PREDICTION OF REGENERATION FOREST ATTRIBUTES USING UAV PHOTOGRAMMETRIC DATA IN SOUTH-EASTERN NORWAY
José Luis Solís Parejo
MASTER’S THESIS
FORESTRY (CBU)
JOENSUU 2021
specialisation Forest Inventory and planning, 34 p.
Abstract: The rapid development of the new remote sensing platforms has allowed a new approach for the assessment of natural resources and especially an improvement in the acquisition of data for the purpose of forest inventory. New technologies and aerial platforms such as Digital Aerial Photogrammetry (DAP) and unmanned aerial vehicles (UAV) can drastically reduce the time spent in the field and the economic costs when a forest inventory is needed. All stakeholders in the supply chain can benefit from these advantages in a medium and - long term. In addition, it must be noted that this disruptive technology is not exclusive and can be used along with other existing technologies, such as ALS or satellite.
The main aim of this study is to estimate two stand attributes at plot level (stem density - trees ha-1 and mean height - m) of 29 forests stands under regeneration with two tree species classes: - coniferous (Pinus sylvestris L. and Picea abies L. Karst) and - deciduous (Betula sp.). The statistical non-parametric k-MSN method was used with metrics from products derived from high-resolution aerial imagery acquired with a user-grade UAV combined with ground reference data. Spectral and textural data and a precise three-dimensional point cloud were selected as sources of forest structure information. The k-MSN method performs a weighting matrix for choosing the k most similar neighbours from the reference data. It is a canonical correlation analysis that enables the simultaneous modelling of multiple forest variables. A leave-one-out cross validation method was implemented for model assessment.
For all stands the results observed were 2926 trees ha-1 and 1.08 m for RMSE and 50.8 % and 48.5 % respectively for rRMSE.
The basic goal is to demonstrate that the UAV-based data with DAP processing can be used to predict the forest attributes of interest in the current study with enough accuracy to facilitate informed decision for boreal forest managers.
Keywords: UAV, stands under regeneration, digital aerial photogrammetry, remote sensing
their guidance throughout the process of this thesis. Without their help, this project would have not been possible. They always pointed in the right direction to gather the necessary knowledge to complete this work. In addition, I would like to thank the NIBIO research center for supporting my research within their facilities with a welcoming and hardworking environment. Finally, I want to express my gratitude to my friends and colleagues from the University of Eastern Finland for their support. And most importantly, to my family.
1.1 Forest stands under regeneration 6
1.2 From traditional forest inventories to EFIs 7
1.3 UAVs as a remote sensing emergent tool in the forestry sector 8 1.4 Digital aerial photogrammetry 11
1.5 Objectives 13
2. Materials and methods 14
2.1 Materials 14
2.1.1 Study area 14
2.1.2.Field data 15
2.1.3 UAV-based data acquisition 18
2.1.4 Aerial imagery post processing 20
2.1.4.1 Alignment 20
2.1.4.2 Optimization 20
2.1.4.3 Ground Control Points allocation 21
2.1.4.4 Densification 22
2.1.4.5 Ground classification 22
2.2 Methods 23
2.2.1 Prediction of forest regeneration attributes 23
2.2.2 The k-MSN method 24
3. Results and discussion 28
3.1 Prediction of stem density 28
3.2 Prediction of mean height 30
4. Conclusions 33
5. References 34
1. INTRODUCTION
1.1 Forest stands under regeneration
About 40 percent of Norway's land area is covered by forests with an extension of approximately 12 million hectares. Traditionally, forest managers used to carry out extensive logging practices along with long-term investments [1]. As a result, the timber volume stocked in Norwegian forests has increased over a period of 100 years (from 300 to 900 million m3). In addition to other factors, this increase in the growing stock has led to more dense forests and longer rotations resulting in timber with better quality. Stands under regeneration (trees in previous stage of pre-commercial thinnings) have a significant share of norwegian productive forests. Assuming that dominant height does not reach a merchantable size, the stage of these stands suggests that a timely and efficient forest management plan is needed. Despite this, forest managers from Nordic countries usually adopt a subjective approach to final decision-making (e.g. timing and intensity of tending practices in these stands) due to the limited information available.
After a clear-cutting treatment for a given stand, any forest management plan should implement an initial assessment and monitor the regeneration of the recent established tree population. The restoration of harvested areas can be done in two ways: - artificially (tree planting) - and naturally (promoting the natural invasion of new seedlings). The natural regeneration of a forest stand is linked to its ability to produce seeds and originate new trees, achieving their self-perpetuation. It is a fairly common practice in Norwegian forests on medium to low site indices [2]. However, although the natural regeneration of forest stands always exists, it rarely happens in the conditions desired by a forest manager because the seedling establishment is often obstructed by competing vegetation. Availability for environmental resources such as light, water or nutrients between individuals or species sharing the same habitat can often be an obstacle. The main factor affecting tree regeneration is climate. Temperature, solar radiation and humidity constrain the occurrence of species in certain areas. Other abiotic components influencing regeneration are soil and topography. One of the biggest concerns of forest management history has been tree regeneration success since this always has a big impact on the forest stands development [3]. Accurate data collection of tree density and stand canopy mean height is crucial for optimal planning of silvicultural treatments in stands at the regeneration phase [4]. Specific elements must be considered to evaluate the success of new population development because the rapid growth of a young
forest stand could alter the structure of the forest. Indicators such as spacing, tree density, species composition or dominant height often help to assess the evolution of new seedlings.
1.2 The shift from traditional forest inventories to EFIs
Historically, traditional forest inventories have gathered information on basic attributes to undertake forest management according to the stand characteristics. Since the 19th century, the statistical method of sampling was well-established and widely used to estimate the desired physical characteristics of a forest. The vastness of the forest areas brings a great challenge to identify the number of trees in a particular area and usually measuring all the existing stems is not a feasible solution [5]. As a result, the information gathered is usually based on sampling and it varies according to the conditions of the ecosystem type, the goals and priorities of the forest management plan and the available technical, human and economic resources. Diameter at breast height, basal area, site index, dominant or mean height are the main commonly assessed attributes that forest inventories consider in most of the countries.
Unlike traditional forest inventories, EFIs (Enhanced Forest Inventories) offer - plenty of original information - (such as vertical and horizontal structure information) that can be used to make informed forest management decisions [6]. This information becomes more necessary for stands that are in the regeneration establishment stage. It is very important to have clear information during the first years for a more detailed evaluation. Major disturbances (wildfires, diseases, wind damage, silvicultural treatments, etc.) are also powerful reasons to improve traditional forest inventories. One of the main silvicultural lines of research expects to identify new value chains based on versatile and accurate forest data.
With today's new economic scenario and the expansion of new markets, the forest industry must increase the value of the wood harvested and the products and services associated with the exploitation of the forests. Precision forestry can play a key role to improve the competion of wood production because it can supply information on the stem diameter distributions and pre-knowledge of the timber assortments [7]. This new approach to forest inventory can lead to a clear improvement in economic performance and can therefore encourage a mobilisation of private capital in rural areas of southeastern Norway.
Overall, EFIs provide several advantages: lower operating costs, identification of high-value wood products, analysis of variation in forest growth, reduction of negative environmental impacts, improved detection of changes and prediction of accurate forest metrics.
1.3 UAV as a remote sensing emergent tool in the forestry sector
Methods to obtain data through different technologies, including LiDAR (Light Detection and Ranging) and digital aerial photogrammetry (DAP), and platforms such as unmanned aerial vehicles (UAV) have grown rapidly in recent years. LiDAR is usually an airborne remote sensing technology (named airborne laser scanning, ALS) that allows to obtain greater density of terrain three-dimensional measurements when combined with a precise inertial and geopositioning systems. This technology brings a paradigm shift in data capture instruments for forest management planning due to recent growing national availability of ALS data [8].
Several countries in Europe have already completed national ALS mapping campaigns that provide reliable and up-to-date data for administration bodies and private companies.
Sometimes this type of national data collection campaign has two main restrictions: - the sampling density (points m-2) frequently does not meet the specific needs of the forestry sector and - the period for delivering an update for each national coverage (>5 years) is usually much longer than the time needed to use the data for a practical approach (e.g., annual or biannual updates for harvesting areas). Despite this, at the operational level, ALS can provide digital terrain models (DTM) with great accuracy, which has meant a quantitative leap in forest road design and assessment of skid trails inside a logging operation due to reduced costs and increased efficiency.
LiDAR from unmanned aerial vehicles is another recent technology that is increasingly used in the forestry sector as a tool to obtain accurate estimates of timber volume, stem taper, dominant height, growing stock and other forest attributes that have been of great interest for different stakeholders (from small and large logging companies to national forest inventory departments) [9][10]. Its implementation is mainly due to its low cost, high manouverity and great versatility to acquire a wide range of spatial and temporal resolution. Within the wide range of UAVs recently used by the scientific community, fixed-wing and multi-rotor lightweight drones have been the most adopted. The vertical take-off and landing capability is the reason for the very rapid worldwide adoption of multi-rotor aircrafts in various industrial sectors. Fixed-wing normally use batteries with a longer life and can be very suitable for tasks with long-range needs such as forestry. But one of the main setbacks is that the airspace regulation commonly specifies that there must be a visual line of sight from the operator to the UAV. The ability to collect high-resolution aerial imagery of small features with great operational adjustability is a great solution that UAV platforms may provide in operations such as reforestation [12]. They have also been increasingly used in precision agriculture by allowing farmers to determine treatment intensities (water or fertilizer) at fine scale according to their specific conditions.
However, digital photogrammetric techniques require high-precision global positioning
systems (GPS) and inertial monitoring units (IMU) and such equipment is often very heavy and costly. This contributes to limiting the functions of the use of large UAVs [14].
In conclusion, recent literature found that a synergy between these technologies can bring benefits for the classification or assessment of forest resources, and therefore the possibilities they can offer are not exclusive.
1.4 Digital aerial photogrammetry
Digital aerial photogrammetry is fundamentally based on the mathematical relationships of projective geometry and the stereoscopic vision. This discipline allows formulating 3D models from 2D images. DAP uses frequently techniques based on image-matching algorithm. Although photogrammetry is a discipline used since the nineteenth century, the appearance of high-capacity processors along with innovations in digital imaging (including the state-of-the-art remote sensing platforms) have led to the emergence of new techniques based on the correlation of images such as Structure from Motion (SfM). SfM sets up 3D models from overlapped aerial images by automatically deriving key points matches from individual images and then optimizing the 3D location with camera orientation using a bundle adjustment algorithm [15]. The good performance of this algorithm is basic for the correct functioning of SfM as it allows for a significant reduction in errors.
Small UAVs can operate at low altitude above the ground providing data with subdecimeter spatial resolution making this type of platforms very suitable for SfM applications [16].
After the process of the images acquisition, there is a need of making a series of orientations (internal and external) to obtain the stereoscopic model. The internal orientation is given by a set of parameters that allow transforming from the image pixels file to the actual physical dimensions of the image. These parameters are mainly derived from the characteristics of the camera, such as the focal length, the physical dimensions of the image, and certain values of lens distortion. Additionally, others parameters are also used such as flight and terrain altitudes. The external orientation places the bundle light beam of each image in its correct position by intersecting the homologous of two images.
With the emergence of digital photogrammetry, collinearity equations are used to perform this external orientation (fig.1).
Fig.1. Through collinearity equations the main target is to obtain the coordinates of the projection center (X0, Y0, Z0) and the coefficients mxy of the rotation matrix, being known the focal length (c), the ground coordinates (X, Y, Z) and their photo coordinates (x,y).The parameters are a combination of the 3D coordinates with the calibration of the camera positions.
The effectiveness of SfM also depends on the resolution of the captured image. Other factors are the quality of the camera (e.g. the lens used), the degree of image overlapping and the relative motion of the camera in relation to the scene.
The capability of structure from motion algorithms used to collect forest variables data has seen significant progress in recent years. This implementation has been proven successful in different scenarios: stands under regeneration and long-rotation managed forests. The results presented in previous studies with different methodologies and approaches are similar to those obtained from other remote sensing techniques such as LiDAR.
1.5 Objectives
The main objective is to demonstrate that UAV photogrammetric data can be used to predict the forest attributes of interest with enough accuracy to facilitate informed decision for boreal forest managers.
The explanatory variables extracted from the products generated during the DAP analysis of UAV-photogrammetric data will be used to test a non-parametric statistical approach (k-MSN method) in order to estimate the following forest inventory variables:
Prediction of stem density (trees ha-1) at plot level
Prediction of mean height (m) at plot level
2. MATERIALS AND METHODS 2.1 Materials
2.1.1. Study Area
The 29 stands of interest used for the current study were located in the county of Hedmark (South-Eastern Norway), within the municipality of Stange (60.63º N, 11.41º E). The elevation of the terrain ranges from 250 to 600 meters above sea level (m.a.s.l.) in the whole considered study area. Climate is cold and temperate and the closest meteorological station is located in Hamar municipality (132 m.a.s.l) 10.3 km away from the study area. The average amount of annual precipitation is 864 mm and the highest daily value in 2018 was 15.8 mm measured during 24th in January. Mean monthly minimum and maximum temperatures during 2018 were -7.2°C in February and 20.8°C in July (Norwegian Meteorological Institute).
Fig. 2. Different scales of maps from the area of study. From left to right; the location within Norway, stands locations at regional scale and one of the field plot cluster.
The study area is classified as Warm-summer humid continental climate (dfb) according to Köppen-Geiger climate classification. Sandy and silty soils dominate the soil texture and based on the Soil Map of Norway the most common types are Podzols and swamps. The varied conditions throughout the county make the quality of the region's timber the best in Norway. Deciduous species (Betula sp.), Norway spruce (Picea abies (L.) Karst) and Scots pine (Pinus sylvestris L.) were the main dominant species within the study site and Norway spruce dominated over Scots pine in 67 % of the stands. The regeneration stands were located in areas with a site productivity that ranges between 2 m3/ha/year to 4 m3/ha/year.
2.1.2 Field data
Local experienced foresters selected the stands. They focused on several factors: tree species composition, heterogeneity and site productivity between June and September 2018.
The size of the 29 selected stands varies from 0.4 to 13 ha with a mean size of 4 ha. Field data were collected based on a systematic clustered sampling design. Each cluster of plots consisted of 5 circular plots of 50 m2 (one in the center and the rest in the four cardinal directions) with a distance of 10 meters between them. The total amount was 580 plots used as ground reference data (fig.4). The evaluation within each plot considered the measurement of trees than higher than 50 centimetres. The variables or characteristics measured in the field included the number of trees per species (Norway Spruce, Scots pine and broadleaved species) and a sampling of the heights of the three species mentioned above. A vertex hypsometer was used to measure the heights for the larger trees in all the plots, and for the small trees, it was only necessary to use a height pole. The number of trees was transformed from plot level to trees per hectare (Table 1). For better understanding of the mean height for coniferous, there was a need to weight the height by stem density (just like Lorey's height is weighted by basal area, see the formula below) of Picea abies (L.) Karst and Pinus sylvestris L.
Where: hm is the mean height value for coniferous weighted by tree density, ni and hi are the observed values of stem number and height for each species on a given plot.
A Topcon GR5 GNSS receiver was used as a mobile equipment to locate and position the center of each plot with sub-decimetric precision using RTK (real-time kinematic) correlation with a GSM network.
Once the center of the plot was fixed, the rest of the cardinal points that delimited it were calculated considering that they were 10 meters away from the plot center.
Table 1. Summary statistics of tree density and mean height for coniferous and deciduous from the field data.
Variable Min 1st Quartile Mean 2nd Quartile Max
N (trees ha-1) 0 2400 5673 8000 21600
Conifer mean height (m) 0.5 0.9 1.6 2.4 13
Deciduous mean height (m) 0.5 0.6 0.9 1.4 7
Fig. 3. Mean height frequency distribution for the plots within the study area.
Fig. 4. Frequency of number of plots distribution according to tree density.
2.1.3 UAV data acquisition
A DJI Phantom 4 PRO was used in order to acquire the imagery dataset in the assessed regeneration stands for this study (fig.5). There are several factors to consider when collecting a good set of images that meet the objectives of the study. The atmospheric conditions are essential, and the acquisition of the images took place with different intensities of light. For logistical reasons the only conditions that restricted UAV flights were strong winds (with continuous and intense gusts higher than 10 m/s) and heavy rain. The sensor used to capture the images was a 1-inch 20 effective megapixels red, green and blue sensor (RGB) with an ISO range from 100 up to 12800, a focal length of 8.8 mm and a mechanical shutter between 1/2000 s. The ISO value was adapted to the light conditions and the shutter speed was fixed to 1/1000 s in order to avoid blurriness.
Due to restrictions of the civil aviation authority of Norway, the flight altitude was established at 110 meters, with a ground sampling distance (GSD) of 3 cm. The flight speed was set at about 5 m/s, being constant during all flights. As a result, the duration of data collection was practically the same in stands with similar extension. Flying in strips with a pre-arranged number of along and across-track overlap is the most frequent approach used for planning imagery acquisitions [7]. The adjustment of the image overlap is an important decision because the quality of the stereo-matching analysis depends on the overlap amount (>80 %) and the size of the data collected has an impact on the computing time to obtain the desired end-product. The flight duration range and data volume capacity of the UAV must also be taken into account. The along-track overlap used in every flight was set to 90 % and the across-track overlap to 80 %. The whole UAV campaign took place in two weeks and the total duration of the flights for all the stands was around 5 hours, counting the setup of the UAV, the flight itself and the packaging of the UAV. Occasionally, the time invested for each flight was increased due to expected difficulties with specific stands. These problems were caused by either a lack of safety on the selected location or poor radio signal to manouver the drone. There was a need for a detailed planning design of the location of several Ground Control Points (GCP). A GCP is a physical site located on the ground with known position (x, y, z). The coordinates are then used to provide to the photogrammetric software accurately position of every object in relation to the real world. The aim is to establish the most accurate correspondence between the images and the precise geographic coordinates. Because of this, during GCPs allocation a sub-decimetric precision system was used to establish the geographical location of each point. They must be clearly visible from the air, and ideally,
accessible in several photos when the digital photogrammetric analysis is performed. This stage is very crucial and it must be taken into account in the flights planning depending on the forest stand size, shape and position. Sometimes it was necessary the use of the same GCPs for stands that were too close between them in order to save time.
Fig. 5. DJ Phantom 4 Pro used for the aerial imagery data acquisition.
Fig.6. Stand under regeneration located in the area of study.
2.1.3 Aerial imagery post-processing
The processing of the images was performed using the photogrammetric software Agisoft Photoscan Professional version 1.4.1 (Agisoft LLC, St. Petersburg, Russia). It is a software that combines a set of images to create a 3D model. Due to the robustness and accuracy of its algorithms is one of the most used in digital photogrammetry. There are three main products that were generated for each stand using the software: An extremely dense point cloud with a nominal density of 321 points/m2, an orthomosaic with a pixel size of 3 cm and a Digital Surface Model (DSM) with a pixel size of 6 cm. DAP analysis includes several steps:
alignment, optimization, GCP allocation, densification and Ground points detection. This last step classifies the dense point cloud into ground and non-ground points so the resulting digital terrain model (DTM) will be used for point cloud normalization.
2.1.3.1 Alignment
At this stage, the DAP software finds matching points between overlapping images, estimates camera position for each photo using the external and internal orientations mentioned in the previous chapter of DAP workflow and builds a sparse point cloud model. The software allows to set the accuracy level during this process; the highest option applies a multiplier of 4 to the original data, while the medium and low options apply a factor reducing the size of the photo of 1/4 and 1/16 respectively. The highest option of the accuracy level was chosen based on the purpose of this study in order to obtain a very dense point cloud and the subsequent extracted metrics.
2.1.3.2 Optimization
This procedure must be executed in order to achieve greater accuracy in the calculation of the external and internal parameters of the camera and to correct possible distortion. This step strongly depends on the quality of the ground control points [17] and it is crucial in order to avoid optical displacement. Because of this, precise ground control points were fixed on the study area using a GNSS RTK receiver.
Fig.7. Point cloud generated with Agisoft PhotoScan from the imagery data.
2.1.3.3 Ground control points allocation
Regarding a coordinate system, the control points are important since the georeferencing depends on them in order to guarantee that final product as DTM will not be deconfigured neither in position nor in height. The location of the points must be planned before the flight mission, the number of control points should be three or more (five or six are recommended according to data and experience from previous studies). It was required to mark the control points in an identifiable way when viewing the images so it is advisable to use natural objects or spray paint (of a different colour than the one on the ground), forming a cross at the control point (fig.8). It should be quite visible because the coordinate at which it was recorded in the field will be assigned to the pixel that represents the control point in the image. Another factor to be considered is the surrounding area. When locating several control points there should be no obstacles interfering with the signal (dense and high vegetation, infrastructures, etc.). In open forests, conditions for GPS position data collection are assumed to be relatively accurate compared to dense forests.
Fig. 8. A red cross marked with spray paint used as ground control point. The mark should be clear and visible (at least bigger than GSD used in the study) in order to perform a precise imagery georeferencing.
2.1.3.4 Densification
The sparse point cloud is valid to obtain the orientation of the images, but it is not enough in order to create an accurate reconstruction. Due to this restriction it is necessary to densify the point cloud, previously defining some parameters such depth filtering. This parameter helps to remove some factors as noise, poorly focused images or some possible outliers values between the points. For this study a mild depth filtering intensity was selected. It did not minimize the time used for computational processes when compared with other available choices. It was decided to avoid a strong filter on the sparse point cloud due to forest stand variations and to ensure the collection of all available information.
2.1.3.5 Ground classification
The classification of points is a necessary process to obtain the digital terrain model (DTM) in order to normalize the point cloud and calculate the variables of interest for this study such as the forest canopy height. The software enables to classify points with great efficiency to detect ground points and low noise within the point cloud. Basic parameters for the ground points classification step were set using a trial and error approach, choosing a maximum angle of 10º, a maximum distance of 1 m and a cell size of 50 m.
2.2 Methods
2.2.1 Prediction of regeneration forest attributes
A set of 94 candidate variables were obtained from the data generated with the UAV platform. During statistical analysis, the best explanatory variables were used as predictors to model the forest attributes of interest in this study. Depending on the origin of these variables they can be categorised into four types: Orthomosaic (spectral and textural variables), DSM (textural variables), normalized point cloud and ground points percentage (Table 2). The spectral information was extracted from the digital numbers of the red, green and blue bands from the pixels in each assessed plot and then the mean and standard deviation of each band was calculated. The textural variables were computed including the mean and standard deviation using the DSM (dsmMEANm, dsmVARm, etc.) and the raster product from the RGB bands (spcMEANm, spcVARm, etc.). In order to limit the time spent in the computation stage the R package glmc was used defining as constraints 45º degree shift (the direction in which the textural variables were calculated) and a window size of 3x3 pixels. The results found in the researching of Giannetti et al. [18] demonstrated that a UAV-based DSM can be effectively used to predict various forest attributes of interest even when high-resolution DTMs are not available, increasing the potential for these types of platforms for the production of Enhanced Forest Inventories. The extracted variables from the normalized point cloud that describe the vertical structure of the forest stands are mainly percentiles of heights (P10, P20, P30,..., P100), fraction of point density variables compared to the total number of points along the vertical layers (d1, d2, d3,…, d9), fraction of intensity points (IP10, IP20, …, IP100), percentage of ground points (prcGrnd, HartgDp) and the number of local maxima points (locMax3, locMax5,..., locMax29).
The modelling of the variables of interest at plot level for the present study was done through the non-parametric method k-MSN and the resulting models were validated with leave-one- out cross validation (LOOCV).
Table 2. Set of explanatory variables from the products generated in the photogrammetric workflow: Orthomosaic (Ortho), Digital Surface Model (DSM), Normalized Point Cloud (PC) and Ground Points percentage (GP).
Variable Category Variable Category Variable Category
"Rm" Ortho "Gm" Ortho "dsmMEANm" DSM
"d10" PC "Bm" Ortho "dsmVARm" DSM
"d20" PC "Rsd" Ortho "dsmHOMOm" DSM
"d30" PC "Gsd" Ortho "dsmCONTRm" DSM
"d40" PC "Bsd" Ortho "dsmDISSm" DSM
"d50" PC "spcMEANm" Ortho "dsmENTRm" DSM
"d60" PC "spcVARm" Ortho "dsmSECMOMm" DSM
"d70" PC "spcHOMOm" Ortho "dsmMEANs" DSM
"d80" PC "spcCONTRm" Ortho "dsmVARs" DSM
"d90" PC "spcDISSm" Ortho "dsmHOMOs" DSM
"d100" PC "spcENTRm" Ortho “dsmCONTRs" DSM
"prcGrnd" GP "spcSECMOMm" Ortho "dsmDISSs" DSM
"HartgDp" GP "spcMEANs" Ortho "dsmENTRs" DSM
"locMax3" PC "spcVARs" Ortho "dsmSECMOMs" DSM
“locMax5” PC "spcHOMOs" Ortho "DzP10" PC
"locMax7" PC "spcCONTRs" Ortho "DzP20" PC
"locMax9" PC "spcDISSs" Ortho "DzP30" PC
"locMx11" PC "spcENTRs" Ortho "DzP40" PC
"locMx13" PC "spcSECMOMs" Ortho "DzP50" PC
"locMx15" PC "DSMmin" DSM "DzP70" PC
“locMx17" PC "DSMmean" DSM "DzP80" PC
"locMx19" PC "DSMmedn" DSM "DzP90" PC
“locMx21 PC “DSMrang" DSM "DzP95" PC
"locMx23" PC "FLOWDIRs" DSM "DzP100" PC
"locMx25 PC "SLOPEm" DSM "IP10" PC
“locMx27” PC "TPIm" DSM "IP20" PC
"TRIm" DSM "IP30" PC "ROUGHm" DSM
"IP40" PC “SLOPEsd" DSM "IP50" PC
"FLOWDIRm" DSM "IP60" PC "TPIsd" DSM
“IP70" PC "TRIsd" DSM "IP80" PC
"ROUGHsd" DSM “IP90" PC "IP95" PC
“IP100" PC
2.2.2 k-MSN statistical method
According to Packalen and Maltamo [19], “The k-MSN (most similar neighbour) method is a statistical non-parametric nearest neighbour approach. It applies canonical correlation analysis to perform a weighting matrix for choosing the k most similar neighbours from the reference data”. In forestry application, k-MSN is useful for predicting values from observations in forest inventory units to others where these measured values are missing [20].
The k-MSN method enables the simultaneous modelling of multiple variables while highlighting the correlation between them. However, if there is no prior limitation on the
maximum number of dependent variables used in the analysis, there may be a tendency to construct overly complex models that may not be accurate for all observations. Therefore, the selection of the number of dependent variables is a key consideration in this type of non- parametric approach [21].
As previously mentioned, k-MSN produces a weighting matrix for selecting the k most similar neighbour. In order to generate this matrix some requirements are needed, such as the specification of the distance measure used to establish the proximity of a particular point, being the Euclidean or Mahalanobis distance the most used [19]. The R software package yaImpute was used to perform this matrix from the dataset. This package is widely implemented in many disciplines using k-nearest neighbour imputation and offers a variety of diagnostics for comparison between results generated by different imputation analysis. There are two main functions contained in yaImpute: yai (canonical correlation and distance measures calculations) and impute.yai (impute calculation) [22].
The analysis started by removing the plots without any stem within the dataset. Subsequently, we defined the dependent variables and the weights for each variable, in this case 1.6 for mean height and 1.4 for tree density. The choice of weights was iterated following a trial and error approach, using different combinations each time the whole analysis was run. Within the range of different distance metrics available within the yaImpute package (msn, raw, Mahalanobis and Euclidean), msn was chosen due to its good performance in the analysis.
This distance is computed in a projected canonical space. The number of neigbours (k) was bounded to 3 and the number of predictors used in each iteration was limited to 18.
After the imputation analysis the simulated annealing (SA) algorithm was used to optimize the functions provided in the previous step. This is a probabilistic technique for approximating the global optimum. The GenSA R software package was used for this purpose. This technique is heuristic, so it was unlikely that the overall optimum would be found, but instead some reasonably good local was selected. A different solution was given each time applying a lower and upper limit to the optimization range for each dimension (as well as the maximum number of iterations, in this case 9000). Therefore, as the algorithm works, there is a slow decrease in the probability of accepting worse solutions as the space is explored. Figure 9 graphically shows this decrease.
The last step was the model assessment using the leave-one-out cross validation method (LOOCV). This approach considers a single sample as a validation sub-block (the iterative removal of a single plot) using the rest as a training sub-block fitting a k-MSN model. This
process is repeated sequentially up to the total number of plots to provide the prediction of the variables of interest.
Fig. 9. Optimization criterion development according to numbers of iterations and decreasing of criterion value, showing a threshold of 4000 iterations.
Absolute and relative root mean square error (RMSE and rRMSE) was used to evaluate the predictions and enable comparison with other studies. This is a measure of the difference between values provided by estimators and the values observed. It indicates in the next formula:
Where; yi is the predicted value xi is the observed value and n is the number of plots.
3. RESULTS AND DISCUSSION
Two main variables were evaluated in the present study: stem density (N, trees ha-1) and mean height (hm, m). The results observed in this study were 2926 trees ha-1 and 1.08 m for RMSE (50.8 % and 48.5 % respectively for rRMSE) at plot level. A total of 9 explanatory variables were used to model the variables of interest. These variables corresponded to all categories except those based on spectral information extracted from the orthomosaic product.
Metrics based on the normalized point cloud were the most selected for the modelling of the two variables because 5 variables were included as predictors. The explanatory variables selected were DSMmax, dsmSECMOMm, DzP100, dsmDISSs, locMax5, locMax7, dsmHOMOs, IP50 and locMax15.
3.1 Prediction of stem density
Stem density estimation remains a challenge using data from different remote sensing platforms (from ALS to UAV) offering low levels of accuracy comparing with others forest attributes such as basal area or timber volume. This is particularly evident when a high- density point cloud is not available. Most of the related literature obtained similar results to those of this study when considering stands with high stem density, despite the variability in predictions. In the current research, the relative RMSE revealed a value of 50.8 % and the RMSE was 2926 stems/ha at plot level.
Puliti et al. [5] analyzed the same UAV-based dataset used in the present study to compare its performance with respect to ALS data applying a different approach, the Random Forest (RF) model, providing an excellent opportunity for a direct results comparison. Their values in terms of RMSE% and RMSE for tree density were 36.4% and 2024 trees ha-1 respectively, being more accurate than those found with the k-MSN method. In addition, Puliti et al.
obtained better results in terms of RMSE values when the analysis was scaled up from plot to stand level because of the reduction of extreme observations.
Iqbal et al. [23] used the non-parametric modelling approach Random Forest (RF) algorithm with different datasets (small and medium-format digital aerial photography, SFP and MFP) and showed a relative RMSE of 68.9 %. It is not easy to compare studies when sampling designs, data availability and tree species are different. Even so, a comparative framework of similar variables within different studies could be developed. Considering that obtaining a
high level of precision to estimate the tree density is quite challenging, future lines of research should focus on this forest inventory variable.
Gobakken et al. [24] used 151 training plots in young forests in a study area of 852.6 ha and they proved that UAV-based photogrammetric data are better estimates for mean height than ALS data. However, the UAV-based photogrammetric data was not a better estimate for stem density (N). They yielded a 43.7 % of relative RMSE and it is consistent with the results for this study. There is still a lack of literature regarding the use of photogrammetric data to estimate the number of stems and the few existing studies showed very different results.
Kukkonen et al. [25] compared two image matching algorithms (semi-global matching and next generation automatic terrain extraction) with ALS data in order to assess the performance in making plot level predictions of stem number in a typical managed boreal forest in Southern Finland. The area-based approach estimates were in lines with the findings on our research and the results revealed that spectral information decreases the prediction error regarding the stem number more than in any other variable. Despite this, the ALS data better predicted tree density and neither algorithm was better than the other for estimating this variable.
Fig. 10. Scatterplot showing predicted values (y-axis) versus observed values (x-axis) for tree density (trees ha-1) at plot level after leave-one-out cross validation.
3.2 Prediction of mean height
After the leave-one-out cross validation the prediction for mean height was used to reveal a RMSE of 1.08 meters and 48.5 % for rRMSE for all stands at plot level. Observing figure 11 of the predicted against the ground reference values, it is remarkable that there is an over- estimation for height values ranging between 0.5 and 2 m. It can be assumed that the model is not able to accurately predict the mean height of the smallest trees in regeneration plots where tree density is quite large.
As previously mentioned, Puliti et al. [5] achieved better results in comparison with this study using the same UAV photogrammetric data but utilizing a different methodology and they found a rRMSE and RMSE values of 30.9 % and 0.77 m respectively. The approach to modelling forest attributes was different in both investigations. In the present study, we made a simultaneous prediction of different biophysical variables with the same explanatory variables while Puliti et al. modelled each stand attribute separately.
Nurminen et al. [26] evaluated 89 test plots in Finland for the same species as the present study using different combinations of forward overlap percentage in aerial images acquisition (60-80). They predicted the mean height (hm) at plot level among other variables required for a forest inventory and the results were used to assess the performance of digital surface models extracted from a high-density photogrammetric point cloud. They proved that aerial images were about as accurate as ALS in estimation of forest resources as long as a precise terrain model was available (e.g. ALS-based DTM or DAP-based DSM). Thus, to improve the absolute accuracy of the DSMs generated, it is necessary to use precise ground control points to reduce the errors obtained in absolute coordinates.
Järnstedt et al. [27] presented a study in a 2000 ha state-owned forest with 402 circular plots.
The estimation of the forest attributes was acquired through k-NN imputation with Euclidean distance weighting average and the accuracy assessed via cross-validation.They demonstrated the effectiveness of the UAV photogrammetric data for providing a good-quality DSM in order to predict forest vertical structure variables (such as dominant height) compared to a surface model based on ALS data. The non-parametric method used for modelling the estimation of the biophysical variables helps to produce an idea of the capacity of these methods to generate accurate predictions compared to other non-parametric techniques. A canopy height model (CHP) based on both surface models was used to perform the forest attributes predictions and they found a rRMSE of 28.23 % for dominant height at plot level.
Fig. 11. Scatterplot showing predicted values (y-axis) versus ground reference values (x-axis) for mean height (m) at plot level after leave-one-out cross-validation
According to Vastaranta [28], the performance of models based on digital stereoscopic images (DSI) to predict mean height compared to ALS data is very accurate, reaching 11.2%
for rRMSE in a 2000 ha managed boreal forest. Species composition (even-aged stands of Norway spruce and Scots pine) and UAV flight parameters were similar to the present study.
Bohlin et al. [11] reported that an increase in image overlap and also in GSD resulting from lower flight altitudes does not improve the estimation accuracy of the assessed forest variables. A k-MSN method was used in this study for modelling mean height and thus enables a direct comparison with our results. Measured in relative terms (rRMSE) Bohlin et al. yielded better accuracy (7.4 %) but the k value (k=1) during k-MSN imputation process was different to this study (k=3). This discrepancy is a possible cause for the difference in accuracy levels. They established a low k value in order to maintain the natural dependencies between the forest variables of interest (e.g. tree height, basal area and volume).
Zarco-Tejada [29] used a low-cost camera on board on a fixed wing platform with similar UAV data acquisition parameters (overlap percentage and GSD) but different - scope (assessment of the effects of different spatial resolution on DSM generation) and - methodology (tree height was extracted from the DSM assuming that a tree top is probably
represented by a local maxima). Nonetheless, they performed the quantification of tree heights in a discontinuous canopy with excellent results (11.5 % rRMSE for tree heights with a range between 1.16 and 4.38 m).
In accordance with Naesset et al. [30] errors in the positioning of the plots can induce a lack of precision in certain variables such as basal area, volume and canopy heights of the forest in which these variables are being assessed. This reason suggests that better sampling design by increasing the number of plots and clusters and longer time periods of GPS data collection can increase the prediction of this variable in regeneration stands such as the case study.
Another investigation that focuses on the estimation of the biophysical properties of the canopy such as canopy height is the research carried out by Krause et al. [31]. This study was carried out in the Schorfheide Biosphere reserve located in Brandenburg (Germany). This research station has 9 fixed plots with different tree species (mostly Scots Pine) and forest structure. The researchers extracted metrics from canopy height models and DTMs using digital aerial photogrammetry based on images captured with a RGB sensor. Linear regression was performed in order to provide estimates to compare them with field measurements. As a result, a low RMSE of 0.34 m (rRMSE of 2.07 %) was found.
As in the present study, the tree height extraction in a dense forest stand using DAP often leads to overestimation of predictions due to occlusion between trees. Assuming the advantage of ALS data over UAV data in describing the vertical structure of the forest under certain conditions (e.g., sun angle or very dense canopy cover), it is relevant to note that the estimation of forest variables based on photogrammetric data can be a low-cost alternative for small forest owners who carry out extensive silviculture in sparse forest areas, taking into account LiDAR technology remains a very expensive solution for this approach.
Another research is recommended to replicate the present experience but implementing the following scenario: mature stands and with accurate information of stem diameter and tree height. In that scenario the k-MSN non-parametric statistical method would perform more accurate estimates as forest variables such as stem diameter and tree height are expected to be more correlated.
4. CONCLUSIONS
The purpose of this study is to contribute to the recent scheme of the unmanned aerial vehicles used in forest data acquisition for enhanced forest inventory purpose providing a new approach and different methodologies. The results were not so accurate in terms of RMSE comparing with other studies based on a combination of ALS+DAP data, but they were promising if we consider that the forest characterization data was based only on photogrammetric data acquisition. This technology is not recommended to capture data when high temporal resolution is required whereas LiDAR does not get affected by weather changes or natural light availability. The costs that are produced by silvicultural treatments can be predicted with the information provided by this approach so this may be favorable to encourage stakelholders involved in the wood products value chain to consider this technology. As recent methodological approaches are developed, the use of unmanned aerial vehicles in the forest sector is expected to increase, which may lead to regular use for small- scale assessment purposes [32].
We need better, faster and more accurate methods of keeping track of the forest resources.
Currently, the sector faces a challenge: how UAV-based data can provide more understanding about tree ecodynamics and its influence on forest management and forest policies. It should be noted that there are several factors that can delay the rapid development of these platforms.
Country-specific legal and operational restrictions on altitude and flight zones may limit their application. Despite this, the new regulations recently approved by the Council of the European Union allows optimism due to the future regulation of standards at European level, this will result in richer interactions between different countries, which will boost the creation of new business models based on bioeconomy strategies. UAV remote sensing community should move towards a common standardization framework in terms of flight parameters (GSD, flight altitude, sensor optical parameters, images overlapping percentages, etc.) and the application of new research findings into an operational environment across different Forest- based value chains and types of ecosystems.
Finally, advances in image-matching algorithms, Earth-observing satellite programs (e.g.
Copernicus, Landsat) and the increased availability of nation-wide LiDAR data could provide in the future the basis for global forest monitoring as daily weather forecasting is currently performed.
5. REFERENCES
[1] Government of Norway. Minisitry of Agriculture and food. 2014. [Online article]
Available at: https://www.regjeringen.no/en/topics/food-fisheries-and- agriculture/skogbruk/innsikt/skogbruk/id2009516 (accesed on 3 March 2014)
[2] Meyer, P., Janda, P., Mikolas, M., Trotskiuk, V., Krumm, F., Mrhalova, H., Synek, M., Labusova, J., Kraus, D., Brandes, J., Svoboda, M. (2017) ‘A matter of time : self regulated tree regeneration in a natural Norway spruce ( Picea abies ) forest at Mt . Brocken, Germany’, European Journal of Forest Research. Springer Berlin Heidelberg, 136(5), pp. 907–921. doi: 10.1007/s10342-017-1079-7.
[3] Fløistad, I. S. and Granhus, A. (2017) ‘Germination and seedling establishment of Norway spruce ( Picea abies ) after clear-cutting is affected by timing of soil scarification’, New Forests. Springer Netherlands, (1). doi: 10.1007/s11056-017-9616- [4] Korhonen, L., Pipputi, I., Packalen, P., Heikkinen, V., Maltamo, M., Heikkilä, J.
(2013) ‘Detection of the need for seedling stand tending using high-resolution remote sensing data’, Silva Fennica, 47(2), pp. 1–20.
[5] Puliti, S., Solberg, S. and Granhus, A. (2019) ‘Use of UAV Photogrammetric Data for Estimation of Biophysical Properties in Forest Stands Under Regeneration’. doi:
10.3390/rs11030233.
[6] Kangas, A. & Maltamo, M. (2006). Forest Inventory: Methodology & Applications.
10.1007/1-4020-4381-3
[7] Goodbody, T. R. H., Coops, N. C. and White, J. C. (2019) ‘Digital Aerial Photogrammetry for Updating Area-Based Forest Inventories : A Review of Opportunities , Challenges , and Future Directions’. Current Forestry Reports, pp. 55–
75.
[8] Holopainen, M., Vastaranta, M. and Hyyppä, J. (2014) ‘Outlook for the Next Generation’s Precision Forestry in Finland’, (July). doi: 10.3390/f5071682.
[9] Tang, L. and Shao, G. (2015) ‘Drone remote sensing for forestry research and
practices’, Journal of Forestry Research. Northeast Forestry University, 26(4), pp.
791–797. doi: 10.1007/s11676-015-0088-y.
[10] Floris, A., Clementel, F., Colle, G. & Gubert, F. (2014) ‘Stima di volumi legnosi forestali con dati fotogrammetrici telerilevati da UAV su piccole superfici : un caso di studio in Trentino Stima di volumi legnosi forestali con dati fotogrammetrici telerilevati da UAV su piccole superfici : un caso di studio in’, (November 2012).
[11] Bohlin, J., Wallerman, J. & Fransson, J. (2012) Forest variable estimation using photogrammetric matching of digial aerial images in combination with a high- resolution DEM, Scandinavian Journal of Forest Research, 27:7, 692-699, DOI:
10.1080/02827581.2012.686625
[12] Feduck, C., McDermid, G. J. and Castilla, G. (2018) ‘Detection of coniferous seedlings in UAV imagery’, Forests, 9(7), pp. 1–15. doi: 10.3390/f9070432.
[13] Anderson, K. and Gaston, K. J. (2013) ‘Lightweight unmanned aerial vehicles will revolutionize spatial ecology’, Frontiers in Ecology and the Environment, 11(3), pp.
138–146. doi: 10.1890/120150.
[14] Sieberth, Till & Wackrow, Rene & Chandler, Jim. (2015). UAV image blur - its influence and ways to correct it. ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. XL-1/W4. 33-39.
10.5194/isprsarchives-XL-1-W4-33-2015.
[15] Triggs, B., Mclauchlan, P., Hartley, R., Fitzgibbon, A. (2010) ‘Bundle Adjustment – A Modern Synthesis. Id : inria-00548290 Bundle Adjustment — A Modern Synthesis’.
[16] Mlambo, R., Woodhouse, I., Gerard, F., Anderson, K. (2017) ‘Structure from Motion (SfM) Photogrammetry with Drone Data : A Low Cost Structure from Motion (SfM) Photogrammetry with Drone Data : A Low Cost Method for Monitoring Greenhouse Gas Emissions from Forests in Developing Countries’, (January). doi:
10.3390/f8030068.
[17] Agisoft. Agisoft PhotoScan User Manual Professional Edition, Version 1.4 Publication date 2018 Copyright © 2018 Agisoft LLC
[18] Giannetti, Francesca & Chirici, Gherardo & Gobakken, Terje & Næsset, Erik &
Travaglini, Davide & Puliti, Stefano. (2018). A new approach with DTM-independent metrics for forest growing stock prediction using UAV photogrammetric data. Remote Sensing of Environment. 213. 10.1016/j.rse.2018.05.016
[19] Packalen, P. and Maltamo, M. (2007) ‘The k-MSN method for the prediction of species-specific stand attributes using airborne laser scanning and aerial photographs’, 109, pp. 328–341. doi: 10.1016/j.rse.2007.01.005.
[20] Moeur, M & R. Stage, A. (1995). Most Similar Neighbor: An Improved Sampling Inference Procedure for Natural Resource Planning. Forest Science. 41. 337-359.
[21] Maltamo, M., Gobakken, T. and Næsset, E. (2013) ‘Comparing parametric and modelling of diameter distributions on independent data using airborne laser scanning in a boreal conifer forest ˚’, (July), pp. 493–501. doi: 10.1093/forestry/cpt020.
[22] Crookston, Nicholas L.; Finley, Andrew O. (2007). yaImpute: An R Package for k-NN Imputation. Journal of Statistical Software. 23(10):1-16.
[23] Iqbal IA, Musk RA, Osborn J, Stone C, Lucieer A. A comparison of area-based forest attributes derived from airborne laser scanner, small-format and medium- format digital aerial photography. Int J Appl Earth Obs Geoinf. Elsevier;
2019;76:231–41.
[24] Gobakken T, Bollandsås OM, Næsset E. (2014). Comparing biophysical forest characteristics estimated from photogrammetric matching of aerial images and airborne laser scanning data. Scand J For Res. 2015;30:73–86
[25] Kukkonen, M., Maltamo, M. and Packalen, P. (2017) ‘International Journal of Applied Earth Observation and Geoinformation´ Image matching as a data source for forest inventory – Comparison of Semi-Global Matching and Next-Generation Automatic
Terrain Extraction algorithms in a typical managed boreal forest environment’, International Journal of Applied Earth Observations and Geoinformation. Elsevier B.V., 60, pp. 11–21. doi: 10.1016/j.jag.2017.03.012.
[26] Nurminen K, Karjalainen M, Yu X, Hyyppä J, Honkavaara E. (2013) Performance of dense digital surface models based on image matching in the estimation of plot-level forest variables. ISPRS J Photogramm Remote Sens. International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). 2013;83:104–15
[27] Järnstedt J, Pekkarinen A, Tuominen S, Ginzler C, Holopainen M, Viitala R, et al.
Forest variable estimation using a high-resolution digital surface model. ISPRS J Photogramm Remote Sens. 2012;74:78–84.
[28] Vastaranta, M., Wulder, M., White, J., Pekkarinen, A., Tuominen, S., Ginzler, C., Kankare, V., Holopainen, M., Hyyppa, J. (2013) ‘Airborne laser scanning and digital stereo imagery measures of forest structure : comparative results and implications to forest mapping and inventory update’, (February 2014). doi: 10.5589/m13-046.
[29] Zarco-Tejada, P.J., Diaz-Varela, R., Angileri, V., Loudjani, P. (2014). Tree height quantification using very high-resolution imagery acquired from an unmanned aerial vehicle (UAV) and automatic 3D photo-reconstruction methods. European Journal of Agronomy, vol. 55, pp. 89-99.
[30] Næsset, E. & Gobakken, T. (2009). Assessing the effects of positioning errors and sample plot size on biophysical stand properties derived from airborne laser scanner data. Canadian Journal of Forest Research, vol. 39(5), pp.1036-1052.
[31] Krause, Stuart & Sanders, Tanja & Mund, Jan-Peter & Greve, Klaus. (2019). UAV- Based Photogrammetric Tree Height Measurement for Intensive Forest Monitoring.
Remote Sensing. 11. 758. 10.3390/rs11070758.
[32] Torresan, Chiara & Wallace, Luke. (2016). Forestry applications of UAVs in Europe:
areview. International Journal of Remote Sensing.
