2.1.

Study Area

The study area of Dayekou forest hydrology experimental area is located in the Qilian Mountain area, with geographic coordinates ranging from N38°29′ to 38°35′ in latitude and from E100°12′ to 100°20′ in longitude within Gansu province, western China (Fig. 1). The elevation varies from 2500 to 3800 m above the sea level. This area has a typical temperate continental mountainous climate. In winter, when the atmospheric circulation is controlled by the Mongolia anticyclone, the climate appears to be cold and dry, with little precipitation. But when the atmospheric circulation is controlled by the continental cyclone in summer, the diurnal difference of temperature is dramatic. There is large difference of precipitation between summer and winter, and annual precipitation mainly takes place in summer. Influenced by the climate and terrain, the prevalent vegetation types are mountainous pastures and forests. The dominant vegetation types include Picea crassifolia, Sabina przewalski, and grassland. Vegetation density varies with terrain, soil, water, and climate factors. In this study, the coniferous tree species of Picea crassifolia was selected as a target.

Fig. 1

The study area of Dayekou forest center in Gansu province of China.

2.2.

LIDAR Data Acquisition

An airborne laser scanning flight was carried out over the study area in June 2008. The airborne laser scanning system used is LiteMapper-5600 developed by the German company IGI. It is the first batch of commercial airborne LIDAR terrain mapping systems using waveform digitization. Its laser scanner is RIEGL LMS-Q560, the sensor specifications of which are shown in Table 1.

Table 1

Specifications of RIEGL LMS-Q560.

The flight was conducted with a nominal height of 700 to 800 m over ground, leading to a pulse density of 0.36 to 1.6 points per square meter. To increase the pulse density, repetitive flights with the same height above ground were carried out over the sample study area. The flights in the sample plot were five times more than those in other areas, so the pulse density was increased to 2 to 7 points per square meter. Additionally, as the laser scanner records a waveform data, multiple returns need to be sampled. The site of the LIDAR data was shown in Fig. 2. The data used in this study was the high-density airborne LIDAR data covering the blue area.

Fig. 2

The site of the LIDAR data. The green shows the low-density LIDAR point data area; the blue denotes the high-density LIDAR point data area.

2.3.

Field Measurement Data

During the field survey, we selected a large sample plot of $100\times 100\mathrm{m}$ in an area with slope less than 20 deg. In order to ensure the accurate position of our plots, we located them with DGPS stations within the open space of standing forest. As shown in Fig. 3, there are six red points representing six DGPS measurement datum points, two of which are in the interior of the large plot, while the other four are in the external area. Besides, we still used the mobile stations to determine the accurate positions of all the trees. As the communications between datum points and mobile stations were accomplished via satellite signal transduction, it was not necessary to “see” all trees from these DGPS points. For the convenience of field measurement of each timber and their statistics, we had subdivided the large plot into 16 small subplots. The 16 small squares in Fig. 3 represented the 16 subplots with a size of $25\mathrm{m}\times 25\mathrm{m}$, which were numbered from 1 to 16.

Fig. 3

Field measurements of individual trees and their spatial distribution.

Based on the subdivision, the large plot was separated into 16 subplots, and there were 1,456 trees in all subplots in total. The structural parameters of each individual tree, including the height, crown width as well as diameter at the breast height (DBH), and height under the branch, were measured using an altimeter rod, a tape measure, a ribbon tape, and so on. The geographic position of each individual tree also was measured using the stations of TOPCON GTS-602 and TOPCON GPT-7002.

3.1.

Generation of CHM

The LIDAR intensity represents the reflectance characteristics of the surface in the near-infrared spectra between wavelengths of 800 and 1550 nm. The ratio of received to transmitted laser energy thus increases with the increase of reflectivity of the target and decreases with the decreasing distance between sensor and target. In a forested area, there are differences between vegetation and ground soil in both reflectivity and height. LIDAR intensity is also an important information source, which can be exploited in forest characterization. In this study, we separate the ground points from the vegetation points using the skewness-change algorithm based on the intensity information of the high-density LIDAR data.²

In the skewness-change algorithm, the changes of statistical characteristics of skewness have been used. Based on the central limit theorem, naturally measured samples will follow a normal distribution. The object points may disturb the normal distribution. The statistical results of LIDAR intensity from several different area samples have been used to prove it. In our study area, the distribution of LIDAR intensity could be considered as a combination of two normal distributions. The skewness and kurtosis of this distribution are two characteristics, which are used in statistical analyses to describe the distribution of LIDAR points based on its intensity. Based on the change curves of skewness and kurtosis, the vegetation points are separated from ground points. The method, which makes use of intensity of laser scanner data, is especially applicable in steep and forested areas.

The ground points separated from LIDAR points using skewness change algorithm still include a few aboveground points, which are indicated as noise. After removing the noise, DTM as well as fractional vegetation cover are generated, respectively, according to whether or not pulses have their first echo on the ground. First, a triangulated irregular network (TIN) is constructed for the ground point based on a Delaunay triangulation of its elevation data. Then a rectangular grid of pixels is extracted from each TIN using linear interpolation with a constant sampling interval of one meter. Finally, the raster DTM of one-square-meter spatial resolution is generated.

Then the DSM is obtained through filtering out the uppermost echoes and then the CHM is acquired from Eq. (1). Figure 4 shows the DTM, DSM, and CHM of the study area with a spatial resolution of 1 m, respectively.

Fig. 4

The DTM (left), DSM (mid), and CHM (right) of the study area.

3.2.

Determination of Variational Window

Before the extraction of tree heights and crown widths from CHM, we should first determine the detection window, including its shape and size. When Kini and Popescu²³ extracted an individual tree’s structural parameters using local maximum filter algorithm, a variational window was used. Its size was determined according to the relationship between tree heights and crown widths, which were influenced by the species (deciduous or coniferous forest). As there is only one kind of tree (Picea crassifolia) in our sample plot, we only need to determine one kind of relationship between tree height and crown width. According to the structural parameters of 1,456 trees, we established a linear relationship between the heights and crown widths, which was shown as Eq. (2) and in Fig. 5. In Eq. (2), $D_{\text{crown}}$ is the crown width, and $H_{\text{tree}}$ is the tree height. The $R$ square of the relationship is 0.6103, which meets the requirement of accuracy for variational window determination. Based on this relationship, we could determine the size of window; however, its shape is still uncertain. In most previous studies, it was set as square. Considering that the crown projection is circular no matter what is the shape of crown itself, we finally select the circular window,²³ which is more appropriate for the extraction in our study.

Fig. 5

Regression relationship between tree heights and crown widths from field measurements.

3.3.

Extraction of Structural Parameters

Finally, based on the variational window, the local maximum filter algorithm was performed to extract individual tree heights and crown widths from the CHM.¹⁶ The flow chart of this algorithm was shown in Fig. 6. The extraction result was partly shown in Fig. 7 in which the green cross marks indicated the tree locations with a base map of CHM. The right one showed the corresponding aerial photo.

Fig. 6

Flow chart of the local maximum filter algorithm for locating trees and measuring height and crown width.¹⁸

Fig. 7

Tree locations of individual trees from local maximum filter (left) and its corresponding aerial photo (right).

4.1.

Discussion

After extracting the heights and crown widths of individual trees, we performed the validation using field-measured values. The validation results are shown in Figs. 8 to 10. Figure 8 shows the validation result of tree heights. From it we could find that the determination coefficient ($R^2$) between extracted heights and measured values was up to 0.8568, which indicated the efficiency of height extraction.

Fig. 8

Comparison of tree heights from Local Maximum (LM) algorithm and field measurements.

Fig. 9

The histogram of deviation of the tree heights between measured and extracted values.

Fig. 10

Comparison of crown widths from LM algorithm and field measurements.

From Fig. 8, we could find that the extracted heights were among 8 to 20 m while the field-measured values were between 4 and 25 m. Comparing the minimum height, we concluded that some of the lower trees cannot be effectively extracted using the method in this paper. It probably resulted because the lower and younger trees had been covered by higher and lusher surrounding trees and thus could not be detected by airborne LIDAR. It also accounted for that the number of extracted trees from CHM was less than that from field measurements. At the same time, by comparing the maximum height, we were aware that a few of the highest treetops had also not been detected. The reason may lie in that LIDAR-laser points didn’t really hit the apex of the crown. Although there were some undetected trees, in all stands, more than 51.5% of all trees had been detected, and the percentage was larger than 70% in most areas. So the airborne LIDAR technology was indeed effective and efficient in the extraction of tree heights. In order to adequately demonstrate this point, we made the histogram of deviation of the tree heights between measured and extracted values, which is shown in Fig. 9. From this figure, we found that the deviation followed normal distribution, which indicated the reliability and efficiency of our extraction method.

Figure 10 shows the extraction precision of tree crown widths. The determination coefficient ($R^2$) for tree crown widths was only 0.3923, which needed to be improved. The points in Fig. 10 also appeared dispersed. It may be due to the crown’s edge being susceptible to the crown of surrounding trees when the texture feature was used to extract the crown width in the local maximum filter algorithm. So new algorithms should be developed using airborne LIDAR data. In the future studies, aerial photos and other high-resolution optical images and even radar data could be combined to improve the extraction accuracy. Now we are attempting to combine LIDAR data with high-resolution aerial image to detect the crown’s edge more effectively.

4.2.

Conclusion

As LIDAR has been a novel remote-sensing means applied to obtain accurate vertical information, it has been quickly developed. In this paper, we first generated the CHM using the skewness-change algorithm based on the intensity information of high-density airborne LIDAR data and then determined the shape and size of the variational window using relationship from field-measured structural parameters, finally extracted the tree locations as well as tree heights, and crown widths using local maximum filter algorithm. But there was a little deviation between some of the extracted and measured locations, so we only validated the extraction precision for the trees whose extracted location was close to the measured position. The validation was performed based on the field-measured structural parameters.

The experiments in our work provided a useful exploration in the extraction of an individual tree’s structural parameters using airborne LIDAR data. The results showed that the airborne LIDAR data was significantly effective in the extraction of tree heights while it seemed disappointing in extracting the crown widths. On one hand, it verified that the LIDAR data have more advantages in obtaining vertical information than acquiring horizontal information; on the other hand, it also indicated that new algorithms should be developed and other aerial photos or radar data could be introduced to improve the application effect.