Introduction

Forests are complex terrestrial ecosystems that are dynamic in time and space. They are home to 80% of the terrestrial biodiversity of the planet [1]. Since time immemorial, people have benefited from the numerous functions of the forest in the form of goods and services provided by forest ecosystems, such as the provision of timber, clean water and air, protection against natural hazards, and many others. However, forests in Central Europe have not always been treated with the same care as they are today, and the complexity and interdependence of their functions have only recently become valued. By the late 14th century, Central Europe’s forests were severely damaged by over-exploitation, resulting in a timber shortage. This shortage led to the first attempts at planned reforestation, and the subsequent genesis of the fields of forestry and forest sciences [2]. By the late 18th century, the principle of sustainability was developed, advocating the creation and conservation of forests and the use of wood in a stable and sustained manner [3]. Ensuring sustainability required an intimate knowledge of the current condition and extent of forests, as well as their development; this need formed the foundation for the field of forest inventory [4]. Forest inventory is defined as the systematic collection of data and forest information for their assessment or analysis. Basic information collected in forest inventories includes species, diameter at breast height, height, age, defects, and site quality. Such detailed inventories are still carried out today both at the national and local level. However, they are labour intensive and require trained personnel, making them very costly. In the 20th century, new and more efficient methods using 3D mapping technology have been developed, but their application has focused mainly on assessing timber volumes and the potential for timber harvesting.

With an increasing population and growing resource demands, managing forests only for the provision of timber is no longer sufficient. In combination with rising labour costs and declining timber prices, there is now a clear need for more affordable yet detailed solutions. Additionally, the major impact of climate change is leading to initiatives such as climate-smart forestry [5]. The emphasis is on creating resilient forest ecosystems where timber is no longer the main product. That being said, such a forestry approach will lead to even more demanding forest management and inventory work.

The development of 3D capture technology has triggered a revolution in the way forest resources are surveyed. Airborne Laser Scanning (ALS), for example, has made it possible to map extensive areas of forests and even whole countries [6,7,8]. The quantification of point cloud in terms of different types of statistics has facilitated the development of statistical models that make it possible to predict many biometric features of trees and to characterise forest areas continuously (wall-to-wall maps) [9].

High-altitude aerial methods, such as ALS and photogrammetry, provide a large-scale forest perspective but with sparse detail. On the other hand, Terrestrial Laser Scanning (TLS), Mobile Laser Scanning (MLS), and close-range photogrammetry technologies, deployed in either a ground-based or a close-range aerial manner via Unmanned Aerial Vehicles (UAVs), provide a small-scale but very detailed perspective on forests [10]. This makes ground-based methods able to map the shape and dimension of individual trees more precisely than aerial methods and to obtain information about the forest understorey and regeneration [11,12,13,14].

These 3D mapping technologies may therefore be considered an alternative to traditional forest measurements and are often used in forestry and forest ecology studies, with a trend towards more use of laser scanning [15]. In the last two decades, many studies have demonstrated the high accuracy of direct measurements of forest parameters when using TLS technology [16, 17]. However, its practical application remains a challenge due to the variety of devices, the limitations imposed by the cost of implementing these technologies, and, most importantly, the lack of user know-how and lack of standards regarding data collection and processing. Furthermore, depending on the scale level of the inventory (national, local, or anything in between), different 3D technologies may be considered. As no standard currently exists on the levels of scale and detail, it may be difficult for users to determine which sensor to use in which circumstances. In many cases, there is also a need to develop algorithms for detecting and determining target characteristics of forest ecosystems, due to the highly fragmented processing solutions. However, the intensification of scientific work and technological developments in recent years suggest that these technologies will see considerable use in the near future.

Three-dimensional point clouds have provided forest practitioners and scientists with a completely new way of assessing and monitoring forest resources and services, and of conducting research that was previously impossible. As a result, more and more scientific groups and practitioners are intensively developing, often in parallel, methods and technologies to automate the surveying of ground plots and the determination of stand characteristics using point clouds. Here, again, there is a lack of standardisation and dynamic comparison with a focus on end users, such as foresters, ecologists and scientists. There is therefore a need for a joint initiative to manage the new findings and make standards for the above-mentioned end users.

In this paper we describe the results of one such initiative, conducted within the context of the 3DForEcoTech COST Action. In this initiative, our objectives were: (1) to compile a list of available ready-to-use processing solutions to derive forest characteristics from ground-based point clouds based on criteria such as availability, focus and relevance, and (2) to introduce a web platform with information about the identified processing solutions, their availability, technical guides on installation and general use, and benchmark results. Based on responses to a questionnaire distributed within the vast network created by the COST Action, we identified a total of 24 solutions.

In this review, we formulate three main aims: (1) to explain the use of point clouds in forestry; (2) to summarise forest point cloud processing and various approaches used by the different algorithms; and (3) to describe the 24 solutions compiled in the COST action survey. We also provide an overview of the potentials and limitations of the compiled solutions, for use by practitioners and researchers who would like to process point clouds for forestry applications. The remaining sections of this paper are organised as follows. In Sect. 2 we explain the use of point clouds in forestry. In Sect. 3 we describe a literature study on the state of the art in forest point cloud processing and the different approaches used by the different algorithms. In Sect. 4 we describe the 24 compiled solutions and discuss some of the main observations. We finish with concluding remarks in Sect. 5.

Point Clouds for Forest Applications

Common Point Cloud Acquisition Techniques

A point cloud describes a collection of points known in a cartesian tridimensional system and together forming a 3D object [18]. As such, a point cloud is by nature a geometric entity. Early conceptions of a point cloud already existed in traditional land surveying [19]. However, the generation of dense point clouds -- as the term is commonly understood today -- only started with the advent of lidar [20]. Lidar, or laser scanning, is today one of the techniques most commonly implemented in generating point clouds of real-world objects [16, 21, 22*]. As an active range-based sensor, a lidar device emits laser waves and records the distance between an object and the origin, along with sweeping angles, thus computing discrete 3D coordinates which form the backbone of a point cloud. A distinction is generally made between aerial and ground-based lidar [23, 24]. Aerial lidar, or ALS, may be distinguished according to its platform, with UAV [25] being pertinent within the context of close-range sensing. TLS and MLS are the most prominent examples of ground-based lidar [18]. The term “lidar” refers to the technology used, but is most commonly associated with and sometimes even considered interchangeable with ALS, while TLS and MLS are sometimes referred to as simply “laser scanning” [26]**. For ground-based forest mapping, TLS may be considered as the reference, due to the high quality of the data that may be achieved using this technique [16, 27, 28].

The other major alternative to lidar is photogrammetry. Photogrammetry is a much older technology, dating back to the first use of aerial photography [29]. Unlike lidar, it involves a passive image-based sensor which captures electromagnetic waves reflected by the surveyed object. Photogrammetry originally relied on empirical principles, and later on mathematical ones, to infer 3D coordinates from 2D images [30]. It was not until the last few decades that it managed to rival lidar in the generation of dense point clouds, thanks to new developments in the field of computer vision. Automated image orientation was developed in parallel with Structure from Motion (SfM) methods [31], while Multi-View Stereo (MVS) and dense matching principles [32, 33] truly boosted photogrammetry’s popularity. Recent developments also saw an increasing interest in learning-based MVS [34] and novel 3D rendering methods, such as Neural Radiance Fields (NeRF) [35] and 3D Gaussian splatting [36]. Similar to its lidar counterpart, photogrammetry may be implemented both from an aerial and from a terrestrial perspective. Aerial photogrammetry traditionally involves the acquisition of nadir images from an aerial platform, which includes drones. However, oblique views are also common, especially in close-range photogrammetry [37]. Terrestrial close-range photogrammetry is especially known to be able to deliver high-precision results with a relatively low initial investment [38, 39]. However, it is not applied often in a forestry setting, mainly due to its difficult set-up in a forest environment. Indeed, traditional pinhole photogrammetry relies on multiple overlapping images taken in a convergent network, something which is difficult to achieve in a heterogeneous and uneven environment [40].

In recent years, novel sensors have been developed with a focus on portability and low cost, at the expense of precision. This philosophy of sensor development generally tries to fill the gap between very high precision, expensive solutions and low cost, generally lower quality ones. An interesting example can be seen in the development of MLS, which combines lidar technology with Simultaneous Localisation and Mapping (SLAM) methods. MLS has recently seen many applications in forestry, thanks to its portability [16, 41]. While its precision is generally lower than stationary TLS, in many cases it is high enough for mapping forest attributes. The same reasoning has also pushed the use of low-cost sensors in forestry, for example, depth cameras [42], spherical and fish-eye photogrammetry [40, 43], and the novel Solid-State Lidar (SSL) [44]. Figure 1 summarises the different categories of close-range point cloud generation techniques. In this paper, we focus on solutions for processing ground-based point clouds.

Fig. 1
figure 1

The different close-range point cloud acquisition techniques available for forest mapping, categorised according to the type of data acquired and then further divided into sub-categories according to the sensor and algorithm (for image-based techniques) and hardware type (for range-based techniques)

Full size image

In terms of visualisation, the increasing availability of affordable online platforms for processing large 3D point clouds has facilitated the integration of point cloud data with cutting-edge visualisation technologies, such as Virtual Reality (VR) [45]. A major challenge in 3D rendering is related to memory requirements. To overcome this issue, many methods involve converting point clouds into meshes to optimise memory usage and ensure smooth visualisation [46, 47].

Data Types and Formats

The 3D representation of an object may take several forms, with the point cloud being one of the most common and the simplest in structure: point clouds are at their geometric base simple lists of coordinates. Other forms of 3D representations, like meshes, as well as volumetric and parametric models, are also commonly used, depending on the requirements. 3D meshes often consist of triangles, whose vertices are extracted from the point cloud. Volumetric and parametric 3D models can use simple geometric primitives and are also commonly used in information systems, e.g. Building Information Models (BIM), or Quantitative Structure Models (QSM). Despite their geometric simplicity, point clouds can be stored in various formats: binary files, which are usually fast to read/write and allow compact storage; and text files, which are more inefficient but simpler to use and adapt. Point clouds can also be stored in a structured or an unstructured manner. Unstructured point clouds are simply lists of coordinates and attributes that can be conceptually pictured as a data table with as many rows as points and as many columns as dimensions and attributes. All the points in an unstructured point cloud must be in the same coordinate system. Conversely, structured point clouds have a more complex arrangement: they store the data as they were gathered in the field, together with all the additional information needed to generate a coherent point cloud with a unified coordinate system. Structured point clouds are frequently generated in ground-based laser scanning (specifically in TLS) and with depth-cameras but are not so common in ALS and photogrammetry.

Table 1 lists some of the most common point cloud formats on the market. Additionally, some formats support mesh and volumetric model representations on top of the point cloud. LAS is a binary and unstructured format that is used as a general exchange file format. It was initially designed for ALS point clouds, but due to its simplicity it is now used for any point cloud type. Most software for processing point cloud data includes reading and writing capabilities for this format. LAZ is a very common variant of LAS that allows data compression (both with or without information loss). Text files are also frequently used for point cloud storage and exchange. In most cases, data are stored in ASCII code and in an unstructured manner, with one point per line in the text file and the coordinates and attributes separated with commas, spaces or tabulated spaces. However, although there are some predefined text-file formats for point clouds, there is no clear standard for extended use, not even for the inclusion of metadata and/or headings.

Table 1 Features of some of the most common 3D formats available for 3D point cloud data
Full size table

Regarding structured point clouds, almost every manufacturer of ground-based laser scanners has developed their own format. These formats may be used for point cloud registration, denoising, colouring, and initial shape detections. However, in most cases, especially in forestry, these formats are only used for pre-processing the point clouds before transforming them into other exchange formats, such as LAS/LAZ or text files. E57 is another popular manufacturer-agnostic structured point cloud file format with read/write support in many software solutions and algorithms. This format is often encountered with TLS in the fields of engineering, heritage and architecture, but it is not yet popular in forestry. Very few software solutions related to forestry allow the use of this format.

Examples of Point Cloud Applications in Silviculture and Forest Ecology

Point cloud data obtained through laser scanning or similar technologies that generate 3D representations have numerous applications in silviculture and forest ecology. Furthermore, 3D data become increasingly important within the monitoring context of Essential Climate Variables (ECV) and Essential Biodiversity Variables (EBV) [48, 49]. Some examples of point cloud applications are:

  1. 1.

    Forest and tree attribute inventory: Point cloud data can be used to estimate conventional to complex forest attributes, such as tree height, Diameter at Breast Height (DBH), canopy cover, leaf area distribution, stem volume and Above Ground Biomass (AGB) [17, 50**, 51]. This information is crucial for forest management and monitoring purposes. For example, lidar data, combined with allometric models, can be used to estimate above ground carbon stocks in forests [52]. This information may be used to assess the carbon sequestration potential and to evaluate the effectiveness of climate change mitigation strategies.

  2. 2.

    Tree species classification: By analysing the structural characteristics of point cloud data, such as point density and canopy shape, we can apply machine learning algorithms to retrieve information at the individual tree level, e.g. using semantic and instance segmentation of point clouds [53].

  3. 3.

    Forest structure analysis: Point cloud data enable the quantification of forest structural parameters like canopy height profiles, vertical vegetation layers, and canopy gap distribution [54, 55]. By using lidar data, we can also provide valuable information on the need fire modelling, estimate forest canopy fuel parameters, map fire risk, and evaluate the effectiveness of fire management strategies [56].

  4. 4.

    Forest regeneration monitoring: Point cloud data can help assess the success of forest regeneration efforts by quantifying sapling density, height and spatial distribution within a forested area [57, 58]. These data aid in the evaluation of forest recovery after disturbances such as logging and fire.

  5. 5.

    Forest visualisation: Recent advancements in 3D scanning technologies and the increasing availability of affordable online platforms for processing large 3D point clouds have facilitated the integration of point cloud data with cutting-edge visualisation technologies, such as VR [45]. These applications can play a crucial role in supporting forest management practices and have the potential to contribute to the education of future foresters [59].

State of the Art of Processing Algorithms

Point Cloud Processing Pipeline in Forestry

A main goal of the pipeline of point cloud processing for forestry is to derive or directly measure information about essential parameters of the forest on an individual tree basis from captured point clouds. We consider measurements of tree dimensions, such as DBH, tree height, volume and selected crown parameters, essential parameters for forestry and precision forestry. The post-acquisition pipeline can be divided into general pre-processing and thematic (forest-specific) processing. The pipeline is specific for each of the processing solutions that are included in this paper, first because there are various goals of these solutions and second because the developers took different paths, for example in the selection of the programming language.

As the first step, general pre-processing is normally done within the dedicated software of the scanning device, depending on the acquisition method. For example, manufacturers of laser scanners (whether terrestrial or mobile) provide robust software solutions focusing on pre-processing and many additional post-processing options. This usually covers the registration of multiple scan positions for TLS or the application of SLAM post-processing for MLS data, in order to register the point cloud accurately. Another important process is the georeferencing, filtering and classification of the point clouds. These solutions are however commercial in nature and is thus unavailable for users who do not possess a specific license. In the case of photogrammetry, the 2D images are processed into 3D point clouds. This task is fully handled by photogrammetric software. Some of the more popular options include Agisoft Metshape (Agisoft LCC, Saint Petersburg, Russia) and Pix4D (Pix4D S.A., Prilly, Switzerland) [60], both of which are also commercial solutions. The pre-processing of photogrammetric point clouds is more computationally demanding than that for lidar point clouds [61], but it generally has the major advantage of lower cost.

In general, these methods or even tool-specific software packages provided by manufacturers do not have tools for individual tree measurements. They therefore constitute a first step in the pipeline, whose aim is to prepare the point cloud for further processing more specifically targeting individual tree measurements. In this regard, the thematic processing of the pipeline has the goal of measuring the parameters of individual trees. The method with which this is done varies based on the software or algorithm used. It usually starts with the segmentation or classification of the point cloud. This can be approached with different levels of complexity. For example, the FSCT pipeline [62] starts with semantic segmentation, where the point cloud is segmented into four categories using deep learning. On the other hand, simpler approaches that do not use deep learning first focus on the classification of terrain points. For example, Dendrocloud [63] divides the 3D point cloud by raster projection with a specified cell size, where the minimum z-value is searched and assigned to that particular cell. 3DForest [64] on the other hand, uses a voxelisation of the point cloud, where the minimal z-value is iteratively searched through neighbouring voxels. Based on these points, digital terrain is created, which is then used as a normalised surface from which cross-sections are generated.

The most important parameter to measure is DBH, as evidenced by the overwhelming availability across the processing solutions from the compiled list described in the following sections. In most solutions this is done on spatially grouped cross sections, often using either circle fitting or cylinder fitting with the help of the Random Sample Consensus (RANSAC) algorithm. In this regard, individual tree detection is a prominent functionality which would enable the solutions to compute the DBH as well as other tree parameters such as diameter at multiple heights, tree height, and stem volume.

Notions of Levels of Detail and Scene Scale

Different scenes may require different types of sensors and various kinds of processing strategies, depending on the scale of the scene and the requirements of the application. While a systematic formal definition of levels of detail in the general use of 3D data, particularly in urban environments, was presented in e.g., the CityGML paradigm [65], similar attempts for formal but specific forestry definitions are lacking in the literature. One such attempt was presented in [66], summarised in Fig. 2. [67*] also presented an interesting approach to categorise 3D data generation techniques based on the complexity and size of the scene; this approach influenced the creation of Fig. 2, in which the levels of detail are divided according to the spatial scale.

Fig. 2
figure 2

Adapted from [66]

Categorisation of the different levels of detail in 3D forests based on their spatial scales and how several 3D reconstruction techniques can address the requirements of these scale levels

Full size image

In general, definitions of scales and levels of detail exist in discussion of forests. While these definitions are not directly analogous to similar ones used in urban settings, it is possible to propose a sufficiently descriptive categorisation example, at least for the purposes of this paper. This notion of level of detail in forest point clouds is by no means authoritative, in part because the definitions of forest scale levels may also be subject to different interpretations. As can be seen in Fig. 2, five scale levels have been identified, ranging from very small objects (e.g. microhabitats) to large scenes. Figure 2 likewise proposed a categorisation of several 3D techniques in responding to the needs of each scale level.

Figure 2 refers to five scale levels, namely micro, small, medium, large, and very large. These levels were identified based on a purely spatial data point of view; this means that the levels’ definitions refer to both the area of the forest to be covered by the 3D mapping and the expected geometric accuracy of the point cloud. In this context, note that the measurement of tree parameters such as DBH, tree height, or tree position are sensor-agnostic in nature since they are computed as derivatives of the point cloud as the main result of 3D sensors. However, the precision and accuracy of those parameters will be highly related to the quality of the point cloud and therefore choice of sensor; hence the proposal suggested by Fig. 2 to help future new users of the technology decide which sensor is best suited to their needs.

Nevertheless, an important notion in the discussion of levels of detail is the relationship between the expected quality (be it in terms of point cloud resolution, precision or accuracy) and the most appropriate technology to attain it. This in turn influences the way processing algorithms are developed. It is worth noting that in Fig. 2 both TLS and MLS represent a “middle-ground” compromise between details and scale. This explains their popularity in forest applications, as highlighted by the identified processing algorithms. Figure 2 does not, however, take into account other factors, such as occlusion in the forest.

Heuristic vs. Machine Learning Methods

Point cloud processing algorithms can be roughly split into two groups: heuristic and machine learning algorithms. Heuristic algorithms represent a set of logical rules that guide the user step by step toward the target result. In point cloud processing routines, heuristic approaches usually operate on the fitting of geometric primitives (lines, circles and cylinders), the calculation of statistics/features per area unit (e.g. cells) or space unit (e.g. voxels), and feature thresholding. Due to their logical and understandable nature (hence the term “knowledge-based”), heuristic approaches often serve as a starting approach for extracting target information from a point cloud. They are especially suitable when the amount of data is limited [68]. Today, in the forest domain, heuristic approaches dominate point cloud processing routines and are often used to extract a wide range of forest characteristics, e.g. the identification of individual trees and tree stems [69,70,71], DBH and tree height [72,73,74,75], forest structure characteristics [76], and Leaf-Area Index (LAI) [77,78,79].

A prominent example of the great success of a heuristic approach to point cloud processing is the QSM, which comprise a set of rules to reconstruct tree architecture using cylinder-based models [80, 81]. These models are widely used to derive the total volume and AGB of the tree, as well as its components [82,83,84]. However, heuristic algorithms may suffer from generality issues (intra- and extra-technological transferability/scalability). When they are applied to new data, some processing steps in heuristic algorithms might need adjustments (e.g. reconsidering thresholds and adding or removing processing steps), contributing an empirical aspect to the point cloud processing.

In contrast to heuristic algorithms, Machine Learning (ML) is generally used to extract forest characteristics that do not follow a clear geometric pattern and are hardly describable using a set of logical rules. ML implies supervised or unsupervised learning on a variable space, which is usually compiled using engineered features or real measurements (e.g. XYZ coordinates, spectral response). This approach previously operated on classic machine learning algorithms (e.g. Random Forest, Support Vector Machine, and XGBoost) and a set of engineered features to identify tree species [85,86,87] or to separate leaves from wood [88,89,90]. Today, however, Deep Learning (DL) is gaining attention from forest researchers. In other domains, DL has achieved state-of-the-art performance (sometimes even outperforming humans) in classification, segmentation and object detection tasks for both image and point cloud data. Forest researchers have begun to explore its potential for individual tree segmentation [91,92,93], tree species identification [94,95,96], and semantic point cloud segmentation [28, 97, 98]. However, the forest domain is generally a user of existing DL solutions rather than a developer of new ones. Thus, it tends to be a few steps behind the current state-of-the-art. It also suffers from a lack of large and representative public datasets to develop and calibrate DL models and fairly benchmark them against other solutions [99].

Within the context of the algorithms identified in this paper, most use a heuristic-based approach to generate the output. However, both ML and DL have been used in several algorithms in varying levels. Indeed, ML is not always used directly (e.g. for semantic or instance segmentation) but may be used to support the heuristic process, for example in performing individual tree segmentation, before referring to heuristic methods to generate the output parameters.

Identified State-of-the-Art Algorithms

Methodology

The process of compiling the algorithm list for terrestrial point cloud processing software solutions was conducted through a series of structured activities under the 3DForEcoTech Cost Action project. These activities were widely publicised within the 3DForEcoTech community and its extended networks, and they involved various channels and platforms accessible to participants. This effort resulted in an initial list of 65 available software solutions, which was subsequently refined to 24, based on criteria such as availability, focus and relevance. The creation of the initial list involved comprehensive activities conducted within Working Group 3 (WG3) of the COST Action 3DForEcoTech, including the distribution of an online questionnaire and multiple COST Action meetings.

The aim of the questionnaire was to gather preliminary information about algorithm implementations for point cloud processing in forestry, focusing on ground-based point clouds, tree/forest metrics, and publicly available solutions, regardless of their being free, open-source or even commercial. It was distributed to all 450 participants of 3DForEcoTech, representing over 50 countries. Participants were encouraged to share the questionnaire within their professional networks. Additionally, members of 3DForEcoTech convened meetings to complement the questionnaire results by identifying additional software solutions that may not have been considered in the questionnaire.

Following the collection of questionnaire responses and additional research from WG3 meetings, the initial software list comprising 65 solutions was compiled. This list underwent iterative review processes, facilitated by three Short Term Scientific Missions (STSM), each involving a different researcher. STSMs are funded scientific collaborations within the framework of COST Actions. The review process ensured: (i) compliance with the initial questionnaire requirements for participant submissions, (ii) functionality verification of the online-available versions of the software, (ii) evaluation of the software’s capability to process simple point clouds, (iii) assessment of the software documentation, and (iv) specificity for terrestrial point cloud processing. Additionally, technical guides on installation and running instructions, along with relevant supplementary information, were compiled for each software solution.

To create a comprehensive database and overview of terrestrial point cloud processing software solutions, data were gathered from publicly available documentation provided by the authors/distributors of the implementations. Insights obtained during STSMs were also incorporated into the database. This resource compiles essential information for each software solution, including inputs, outputs, processes, and scope of use, and is intended to serve a valuable reference for understanding the functionalities and applicability of each implementation within the context of terrestrial point cloud processing in forest environments. Table 2 contains all the categories and items assessed in the list and database, including their basic information, availability, suitable inputs (point cloud technologies and file formats), scope of the application, and outputs.

Table 2 Categories of the identified algorithms according to the sensor, input, scale and output criteria
Full size table

Identified Software Implementations

Each of the identified software was tested with different configuration environments. The software was mainly tested based on three important requirements, i.e. the ease and requirements of implementation, the main functionality of the software, and the possibility of errors occurring during the installation procedure. Table 3 presents the names of the identified solutions and a few important metadata, including their associated licences. The table also shows that the majority of the solutions are either open-source (20) or free (2), with another 2 available as commercial software. In most cases a relevant scientific publication was identified from the literature, although some contain explanations on the algorithmic background while others focus on its applications. Most of the open-source software solutions are hosted by the git platform www.github.com. Notably, the use of the R language is prevalent, although Python is a close second.

Table 3 General information and availability of the 24 identified algorithms. Green cells denote open-source, yellow free software, and blue commercial solutions
Full size table

Table 4 presents the characteristics of the identified solutions in more detail. In general, of the 24 solutions identified in this paper, all are able to process TLS data. While several solutions do not support SfM and MLS data, most of them are generally sensor-agnostic. LAS is the most prevalent point cloud format, while batch processing is not a common feature. However, it should be noted that most of the identified solutions are in the form of source code. Batch processing is therefore theoretically possible, if not directly available. Note that within Table 4, several cells had either a “probably yes” (PY) or “probably no” (PN). This implies that according to our tests the concerned solution includes respectively availability or non-availability of the criteria mentioned in the column, sometimes in an indirect manner. However, we relegated it to “probably” due to the absence of a formal indication of such capability or lack thereof in the software’s official documentation.

It is also interesting to note that while most solutions provide basic tree parameters, such as DBH (up to 75%) and tree height (up to 54%), a few are highly specialised. For example, Crossing3DForest was designed solely to create QSM models and TLS2trees for stem segmentation. None of the identified software and algorithms provide a feature to compute total leaf area. Figure 3 summarises the findings graphically.

Fig. 3
figure 3

Polar plot showing the distribution of the identified solutions, classified into the pre-determined categories

Full size image
Table 4 Table of the identified software solutions and algorithms, with the availability of features indicated in the columns. “Y” denotes availability and “N” non-availability. “PY” denotes “probably yes” while “PN” signifies “probably no”
Full size table

Online Dissemination Platform

Based on the identified software list and successful testing, an online platform was created for the end users. The information in Table 4 is reflected in the platform and should help users in choosing which solution suits their needs best and meets the required accuracy. As a preliminary system, the platform contains information from Table 4 in the form of a web application. The selection criteria in the platform are based on the categories defined in Table 2 and information from the columns of Table 4 was thereafter fed into it. In this way, the web-based platform may serve the community as a single source of information to select a specific software or algorithm that works for their requirements. Furthermore, the online nature of the platform means that it will evolve in time with regular updates of new algorithms and features. In order to further improve the information presented in the database especially regarding technical capabilities, a benchmark was also performed on the solutions. This benchmarking was performed during a hackathon organised by 3DForEcoTech in September 2023, and its results will be described in a future publication.

The platform is currently hosted within the 3DForEcoTech website (https://3dforecotech.eu/database/ last accessed 24 April 2024), where users may perform queries based on the available categories (represented as columns in Table 4). From the results of this query request, users may then choose a specific algorithm and click on it to see more information on a dedicated page for each algorithm. A conceptual drawing of how the online platform works is given in Fig. 4, and a concrete example of its implementation, taken directly from the website, is showcased in Fig. 5. This information page contains a description of the algorithm, as well as links to the respective codes and/or implementation. Furthermore, for each algorithm identified in the database, metadata were collected during installation and test runs, to assess its applicability for forestry. From these tests, technical guides on installation and general use were written and included in the web platform. These user guides are provided along with installation steps, basic computational configuration requirements, contact details of the author of the tool, and information on how to deal with possible errors in specific computational configurations.

Fig. 4
figure 4

Conceptual representation of how the online platform presents Table 4 as a queriable database

Full size image
Fig. 5
figure 5

A concrete example of how the online query works. In this example, FSCT was queried by the user. The platform provides a description of the software solution and a link to the user guide

Full size image

Comparisons and Discussions

In this review, we leveraged the unique opportunity presented by a community of 450 researchers and practitioners from 50 countries dedicated to and/or interested in the application of close-range technologies for characterising forest environments, along with their extensive networks. We believe there is a pressing need to establish a standardised dynamic database of processing solutions that are dedicated to ground-based point clouds and forest measurements. In this paper we present one option to fulfil this need. It has already been established in many studies that 3D point clouds are well suited to measure individual tree parameters with high levels of detail and accuracy that can even exceed the conventional approaches (e.g. [15, 16]). Furthermore, it is important to note that this technology provides an option to measure on a level of detail that was not possible before. This in turn helps to address questions that previously could only be answered on a theoretical basis. However, these technologies are not yet commonly used by the wider community or relevant stakeholders, such as foresters, forest ecologists and scientists outside the remote sensing field [66].

By conducting a questionnaire and creating a database of processing solutions, we aimed to show what solutions are available and ready to use. More importantly, in this review we hoped to identify what has already been solved properly within the available solutions, thereby aiding the community in avoiding doing work on the same solution in the future. On the other hand, we also aimed to identify gaps in the state of the art to highlight areas where future developers should focus.

Observed Trends

In the discourse surrounding this review, it became evident that the landscape of ground-based point cloud processing in forest environments is primarily oriented towards automating precision forest inventory at the plot level. This involves the meticulous measurement of individual trees, encompassing parameters such as tree location, DBH and tree height. Such an approach closely mirrors the methodology employed in traditional forest inventories, thus establishing a familiar framework for practitioners transitioning into the realm of point cloud analysis. However, despite the prevalence of these automated inventory solutions, there remains a notable gap in the exploitation of the full potential offered by ground-based point clouds. While certain software solutions delve into more complex metrics and analyses, the broader utilisation of these datasets has yet to be fully realised. Ground-based point clouds, by their very nature, offer a spatially explicit and three-dimensional representation of forest structure. This wealth of data holds considerable promise for enabling measurements and estimations that surpass the capabilities of conventional methods, including traditional inventories reliant on manual tree-by-tree measurements, aerial lidar surveys, and other forms of remote sensing. It is imperative to recognise that ground-based point clouds possess unique attributes that distinguish them from other data sources. Unlike traditional inventories, which are often limited by the labour-intensive nature of tree-by-tree assessments, point clouds offer a comprehensive and continuous dataset that captures the intricacies of forest ecosystems in unprecedented detail. Furthermore, their three-dimensional nature facilitates advanced analyses, such as volumetric assessments, canopy characterisation, and habitat mapping, which have the potential to revolutionise our understanding of forest dynamics and biodiversity.

In addition, advancements in Artificial Intelligence (AI) methods applied to point clouds are beginning to usher in algorithms for quantifying and mapping complex variables. However, as emphasised throughout this work, publicly available implementations remain scarce. In the specific case of utilising novel DL methods, additional challenges arise from creating publicly available implementations, including those stemming from the complexity of configuring and executing processes with specific hardware requirements, such as the utilisation and management of GPU-based systems, along with the need for extensive training data and long processing times to ensure functionality. These complexities underscore the ongoing need for further research and development to overcome barriers to widespread adoption, to facilitate user-friendly operability, and to maximise the potential of AI-driven approaches in ground-based point cloud processing.

In light of these considerations, while existing ground-based point cloud processing software solutions have made large strides in automating forest inventory processes, there exists a compelling opportunity to further innovate and expand the scope of analysis. By leveraging the spatially explicit and multidimensional nature of point cloud data, researchers and practitioners can unlock new avenues for ecological research, conservation planning, and sustainable forest management. As such, future developments in this field should aim to harness the full potential of ground-based point clouds, driving forward advancements in forest science and management.

Identified Gaps

This study represents a unique opportunity to gain a comprehensive overview of existing implementations of algorithms aimed at automating forest mensuration, inventory and mapping. Although algorithms can be identified through systematic paper searches, compiling a complete repertoire of available software would require alternative means, which are not always straightforward -- especially for non-specialists. Thus, we encourage researchers to share, along with scientific publications, their point-cloud processing solutions implemented in a way that is as user-friendly as possible. This will foster other researchers to not repeat, but build on existing solutions and develop them further. It is also worth noting that most of the identified algorithms and software are usually focused on a particular problem related to the developer’s needs. Indeed, the solutions are generally good enough in terms of their main functionality but may falter when repurposed for other needs. While this is a logical outcome of the software development process (i.e. to solve a particular problem), there is a growing need for fully automated software which includes all the pre-processing and post-processing steps. The same incoherence can also be seen by the fact that most solutions work with different set-ups in terms of input file format and type (whether plot level or individual tree level). None of the identified software solutions has flexibility for the point cloud input data and file formats, making them quite rigid. Further challenges are also associated with the configuration and implementation of each software solution, due to the specific computational requirements. Furthermore, this ad hoc approach to software development has also hindered the full exploitation of 3D data. As such, in real world applications a 3D mapping mission is often times still accompanied by in situ measurements (albeit reduced), which in some cases may increase the cost, complexity, and required expertise of the mission. This is naturally contrary to the promise of 3D remote sensing technologies of performing simpler measurements.

On the other hand, having several processing solutions that target the same output (e.g. DBH) is natural and welcomed, since different algorithms can be used to derive it and different datasets can be applied while developing the solution. However, a reasonable and fair comparison of the performance of such solutions is highly needed. From this perspective, it is crucial to establish publicly available benchmark datasets that comprise multi-sensor and multi-platform point clouds and accurate reference measurements of forest attributes from various forest ecosystems, optimally from all over the world. Furthermore, such datasets would be crucial for solution development, since they would foster the development of robust, sensor-agnostic and bias-free approaches. The use of a standard dataset for benchmarking purposes is already common practice in other domains, such as computer vision [113, 114] and 3D architecture [115].

Role of a Dynamic and Online Database

The web platform/online database established as a product of this survey is a step in the direction of knowledge consolidation in one place and a groundbreaking opportunity to provide the scientific community with a curated list of algorithms, supplemented by additional metadata. This resource will enable users to select the most suitable software for their needs, circumstances and output data, while simultaneously empowering software creators to avoid reinventing the wheel. By doing so, they can allocate their time and resources more efficiently, ultimately advancing the field of terrestrial point cloud processing and enhancing its accessibility and utility within the scientific community.

The compilation of the list and the database involved meticulous review and analysis of available documentation, as well as direct interaction with the software solutions during the STSMs. By consolidating this information, the database provides a comprehensive reference for researchers, practitioners and stakeholders interested in ground-based point cloud processing. It facilitates informed decision-making and enables comparison among different software solutions based on their capabilities and suitability for specific applications within forestry and related fields.

Conclusions and Outlook

In this paper, we described a review of state of the art point cloud processing for ground-based forest applications, and we presented a list of the available algorithms and software solutions. The aim of the list’s compilation was to collect the scattered information in one place, which we accomplished via the creation of an online searchable database. The paper thus also summarises the state of 3D technology in forestry. We then categorised the compiled list of 24 solutions. Most of the identified solutions are open-source or free, with an observed trend towards the general use of TLS technology. This is evidenced by the fact that while many of the solutions are sensor-agnostic, all of them take TLS data as their default input. Furthermore, a few tree parameters predominate as the computed output, in particular DBH. This may be interpreted as the high demand for such values in forestry applications and, by extension, the ever growing interest in using 3D technologies for forest applications. It is, however, an important caveat that variables such as DBH and tree height are some of the basic tree parameters; it is therefore only natural that solutions would aim to provide them, regardless of the general state of the use of 3D technology in forestry.

On the other hand, the development of software solutions is steadily progressing. Developers are creating software solutions based on the most recent challenges for point cloud processing that they encountered in their work as their principal functionality. However, there is increasing demand for software solutions which can not only carry out a single specific function but also help to assess basic forest inventory parameters with appropriate accuracy. Also, there should be a better solution for the computational requirement of the specific software or tools. To grow user groups and facilitate the use of existing tools by various user types, not just highly trained professionals, developers should focus on the user-friendliness and ease of application of their tools.

In the near future, a benchmarking of the identified solutions will be carried out to assess their geometric quality. This benchmarking is intended to provide future users of the web platform not only semantic information and metadata on the solutions, but also tangible values that determine the applicability of each solution according to the users’ needs. A standardisation of this nature is also envisaged for other aspects of ground-based 3D forest mapping, e.g. sensors and protocols, within the context of the 3DForEcoTech COST Action.