IntroductionTop
Tree structure is a key issue for the understanding of many aspects of plant ecology and is also essential for the characterization and monitoring of cone production of Pinus pinea L. (Calama et al., 2008). Recent developments of operational small unmanned aerial vehicles (UAV) eased their extensive use in forestry, as both the spatial and temporal resolution of UAV imagery suit better local-scale investigation than traditional remote sensing tools (Bohlin et al., 2012; Lisein et al., 2013). UAV imagery emerged as a feasible alternative for the monitoring of the three-dimensional (3D) structure of forests (Puliti et al., 2015). Structure-from-Motion (SfM) techniques allow to extract 3D-information of UAV flights, providing 3D point clouds on the basis of feature matches within overlapping images (Fritz et al., 2013). From the 3D point cloud, a digital terrain model (DTM), a digital surface model (DSM), and a canopy height model (CHM) –computed by subtracting DTM from DSM– can be obtained, and individual tree attributes can be measured in the same way that processing ALS (Airborne Laser Scanning) data (Nurminen et al., 2013).
In Portugal, new irrigation and fertilization experiments were established in 2013 within the framework of PINEA project. This research explores the performance of high spatial resolution UAV imagery to estimate tree height and crown dimensions under different treatments. An individual tree crown approach was performed in P. pinea stands, since could be valuable to examine the variation in the height and size of the crowns, required as input for single-tree level modelling of growth and cone production.
Materials and methodsTop
Study area
This study was conducted in the private forest of ‘Esteveira’, close to Alcochete in the center of Portugal (Figure 1a). The trial was established over an area of 16 ha umbrella pine (P. pinea) forest plantation (Figure 1b), using a randomize design with 2 blocks, subjected to 3 treatments: control and two different levels of fertirrigation. Trees were planted in rows of 8 × 10 m in 1992-1993. In 2013 a systematic thinning was conducted producing a final density around 63 stems ha–1 in rows of 10 x 16 m, in order to form an open canopy of scattered trees. The trial site is characterized by fairly flat terrain (slopes from 0 to 6%, elevations from 78 to 92 m.a.s.l) and no understory.
|
|
Field measurements
Individual tree height and crown width were measured to the nearest decimetre with a Vertex III hypsometer (Haglöf, www.haglof.se) in a subsample of 52 trees (Table 1). To measure crown diameters, the Vertex was positioned just below the crown projection, in each main cardinal extreme, to obtain two cross records (North-South and East-West directions).
Table 1. Summarised data from field and image-measured sample plots
Overall description of the method
The workflow is shown in Figure 2. Firstly, the flight planning and monitoring software eMotion 2 V. 2.4.2 was used to determine the main flight parameters (office-planning phase) (Figure 1c). Secondly, five ground control points (GCPs) were marked to georeference the output 3D models in the photogrammetric processing (First phase, Figure 2).
|
|
Thirdly, images were processed to obtain DSM (Figure 1d). DSM point clouds were classified into ground and non-ground points using Postflight Terra 3D (pix4D, ®Ecublens, Switzerland). Non-ground points were removed using an automatic object classification followed by a subsequent manual refined to remove the remainder noise. Fourthly, the natural neighbour interpolation technique implemented in Postflight Terra 3D was employed to generate the DTM using the ground points. Fifthly, a CHM was obtained by subtracting the DTM from the DSM using QGis V. 2.2.0 (Second phase, Figure 2).
Finally, an object-based image analysis (eCognition Developer 8.7 (®Trimble Gmbh, Munich, Germany)) was applied for the individual crown delineation and field and image-based measurements of the individual tree variables were compared (Third phase, Figure 2).
Image Acquisition and Pre-Processing
The airborne campaigns were conducted on March 12, 2015, using an RGB camera and the fixed-wing UAV SenseFly eBee operated by Terradrone Co. The RGB camera was a Canon Powershot S110 with a 4000 × 3000 pixel detector capturing images at ISO200 and 1/2000 seconds with a 5.320 mm focal length and sensor dimension of 7.4 x 5.58 mm. The UAV weighs approx. 700 g with payload, and has a maximum operational flight time of approx. 45 min.
The flight plan covered the entire study area with a lateral and longitudinal overlap of 80 and 75%, respectively. The flight line spacing was 48 m and the average altitude above ground level was 170 m (Figure 1b). Two perpendicular flight lines were flown in order to improve the accuracy. Each picture covered an area of 240 x 180 m.
A set of 190 images were analysed to generate each orthomosaic and DTM/DSM for the trial. We used pix4D software (®Ecublens, Switzerland), that implements SfM technique for image reconstruction. The orthomosaic had high spatial resolution with ground sampling distance (GSD) values of 6.23 cm.
Calculation of tree variables
Tree position, height, and crown width was retrieved from the CHM by using a modified version of the mixed pixel-and region-based algorithm designed by González-Ferreiro et al. (2013). This algorithm is proposed as a sequence of routines programmed in the integrated development environment (IDE) (eCognition Developer 8.7 (®Trimble Gmbh, Munich, Germany)). All logical procedures in the construction of the canopy delineation algorithm were arranged into five groups: CHM smoothing, segmentation, classification of canopy areas, iterative process, and data export. For this study, some parameters of the algorithm have been tuned, since the size and shape of P. pinea crowns widely differ from those crowns of Pinus radiata D. Don for which the algorithm was originally designed. Thus, a CHM resampling to 20 cm resolution and a subsequent smoothing with mean (5x5) and median filters (5x5) was conducted. Further, the initial maximum search domain in the iterative process (see the Figure 3 in González-Ferreiro et al., 2013) was changed from 5 to 25, and only one interaction was applied. Crown delineation (Figure 3a) was then exported as vector polygons in an ESRITM shapefile. Crown widths of the individual polygons were measured in two perpendicular directions (N-S and E-W) using QGis V 2.2.0. Tree tops and height attributes were exported as a point vector shapefile for subsequent analysis. Finally, linear fits and RMSE were computed comparing field and image-based measurements for the 52 sampled trees.
|
|
ResultsTop
Linear fits of the field and image-measured height showed an R2 = 0.81 (rRMSE = 4.56%), while in crown diameter an R2 = 0.95 (rRMSE = 6.14%) was achieved (Figure 3b). Detection rate was 100%. There were no false positives neither negatives in the sample area.
Discussion and conclusionsTop
This study shows that P. pinea crown variables can be accurately modelled from high resolution UAV imagery. Tree heights and crown width estimations presented similar or higher performance that previous researches for palm plantation (Kattenborn et al., 2014) and olive orchards (Zarco-Tejada et al., 2014; Díaz-Varela et al., 2015). It is unclear if error source were field or imagery data since: i) accurate field measurement of the height of P. pinea is difficult, due to umbrella-shaped crown and lack of a clear apical dominance and ii) tree crown estimation from remote sensors could take into account tree irregularities ignored by the operator in field measurement (Moorthy et al., 2011).
It is important to note that this methodology was applied on flat terrain below sparsely distributed trees without the need of supplementary data points to generate the DTM. Results have demonstrated that forest inventories in P. pinea forests could be supplemented and updated with low-cost UAV imagery. In more dense vegetated areas a poorer performance is expected due to the impossibility of aerial photography for penetrating through vegetation and additional data could be necessary. Future research will replicate this experiment using a UAV-LiDAR system in order to compare its performance against the low-cost alternative presented here and also to enable crown delineation from UAV in more dense forests.
Finally, it is important to highlight that an UAV equipped with low-cost consumer grade cameras has the potential to provide large amounts of information for forest variables mapping at stand level (e.g density, canopy cover, stand volume, stand biomass, etc.) and might be used for surveying spatial variations in growth or estimate yield differences The present study also sets the basis for further research focused on the use of UAV multi-temporal imagery for forest growth assessment by means of time series in order to improve methods to monitor the development of P. pinea stands under different treatments.