.53%, respectively, whereas their CE rates were 49.18%, 13.85%, and 73.44%, respectively (Table 6). Coniferous trees were misclassified as broadleaf trees and bamboo forests; the misclassification of coniferous trees as bamboo forests was the most substantial. Similarly, broadleaf trees were misclassified as coniferous trees, bamboo forests, and fruit trees; the misclassification of broadleaf trees as bamboo forests was the most severe. However, misclassification of fruit trees was observed less, which was attributed to the inclusion of support mapping data. According to Fig. 9, misclassification of broadleaf trees generally occurred in blocks with bright crowns.

Fig. 9

Misclassification of broadleaf trees because of crowns; (a) original image, (b) object image obtained using segmentation scale 200, (c) object image obtained using segmentation scale 400, and (d) image classification.

In a study on the use of digital aerial photographs to detect pests and diseases in trees of different classifications, Meddens et al.²⁴ showed that misclassifications were prone to occur in pixels containing bright crowns as a result of strong sunlight. Similar results were observed in this study in which both segmentation scales 200 and 400 segmented crowns with relatively brighter pixels into independent object blocks; misclassifications were also more likely to occur in such blocks, confirming the effect of bright crown blocks on classification accuracy. Figure 9 shows how the bright crown blocks of broadleaf trees were more prone to being classified as bamboo forests, which contributed to the high CE rates of bamboo forests and the high omission rates of broadleaf trees.

For nonvegetation-covered areas, wetlands showed the highest classification accuracy, followed by buildings, landslide areas, roads, and bare grounds. Misclassifications were effectively lowered for roads and buildings because of the inclusion of support mapping, which limited the distribution range for each terrain feature and lowered the CEs between the terrain features.

Wetlands were sometimes misclassified as landslide areas and vice versa. However, for other terrain features, misclassifications were less observed. Reference 27 performed image classification using an unsupervised classification method in which they found that the misclassification of water bodies as roads (and vice versa) was frequently observed. Conversely, in this study, the use of the knowledge-based classification in combination with classification rules established using support mapping effectively reduced the misclassification of wetlands as roads (and vice versa). Concerning wetlands that were misinterpreted as landslide areas, we compared the original images of such wetlands and found that image blocks with CEs were generally comprised of terrains with brighter image blocks. For example, CEs were mostly observed in wetlands containing a water body reflecting strong sunlight. Bare grounds displayed the lowest classification accuracy. The instances in which they were repeatedly misclassified as landslide areas or roads may be attributed to the three land types sharing similar spectral characteristics.

The aforementioned terrain classification results showed that by adopting knowledge-based classification rules and limiting the distribution areas for each terrain feature, the complexity of terrain classification can be lowered and classification accuracy can be elevated.

6.

Conclusion

In this study, we classified images using the knowledge-based classification method in combination with object-oriented analysis techniques, hierarchical classification rules, and decision tree classification techniques. Subsequently, classification parameters, such as object texture, shape, and spectral characteristics, were used to form a class-based classification method and used together with geographic information system mapping information. The results were then compared with manually interpreted aerial photographs. The result comparisons showed that the above classification method achieved an OA of 78.20% and a kappa coefficient of 0.7597. Such results indicate that using knowledge-based classification method, formulating classification rules using geographical and system data as support mapping, and including the concept of stratified classification, the likelihood of misclassification can be reduced and classification accuracy can be enhanced. On the basis of the image classification results, some classification errors occurred in images of sunlit crowns because the image values for individual trees varied. Such a variance was caused by the crown structure and the incident angle of the sun. These errors lowered image classification accuracy and warrant further studies.