|Abstract (English)|| |
Although the application of aerial stereo photogrammetry in forest inventory has a long tradition, in many countries including Croatia forest inventories are based on labour and time consuming field surveys. Therefore, the objective of this research was to evaluate the applicability of using the digital aerial images of high spatial resolution (ground sampling distance – GSD of 10 cm and 30 cm) for predicting forest stand attributes (basal-area weighted mean dbh – DBHg, Lorey’s mean height – HL, stand density – N, basal area – G, volume – V). This research continues the work of the previous research (Balenović et al. 2013, 2015a), where accuracy of photogrammetrically estimated arithmetic mean diameter and arithmetic mean height of forest stands were evaluated.The research was conducted in the even-aged (sessile oak management class) and the multi-aged stands (European beech and common hornbeam management classes) of a privately owned forest in the management unit Donja Kupčina – Pisarovina, 25 km south of Zagreb, Croatia (Figure 1). Field data were collected during the spring and summer of 2009 as part of the regular forest inventory conducted according to the valid Regulation on Forest Management. A total of 183 circular plots with radii of 8 or 12 m were systematically set in the 14 selected subcompartments. The positions of the sample plot centres were recorded with GPS receiver. Within each plot, the diameter at breast height (dbh) was measured and tree species was determined for all trees with dbh≥10 cm. The height of each tree was calculated by means of the constructed local height curves fitted with Michailloff’s function. The basal area (g) of each tree was calculated from the measured dbhs using standard equation, whereas the merchantable tree volume up to a diameter of 7 cm overbark (v) was calculated from field-measured dbh and estimated h using the Schumacher-Hall function and parameters from Croatian volume tables. The forest stand attributes were calculated by averaging data of all sampled tree within each stand (DBHg, HL) or summing the tree data and dividing it by the total area of all plots for each stand (N, G, V). Stand-level field data were used in the statistical analysis and comparison with photogrammetric data as a ground-truth reference data (Table 1).The colour infrared (CIR) digital aerial images of GSD 30 cm and GSD 10 cm were acquired using a Microsoft UltraCamX digital large-format aerial camera during two aerial surveys in July 2009 (Figure 1, Table 2). The digital terrain data (breaklines, formlines, spot heights and mass points) for the digital terrain model (DTM) generation were collected by stereo-mapping of digital aerial images according to the rules of the Croatian State Geodetic Administration. The whole procedure of image acquisition, aerial triangulation, and collection of 3D data was conducted by Geofoto Ltd. (Zagreb, Croatia).The photogrammetric stereo measurements and the visual interpretation of tree attributes were performed on digital aerial images of 30 cm GSD and 10 cm GSD using PHOTOMOD 5.24 digital photogrammetric system according to procedures described in Balenović et al. 2013, 2015a. The photogrammetric plots were overlaid upon the aerial images based on the spatial coordinates (x, y) of the field plot centres recorded by the GPS receiver. The determination of tree species and crown tops as well as the delineation of crown areas was performed manually for each tree whose top fell inside the plot. The height of each tree was calculated as the difference between the tree top elevations and the corresponding tree bottom elevations determined from the DTM. A raster DTM of 1 m grid size was generated through linear interpolation of a triangular irregular network (TIN) which was previously created from the digital terrain data. The dbh of each tree on the plot was calculated using local regression models with tree height and crown diameter as inputs (Balenović et al. 2012). Crown diameter was calculated from delineated crown area by applying the equation for circle surface area. Further calculations of photogrammetric tree (g, v) and stand variables (DBHg, HL, N, G, V) were identical to previously described calculations of field data.The accuracy of the photogrammetrically estimated stand attributes was evaluated by calculating differences (D), mean differences (MD) and RMSE between photogrammetric- and field-estimates. The relative values of D%, MD%, RMSE% were calculated according to the mean of the field reference values. The D and D% were calculated for each subcompartment, whereas MD, MD%, RMSE and RMSE% were calculated for the whole study area.The results in Table 3 show that photogrammetric measurements of the aerial images of 30 cm GSD (PM30) and 10 cm GSD (PM10) produced reasonable accurate estimates for HL, G, V with relative RMSEs ranging from 3.65% to 5.36%. Similar accuracy was obtained for DBHg estimated by PM10 (RMSE=4.94%), while lower accuracy was obtained for N estimated by PM10 (RMSE=7.71%) and DBHg estimated by PM30 (RMSE=9.460%). The lowest accuracy was obtained for N estimated by PM30 (RMSE=15.90%). Both photogrammetric measurements (PM10 and PM30) estimated HL and G with similar level of accuracy, whereas V was estimated with slightly higher accuracy by PM10 then by PM30. For estimation of DBHg and V, PM10 produced considerably better results, i.e. estimates of approximately twice higher accuracy then PM30. Figure 2 shows relations between D% and field estimates of corresponding attributes for each subcompartment. As can be seen, photogrammetrically estimated HL and V varied between overestimation and underestimation (HL: from -13.6% to 2.8% for PM10, from -12.8% to 3.7% for PM30; V: from -7.0% to 2.2% for PM10, from -10.2% to 8.2% for PM30), but with a slight tendency to underestimate field estimates. Photogrammetrically estimated G also varied between overestimation and underestimation (from -6.2% to 12.9% for PM10, from -5.0% to 10.2% for PM30), but with a slight tendency to overestimate field estimates. DBHg was overestimated for all subcompartments by both photogrammetric measurements (from 1.1 to 9.5% for PM10; from 3.0% to 16.5% for PM30). On the contrary, both photogrammetric measurements underestimated N throughout all subcompartments (from -2.6% to -10.6% for PM10; from -5.1% to -24.4% for PM30). For both DBHg and N, PM30 produced estimates of lower accuracy than PM10. This is a consequence of lesser visibility of details (e.g. crown boundaries) on images of lower spatial resolution (GSD 30 cm) and decreased ability to detect individual trees, especially in the part of stands with greater proportion of younger trees. According to Figure 3, the notable underestimation of N by PM30 was found in the lowest dbh size class (10.0-14.9 cm).The results of this research showed that HL, G and V can be accurately estimated by manual measurements of digital aerial images of high spatial resolution. The use of images of high spatial resolution, along with the use of local dbh models, led to improved accuracy regarding individual tree detection and dbh estimation. Moreover, the errors of N underestimation and DBH overestimation have been mutually abolished, which in the end resulted with reasonably accurate estimates of G and V at stand level. Since the errors of N underestimation and DBH overestimation with PM30 were both proportionally (twice) greater than with PM10, G and V were estimated on both 10 cm and 30 cm GSD images with similar level of accuracy. Although PM10 overall produced the more accurate results than PM30, it should be noted that the price of 10 cm images is several time higher than of 30 cm images. Therefore, a potential user should decide which images to use depending on desired accuracy and available funds.