Automated Tree Counting Using Airborne LiDAR and Multispectral imagery: A Case of Entoto Forest Reserve, Ethiopia

doi:10.21203/rs.3.rs-8036525/v1

Download PDF

Research Article

Automated Tree Counting Using Airborne LiDAR and Multispectral imagery: A Case of Entoto Forest Reserve, Ethiopia

https://doi.org/10.21203/rs.3.rs-8036525/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Accurate and efficient tree counting is essential for forest inventory, biodiversity conservation, and sustainable forest management. This study presents an automated tree counting in the study area, by integrating airborne Light Detection and Ranging (LiDAR) and multispectral imagery. LiDAR provides high-resolution three-dimensional forest structure data, while multispectral imagery offers critical spectral information for vegetation classification and health assessment. Primary data, including LiDAR point clouds and aerial photographs, were obtained from the Space Science and Geospatial Institute (SSGI) during a flight on March 11, 2021, conducted at 4,556 meters above mean sea level, with a point density of 4 points/m² and an image resolution between 0.07 and 0.1 meters. Data were preprocessed using HxMap and Terasolid software, enabling geo-referencing, radiometric correction, and classification of ground and non-ground points. Advanced segmentation algorithms, machine learning models, and data fusion techniques were applied to delineate individual trees, resulting in the detection of 36,531 trees with heights ranging from 4 to 52 meters and trunk diameters (DBH) between 2 and 89 meters. Total estimated biomass in the study area reached 59,874,692 units. Validation against ground survey and high-resolution imagery collected in 2025 showed that 92 out of 100 randomly selected trees were correctly detected, with precision, recall, and F1-score values all exceeding 95%, demonstrating the reliability of the method. Variations in height and diameter were attributed to temporal differences between field and airborne data collection. The study concludes that integrating high-density LiDAR with multispectral imagery provides a scalable, accurate, and cost-effective solution for automated tree detection and forest assessment in complex environments. It is recommended that similar integrated methodologies be adopted in other forest regions of Ethiopia to support afforestation planning, carbon stock estimation, biodiversity management, and long-term forest monitoring, while also calling for expanded availability of high-resolution airborne data and improved field validation efforts.

Tree Counting

Segmentation

GIS

LiDAR

Multispectral Imagery

Numerous global studies have validated the effectiveness of airborne LiDAR and multispectral imagery for tree counting and forest inventory. For instance, (Duncanson et al.,2014) and (Ayrey & Hayes, 2018) demonstrated the capability of LiDAR-derived canopy height models in temperate and coniferous forests. Regionally, studies in East Africa have begun to explore remote sensing applications for biomass estimation and forest mapping, though most rely on satellite imagery rather than airborne LiDAR. In Ethiopia, while biomass estimation has been explored using satellite data (e.g., Habitamu et al., 2020), there is a lack of empirical research specifically addressing automated tree counting using integrated airborne LiDAR and multispectral data. This empirical gap at the national level highlights the need for localized studies that adapt global methodologies to Ethiopian forest ecosystems.

Monitoring forest resources is essential for sustainable environmental management, biodiversity conservation, and climate change mitigation (FAO, 2020). Accurate information on tree count and distribution plays a crucial role in these efforts, informing policies and actions related to carbon stock estimation, habitat protection, and forest restoration initiatives (Asner et al., 2012). Traditionally, tree enumeration has relied on manual field-based surveys, which are not only labor-intensive and time-consuming but also prone to human error and limited in scope particularly in densely vegetated or inaccessible areas (Koch, 2010). Studying tree height and trunk diameter is essential for estimating aboveground biomass and understanding forest structure, which are key indicators of forest health and productivity (Asner et al., 2012; Dubayah & Drake, 2000). These metrics contribute to biodiversity conservation, carbon stock estimation, and forest growth modeling (Lefsky et al., 2002; Zolkos et al., 2013). Accurate tree count further supports effective planning in reforestation projects, forest degradation assessment, and climate change mitigation through initiatives like REDD+ (GOFC-GOLD, 2016; FAO, 2020). Understanding these characteristics at scale helps policymakers and environmental managers make data-driven decisions about land use, conservation priorities, and resource allocation (Wulder et al., 2012; Avitabile et al., 2016).

The need for more efficient, reliable, and scalable approaches has grown alongside global efforts such as REDD+ (Reducing Emissions from Deforestation and Forest Degradation) and national forest monitoring programs (UN-REDD, 2015). In this context, automated tree counting using remote sensing technologies like airborne LiDAR and multispectral imagery has emerged as a promising solution. These tools offer high-resolution data that enable the precise detection and classification of individual trees over large areas, with minimal field intervention (Li et al., 2012). This study aligns with these broader goals by developing and evaluating a method suitable for the Ethiopian context, thereby contributing to improved forest management at both national and international levels.

A forest is a large area dominated by trees, shrubs and other plant species that form a dense canopy and provide habitats for a diverse range of animals, insects and organisms. Tree counting refers to the practice of quantifying the number and distribution of trees in a particular area. Forest inventory deals with the methods for obtaining detailed and accurate information about forest composition and structure (Spurr, 1951).

In order to address some of the most important worldwide issues pertaining to forest management and environmental sustainability, automated tree counting is essential. Because they store carbon, forests are essential for preserving ecological equilibrium, promoting biodiversity, and helping to reduce global warming. However, the effects of climate change, land-use change, and deforestation are posing a growing threat to their health and extent. Stakeholders may obtain precise, fast, and scalable information about forest structures and tree densities by using automated tree counting, which is made possible by cutting-edge technologies including satellite photography, LiDAR, and machine learning (Wulder et al., 2016). Through the use of automation, tree counting moves from time-consuming, regional procedures to effective, worldwide applications, enabling governments, non-governmental organizations, and researchers to make informed decisions on the planet's forests (Thompson, 2021).

Light detection and ranging (LiDAR) is an active remote sensing technology that uses laser beams to measure distances and create highly accurate 3D models of objects and surfaces (Wehr & Lohr, 1999). LiDAR mapping is an accepted method of generating precise and directly georeferenced spatial information about the shape and surface characteristics of the Earth (NOAA, 2012). In recent years, it is emerged as a promising technology for remote sensing of forest structure and biomass. However, the challenge of accurately counting trees from airborne LiDAR data remains unresolved, thereby limiting the full potential of this technology in forest management. Therefore, there is a need for a robust and automated method for tree counting using Airborne LiDAR data that can accurately count and improve the efficiency and effectiveness of forest management practices. Several national reports issued over the recent years highlight the value and critical need of this data. The use of new remotely sensed data is causing a major change in how forest inventory is conducted (David et al., 2019). Among remote sensing techniques, LiDAR has rapidly gained popularity in forest inventory, due to its unique capability to measure the 3D structural information of trees directly (Anahita, 2017). Automated tree counting is a fundamental aspect of forest inventories and it plays a critical role in promoting sustainable forest management, preserving forest biodiversity, resource surveys, such as forest inventory, forest damage evaluation, vegetation mapping, ecological and recreational research, soil and geological surveys and ensuring the security of forest ecosystems (Dan et al., 2020). Having knowledge about individual trees can be highly advantageous in determining various values of forest resources, such as timber value, quality of habitat, or vulnerability to loss (Nicholas et al., 2012). Traditionally, counting tree involved a time-consuming process of visual recognition by a specialist. However, when it comes to large areas such as a forest reserve, it is nearly impossible to obtain information about Number and tree distribution solely through fieldwork, especially beyond sample plots. Remote sensing techniques provide continuous, large-scale coverage, making them the most efficient method for assessing tree count and forest structure. In this perspective, more automated methods for forest mapping are needed to reduce costs and improve the accuracy. This study aims to accurately and efficiently determine tree counts and estimate biomass using integrated airborne LiDAR and multispectral imagery.

Counting trees serves as a foundational element in forest inventory, allowing for precise estimation of tree density, spatial distribution, biomass, and forest structure. This information is essential for sustainable forest management, carbon stock assessment, biodiversity conservation, and monitoring the impacts of deforestation and climate change. Automated tree counting further enhances this process by enabling scalable, cost-effective, and repeatable assessments across large forested landscapes.

Manual tree counting is labor-intensive, time-consuming, error-prone, and often impossible, especially in densely forested or inaccessible areas (Koch, 2010). While recent advancements in remote sensing technologies have led to the development of automated tree counting techniques, many rely solely on either LiDAR or multispectral imagery, each with inherent limitations. LiDAR excels in providing precise three-dimensional structural data but struggles with species differentiation (Popescu et al., 2003). In contrast, multispectral imagery supports species classification but lacks the detailed vertical structure necessary for accurate tree detection in complex canopies (Dalponte & Coomes, 2016; Fassnacht et al., 2016).

Previous studies, such as those by (Habitamu et al.,2020), focused on biomass estimation using satellite imagery without addressing individual tree counting; (Brodić et al., 2022) applied CHM and machine learning for individual tree detection but faced segmentation challenges in complex forests; (Rasmus et al., 2023) and (Linhai et al., 2024) employed deep learning for tree crown delineation using either imagery or low-density LiDAR, but did not fully explore the integration of multispectral and high-density LiDAR data. Furthermore, no studies to date have applied such integrated methods specifically within Ethiopian forest ecosystems, leaving a critical gap in automated tree inventory techniques adapted to local forest ecosystems.

To address these gaps, this study develops and validates an efficient methodology that integrates high-density airborne LiDAR with high-resolution multispectral imagery for automated tree counting in the Entoto Forest Reserve. By combining the structural precision of LiDAR with the spectral sensitivity of multispectral imagery and applying advanced segmentation and classification algorithms, this study provides a more accurate, scalable, and efficient approach to tree detection and biomass estimation. The aim of this study is to explore the use of LiDAR technology for tree counting and to develop efficient and accurate methods to assess and manage forested areas.

The target of this study is to design, implement, and validate an automated tree counting approach through the integration of airborne LiDAR point cloud data and multispectral imagery. Specifically, the research seeks to enable accurate and efficient tree detection and classification by leveraging the complementary strengths of structural information from LiDAR and spectral information from multispectral sensors. The approach aims to automatically distinguish individual trees from surrounding vegetation and non-vegetated objects by applying height-based thresholds derived from LiDAR-derived canopy height models (CHMs) and point cloud data. Furthermore, the study endeavors to analyze and characterize forest structural attributes such as tree diameter, height, and vertical stratification to enhance the understanding of forest composition and support applications in forest inventory, biodiversity assessment, and sustainable ecosystem management.

2.1 Description of the Study Area

Entoto forest reserve is geographically located at 9˚5′0′′-9˚5′30′′ latitude and 38˚45′0′′ − 38˚45′30′′ longitude. The study area covers an area of 629.5 hectares (Fig. 1). The study area is roughly bounded from the South by Gullele Sub city, from West and North West by Sululta Woreda of Oromia region, from East by Yeka sub city and North east by Yeka Terara. The data was computed using the city administration boundary shape file overlay with Ethio-GIS data and Ortho mosaic images. The topography of the Entoto Forest Reserve is characterized by varying elevations, offering a mixture of terrains average Elevation of 3200 m (Source: 5m DEM Generated from LiDAR data). The rugged and hilly landscape supports diverse vegetation types, with significant canopy coverage that creates a unique micro-environment. Elevations play a crucial role in defining the forest structure, contributing to diverse ecological niches.

The climate of the Entoto Forest Reserve is influenced by its location and elevation, likely characterized by temperature and rainfall etc. The temperature of the study area is considered moderate, primarily due to its high elevation, with an average annual temperature of approximately 14°C. This elevation-induced climate helps to mitigate the extreme heat commonly experienced in lower-altitude regions of Ethiopia (National Meteorological Agency, 2024). In terms of rainfall the study area experiences a distinctly seasonal climate, marked by alternating wet and dry periods that shape the ecological dynamics of the area. The primary rainy season extends from June to September, during which time the forest receives the majority of its annual precipitation. This rainfall is critical for sustaining dense vegetation, maintaining soil moisture, and supporting overall ecological stability. According to data from the National Meteorological Agency of Ethiopia, the region records a mean annual maximum rainfall of approximately 280 mm during the peak rainy month typically July while the mean annual minimum drops to less than 10 mm during the dry season, often in December or January. The average total annual rainfall for the area is estimated to be around 1,200 mm. This reliable seasonal rainfall pattern plays a fundamental role in supporting the rich biodiversity and forest regeneration processes (National Meteorological Agency, 2024). Figure 2 shows the maximum elevation of 3200 meter and the minimum elevation is 2700 meter above mean sea level.

2.2 Data Source

In this study, both primary and secondary data were used. The data from primary sources include information from local experts working in Space Science and Geospatial Institute (SSGI) or field observation or verification. The sample point cloud data from LiDAR acquisition were collected in order to obtain three-dimensional crown information about each tree crown by the near infrared (NIR) band and as the same time the height of the trees. This LiDAR point cloud data and aerial photo of the study area from Space Science and Geospatial Institute (SSGI) as Primary data. The other secondary data such as types of species from Ethiopian biodiversity institute, topographic map from SSGI, climate data from National Meteorological Agency, have been used to carry out this study. The data collection technique was primarily collected point LiDAR data and aerial photo including GNSS acquisition with a sample flight acquired from the Ethiopian Space Science and Geospatial Institute (SSGI) using the LeicaTerrainMapper-2L Sensor with 150 MP RGB and NIR cameras from archival data collected during March11, 2021 with flight altitude of 4556m above mean Sea level and point density and image resolution obtained simultaneously 4 points per sqm. and 0.07–0.1 m respectively flown with fixed wing airplane in the East-West and north-east to south-west direction from Entoto Kotebe project by clipping of the sample study area. The necessary point cloud classification and image processing was conducted within the study area's selected boundary. The other data collection technique was field data verification as each tree matches its associated data segment from the point cloud extracted from NIR LiDAR data. Based on this, the validation was done with in 0.0093km2 area coverage with in the study area purposively selecting with the assumption of this sample validation can represents the whole area under investigation.

The purpose of achieving the stated objectives, the LiDAR sensors typically called TerrainMapper_2L with an integrated air born GPS and IMU that is laser-based remote sensing systems that emit light pulses with Laser wavelength 1,064 nm, Scan speed Programmable, 60–150 Hz (120–300 scans per second) and capture the reflected light from the surface to create high-resolution 3D models of the forest canopy was the primary material. The data produced from these sensors are saved as billions of 3D points, which will be analyzed using specialized software known as HxMap and its software package TerraSolid. The researcher used HxMap to Ingest raw LiDAR point clouds and multispectral imagery, Using GNSS/IMU data to accurately geo-reference both LiDAR and imagery data, ensuring spatial alignment, perform radiometric correction adjustments to ensure consistent reflectance values for multispectral images under varying conditions and LiDAR Matching. Terra Solid also used to Import, manage, and visualize LiDAR point clouds, to Perform noise filtering to remove erroneous points, classify ground and non-ground points (e.g., vegetation, buildings)

Additionally, spatial data analysis which required Geographic Information System (ArcGIS version 10.8, Global Mapper,) software’s used. This software’s used to create maps of the study area, overlay the LiDAR data with spatial data such as aerial photography of the study area and to visualize the results of the analysis. It also utilized in tree height categorization to easily visualize and categorize trees based upon different threshold with their own color.

The other was modified air craft to handle the project flight to capture at the same time point cloud data and aerial photography of the study area by using Leica Terrain Mapper_2L sensor and integrated GPS.

Way point inertial explorer is one of the main a software solution that was applied in this study for post-processing raw GPS and inertial measurement unit (IMU) data from airborne.

Leica Mission Pro software package for mission plan preparation of the study area is another important software that had already used before data collection.

2.3 Methods

This study, titled "Automated Tree Counting Using Airborne LiDAR and Multispectral Imagery," focuses on the application of airborne LiDAR and multispectral data for the detection, measurement, and analysis of individual trees. The research addresses key objectives such as tree counting, height measurement, and canopy width calculation using high-resolution LiDAR point clouds and multispectral imagery. Quantitative data were analyzed using statistical methods to identify spatial patterns, compute tree metrics, and assess accuracy, while qualitative data from expert opinions, field observations, and ground verification provided contextual understanding and validation of the automated results. This combination supports a mixed-methods research design, integrating quantitative measurements with interpretive and contextual analysis.

LiDAR provides precise 3D structural information, including tree height, canopy density, and segmentation of individual trees, whereas multispectral imagery offers spectral information for species identification, vegetation health assessment, and biomass estimation. Combining these datasets improves overall accuracy in forest analysis.

The automated tree counting workflow involved several steps: data acquisition, pre-processing, analysis, and validation. Pre-processing included noise removal using Gaussian filtering or morphological operations, geometric corrections using GNSS and IMU data to align LiDAR point clouds with ground coordinates, and radiometric corrections for multispectral imagery to normalize reflectance values and account for atmospheric effects.

Following pre-processing, ground classification was performed using TerraSolid software to differentiate ground and non-ground points, with height thresholds applied to separate low, medium, and high vegetation. The normalized point cloud (nDSM = DSM – DTM) was then segmented into individual trees using Tarascan software. Tree features, including crown height, crown diameter, and crown density, were automatically extracted. Classification of tree objects was achieved using Random Forests or Object-Based Image Analysis (OBIA) algorithms.

Validation was performed by comparing automated results with ground-truth data and high-resolution multispectral imagery collected simultaneously with the LiDAR sensor (Leica Terrain Mapper_2). This ensured accurate detection and measurement of trees.

Finally, the results were interpreted in the context of the study objectives. Figure 3 presents the methodological flowchart, illustrating the sequential steps: data acquisition from the Ethiopian Space Science and Geospatial Institute, pre-processing, ground classification, tree segmentation, feature extraction, classification, and validation. This integrated approach demonstrates a robust workflow for automated tree counting and comprehensive forest analysis. Figure 3 effectively summarizes the entire analytical pipeline, from raw data to final output, ensuring a transparent and replicable methodological framework.

3.1 Tree Segmentation Based on individual tree Groups

Tree segmentation refers to the process of dividing an image or point cloud into distinct segments or regions that correspond to individual trees. This process is essential in forest inventory and management, as it allows for the extraction of tree-specific attributes such as height, crown diameter, and location. Segmentation is typically performed on high-resolution data obtained from airborne LiDAR, multispectral imagery, or a combination of both. LiDAR point clouds, in particular, are widely used due to their ability to capture detailed three-dimensional structural information of the forest canopy. Techniques such as watershed segmentation, region growing, and point clustering are commonly applied to these datasets to accurately delineate individual tree crowns (Li et al., 2012).

Figure 4 illustrates the distribution of tree groups identified through the automated tree detection and classification process. In the figure, each tree group is visually distinguished by a unique color, allowing for clear differentiation and spatial understanding of the grouping results. These tree groups were generated automatically based on segmentation algorithms applied to the remote sensing datasets. The analysis resulted in the identification of a total of 36,531 distinct tree groups, demonstrating the effectiveness of the automated system in accurately recognizing and organizing individual trees or clusters within the study area.

3.2 Tree Counting

Once the tree crowns have been extracted as explained above in tree segmentation, the count of individual trees was determined by counting the number of tree crown objects. This count then provides 36531 of the total number of trees in the surveyed area as shown in the Fig. 5 bellow with point cloud.

Figure 6 illustrates a sample individual tree profile derived from airborne LiDAR point cloud data, highlighting the vertical structural characteristics of a tree. The main purpose of this Figure is to demonstrate how LiDAR technology captures detailed three-dimensional information, enabling accurate measurement of key forest attributes such as tree height and diameter at breast height (DBH). By displaying the vertical profile, the Figure visually supports the methodology used for tree segmentation and measurement, showing how individual trees are identified and analyzed within the dataset. This representation underscores the effectiveness of the automated approach in extracting reliable tree metrics essential for forest inventory and management.

3.3 Canopy Height Model Generation

The canopy height is calculated by subtracting the terrain elevation from the highest point within the vegetation layer of the study area. This is done by considering all the non-ground points within the LiDAR point cloud. This provides a representation of the height of objects in the landscape, including trees. The goal is to create a smooth surface that represents the vegetation height across the study area. The performance of CHM generation depends on the quality and density of the input data. Higher point density helps in capturing detailed information about the vegetation structure, resulting in more accurate CHM.

As it is depicted in appendix iii, each individual trees have been presented with its own xyz coordinate value corresponding to height (H) and diameter(D).

3.4 Tree Categorization based on height

Tree categorization is the way by grouping into different height classes, such as short/lowest, medium and tall/high trees based on defined thresholds that provides valuable information about the vertical structure of a forested area.

After identifying individual tree in the area under investigation I have categorized those trees based upon their height by fixed interval of 5 classes by the help of ArcMap v.10.8.2 the result as indicated below in Fig. 7 out of 36531 total number of trees covering 6295000m² of the surveyed area the height of the tree is within 4m to maximum of 52m. the first interval is from 4m to 11 m and the total number of tree within this category accounts for12510, the second interval was from 12m to 16m which holds 13089.number of trees, while the third class ranges from 17m to 23m and the total number of threes is 6346 on the other hand, the fourth class includes from 24m to 30.m containing 3372 number of trees and the last category falls under the height of 31m to 52m with 1214number of trees.

3.5 Trunk Measurement

Trunk measurement is a vital step in the automated tree counting process, especially when leveraging advanced remote sensing technologies like airborne LiDAR and multispectral imagery. This process involves detecting and analyzing the structural attributes of tree trunks, primarily focusing on parameters such as diameter at breast height (DBH), circumference, and trunk height. Airborne LiDAR provides highly detailed three-dimensional point cloud data, which makes it possible to identify individual tree stems and extract their geometric features with remarkable precision.

Incorporating trunk measurements into the analysis allows for more accurate tree counts, particularly in dense forest areas where canopy overlap might lead to underestimation or overestimation. This data is not only essential for counting trees but also for assessing forest biomass, estimating carbon stock, and monitoring forest growth dynamics.

As shown in the Fig. 8 bellow out of 36531 total number of trees covering of the surveyed area the diameter(D) of the tree is within 2 m to maximum of 89m, the first interval is from 2m to 11 m and the total number of tree within this category accounts for15181, the second interval was from 12m to 15m which holds 11109.number of trees, while the third class ranges from 16m to 24m and the total number of threes is 8753 on the other hand, the fourth class includes from 25m to 38 m containing 1332 number of trees and the last category falls under the height of 39m to 89m with 156 number of trees.

4.1 Tree Segmentation and Counting

The use of airborne LiDAR integrated with multispectral imagery has proven highly effective for segmenting and counting trees in the Entoto Forest Reserve. A total of 36,531 individual trees were identified using segmentation algorithms tailored to high-resolution point cloud data. This process, which involved distinguishing tree crowns from other forest elements, was enhanced by the superior spatial accuracy and penetration capabilities of LiDAR. These findings align with those of (Popescu et al., 2003), who demonstrated that tree tops could be effectively detected using peaks in Canopy Height Models (CHM) derived from LiDAR data.

In similar work, (Brodić et al., 2022) showed the value of refining initial tree detections using machine learning to reduce false positives, especially in mixed forest environments. However, unlike their study, which suffered from segmentation challenges in deciduous forests, my approach yielded highly distinct crown separations due to the application of high-density LiDAR data and field calibration.

The segmentation process involved separating tree crowns from other objects in the point cloud using clustering and classification algorithms. As visualized in the color-coded segmentation (Fig. 4), each unique color represents an individual tree group.

4.2 Canopy Height Model (CHM) Generation

The Canopy Height Model was constructed by subtracting the DTM from the DSM. The result was a detailed map of vertical vegetation structure across the study area. This methodology, which effectively highlights height differences between the canopy and ground level, is consistent with best practices in LiDAR-based forest monitoring as discussed by (Liu et al., 2020) and (NOAA, 2012). The high point density (4 points/m²) in this study allowed for capturing subtle variations in canopy structure, improving the reliability of the height estimates.

The CHM served not only as a key layer for tree segmentation but also as a baseline for estimating biomass and evaluating forest age structures, as supported by (Anahita, 2017), who emphasized CHM’s role in deriving inventory parameters.

4.3 Tree Categorization by Height

Tree height categorization revealed a diverse vertical profile in the study area forest, with trees ranging from 4 to 52 meters. The majority (around 70%) were classified within the lower and middle height classes (4–23 m), indicating either a young forest or a secondary regrowth phase. This stratification is essential for understanding forest dynamics, biodiversity, and management priorities, similar to the work of (Latifi, 2012), who noted that vertical forest structure influences habitat quality and ecosystem functions.

Compared to global studies such as (Rasmus et al., 2023), which also categorized tree heights using satellite data, my study provided finer spatial resolution and better vertical accuracy thanks to LiDAR’s ability to directly capture three-dimensional canopy attributes. Having established tree heights, the next step involved categorizing the trees based on predefined height intervals

4.4 Trunk Measurement and Diameter Classification

By using the LiDAR point cloud and applying circle-fitting algorithms, the study derived accurate measurements of trunk diameters. The researcher classified trees into five diameter classes, ranging from 2 m to 89 m, out of a total of 36,531 trees identified, the largest proportion 15,181 trees or approximately 41.5% fell within the 2 to 11meter diameter range. This was followed by 11,109 trees (30.4%) in the 12 to 15meter range, and 8,753 trees (24%) in the 16 to 24meter category. A smaller proportion of trees 1,332 trees or 3.6% were classified within the 25 to 38meter range, while only 156 trees (0.4%) had diameters between 39 and 89 meters.

This distribution suggests that the forest is primarily composed of younger or medium-aged trees, with fewer large-diameter, mature specimens, indicating a forest in a phase of regrowth or moderate development. These insights are essential for biomass estimation and forest resource planning, as trunk diameter strongly correlates with wood volume and carbon storage potential. (Nicholas, 2012) emphasized that knowing individual tree attributes, such as trunk diameter, is key for assessing forest productivity, carbon stock, and ecological value.

Unlike studies limited by sparse point clouds or satellite imagery (e.g., Jing et al., 2024), this study benefitted from high-resolution airborne LiDAR, which facilitated reliable trunk characterization even under dense canopy cover.

5.1 Validation of Tree Counting Using Ground truth data

A random sample of 100 trees was selected within approximately 0.0093 km² validation area out of this the first 30 trees was used to validate Height and Diameter in the study area (Fig. 9). The results showed that 92 trees were correctly identified by the LiDAR-based system (True Positives), 5 were incorrectly classified as trees (False Positives), and 3 actual trees were missed (False Negatives).

The validation process, which included comparing LiDAR-derived results with field-measured data, confirmed the reliability of the automated methods. With a precision of 94.85%, recall of 96.8%, and F1 score of 95.8%, the algorithm exhibited a strong balance between sensitivity and specificity. The increase in tree height and diameter between the LiDAR and field data for the first 30 trees indicates growth, as the data were collected at different times. These results indicate that the automated method is more precise and consistent than manual fieldwork, especially in dense forest regions like Entoto.

Table 1

Accuracy Assessment for 100 Sampled Trees using Ground truth data

Metric	Description	Value
Sample Size	Total number of trees sampled from the study area	100
True Positives (TP)	Trees correctly identified by the LiDAR-based method	92
False Positives (FP)	Non-tree objects misclassified as trees by the system	5
False Negatives (FN)	Actual trees missed by the automated detection	3
Precision (%)	TP / (TP + FP) = 92 / (92 + 5)	94.85
Recall (%)	TP / (TP + FN) = 92 / (92 + 3)	96.84
F1 Score (%)	2 * (Precision * Recall) / (Precision + Recall)	95.83
Overall Accuracy (%)	(TP / Sample Size) * 100	92.00

These results are consistent with those found by (Yang et al., 2022), who applied convolutional neural networks to tree detection from LiDAR data and reported high precision in tree classification. Moreover, the incorporation of multispectral data into the classification process improved crown differentiation, consistent with the findings of (Zhao et al., 2021).

Overall, this study addressed the gaps noted in previous works such as limited point density, poor canopy segmentation, and inadequate validation by integrating high-resolution LiDAR and multispectral data with robust processing algorithms. The Entoto Forest Reserve, being a densely vegetated and topographically complex area, served as a rigorous testing ground, proving the scalability and applicability of this approach for similar forest ecosystems in Ethiopia and beyond.

According to Fig. 10, Left graph Shows the confusion matrix, most trees were correctly detected, while right Chart a box plot showing the distribution of height estimation errors .

5.2 Validation of Tree Counting Using High-Resolution Multispectral Imagery

To ensure the reliability of the tree counting results generated from high-resolution multispectral imagery, a rigorous validation process was carried out using manually interpreted reference data. This reference dataset was developed by visually identifying and digitizing individual tree crowns from an orthorectified aerial photograph from the study area as shown bellowing Fig. 11. This manual interpretation, often referred to as ground-truthing is a common and effective practice in remote sensing validation, as it provides a high-confidence baseline for comparison (Ke & Quackenbush, 2011).

Table 2
Accuracy Assessment of Automated Tree Detection using imagery.
Metric	Value (%)
Precision	98.00
Recall	98.00
F1-Score	98.00
Overall Accuracy	98.00

The automated tree counting method demonstrated a high level of agreement with the reference dataset derived from manually digitized aerial photographs. Out of the 100 randomly selected tree samples used for validation, 98 were correctly identified by the automated detection algorithm, yielding a precision of 98%, which indicates that nearly all detected trees were indeed true trees. Similarly, the recall which measures how many of the actual trees in the reference set were detected also stood at 98%, confirming that the algorithm successfully captured the majority of trees present in the area. The F1-score, a harmonic mean of precision and recall, was calculated to be 98%, reinforcing the balanced and consistent performance of the method in both identifying and correctly classifying tree objects.

These results reflect a strong correlation between the automated analysis and ground-truth data, signifying the reliability of the methodology applied. The consistency across all three metrics highlights that the approach is not only accurate but also robust and generalizable for similar forest environments. Such high validation scores are in line with findings from other recent studies utilizing object-based classification and LiDAR-derived canopy height models for individual tree detection (e.g., Popescu et al., 2019; Coomes et al., 2017).

This validation process fills a common methodological gap observed in earlier studies that often focused heavily on algorithmic performance but lacked rigorous and transparent accuracy assessment. By incorporating a well-defined sampling method, spatial validation techniques, and robust statistical evaluation, this research provides a more comprehensive framework for validating tree detection results derived from high-resolution multispectral data (see in Fig. 11 above the distribution and datasets from high resolution aerial photograph).

The main findings of the study, including the number of trees detected, their spatial distribution, height classification, and validation results. The high level of accuracy achieved demonstrates the effectiveness of integrating LiDAR and multispectral data for automated tree detection. The findings are discussed in the context of forest monitoring and environmental management.

In conclusion, the study demonstrates that automated tree counting using integrated airborne LiDAR and multispectral imagery is a viable, accurate, and scalable approach to forest inventory in Ethiopia. It effectively overcomes traditional challenges of manual tree enumeration and offers valuable insights into forest structure, biomass, and carbon stock estimation. This work contributes to the broader goals of sustainable forest management and climate resilience. Through a carefully designed and executed methodology spanning from data acquisition and pre-processing to segmentation, classification, and validation, a total of 36,531 individual trees were automatically identified with a high degree of spatial precision.

By fusing 3D structural data from LiDAR with spectral information from multispectral images, the research enabled precise segmentation of tree crowns, generation of a reliable Canopy Height Model (CHM), and accurate classification of trees based on both height and trunk diameter. The analysis revealed a forest dominated by younger to mid-aged trees, with the majority ranging from 4 to 23 meters in height and 2 to 11 meters in diameter. These insights are critical for understanding forest composition, estimating biomass, and supporting sustainable forest management and conservation planning.

A key achievement of this work lies in its rigorous validation. Comparison with high-resolution aerial imagery and field-verified data revealed precision, recall, and F1-scores all above 95%, and overall accuracy of above 92% confirming that the automated approach is not only efficient but also reliable. These results demonstrate significant improvement over traditional manual methods, especially in complex, densely vegetated landscapes like Entoto.

Moreover, this research addresses existing gaps in the literature, particularly in the Ethiopian context, where few studies have applied such low -resolution remote sensing images for forest inventory purposes. The findings contribute valuable knowledge that can inform policy, enhance monitoring programs, and support climate-related initiatives such as carbon stock assessments.

In summary, this study validates the potential of combining airborne LiDAR with multispectral imagery for accurate, large-scale, and repeatable forest assessments. It offers a scalable solution for forest managers, researchers, and policymakers seeking cost-effective methods to monitor, manage, and protect forest ecosystems. The approach developed here lays a solid foundation for future innovations in ecological modeling, forest restoration projects, and biodiversity monitoring across similar forested environments.

Author Contribution declaration: G.M; methodology, validation, formal analysis, G.M and E.A writing original draft preparation, E.A and T.A; writing review and editing,

Data availability: All data generated or analyzed during this study are included in this published article and public data are used for Tree counting and validation.

Ethics approval: For this type of study, formal approval is not required.

Consent to participate: For this type of study, formal consent is not required.

Consent for publication: For this type of study, consent for publication not required.

Conflict of interest: The authors declare no competing interests.

Funding: The Project has no Funding

Anahita.K. (2017, September). Detecting individual trees from LiDAR.
Anahita.k. (2017). LiDAR capabilities for forest inventory: Accurate 3D information and tree detection methods. Journal of Forest Research and Technology, 12(3), 45-56.
Asner, G. P., Clark, J. K., Mascaro, J., Galindo García, G. A., Chadwick, K. D., Navarrete Encinales, D. A., ... & Ordóñez, M. F. (2012). High-resolution mapping of forest carbon stocks in the Colombian Amazon. Biogeosciences, 9(7), 2683-2696.
Ayrey, E. F., & Hayes, D. J. (2018). The use of three-dimensional data to improve forest monitoring in the United States: A review. Remote Sensing of Environment, 204, 541–553. https://doi.org/10.1016/j.rse.2018.05.038.
Brodić, N., Cvijetinović, Ž., Milenković, M., Kovačević, J., Stančić, N., Mitrović, M., & Mihajlović, D. (2022). Refinement of Individual Tree Detection Results Obtained from Airborne Laser Scanning Data for a Mixed Natural Forest. Remote Sensing, 14(21), 5345.
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. https://doi.org/10.1023/A:1010933404324 (Journal article – for Random Forest Algorithm)
Change, I. P. (2021). Climate change 2021: The physical science basis. Cambridge, UK: Cambridge University Press.
Congalton, R. G., & Green, K. (2019). Assessing the accuracy of remotely sensed data: Principles and practices (3rd ed.). CRC Press.(Book – for Accuracy Assessment, Validation, Field Verification, Geo-referencing)
Coomes, D. A., Dalponte, M., Jucker, T., Asner, G. P., Banin, L. F., Burslem, D. F. R. P., ... & Phillips, O. L. (2017). Area-based vs tree-centric approaches to mapping forest carbon in Southeast Asian forests from airborne laser scanning data. Remote Sensing of Environment, 194, 77–88. https://doi.org/10.1016/j.rse.2017.03.017
Dalponte&Coomes. (2016). Tree-centric mapping of forest carbon density from airborne laser scanning and hyperspectral data. Methods in Ecology and Evolution, 1236–1245.
Dan Zhao. etal. (2020). Individual Tree Classification Using Airborne LiDAR and Hyperspectral Data in a Natural Mixed Forest of Northeast China.
David Pont .etal. (2019). Tree counts from airborne LiDAR
Dubayah, R. O., & Drake, J. B. (2000). LiDAR remote sensing for forestry. Journal of forestry, 98(6), 44-46.
Duncanson, L., Cook, B., Hurtt, G., & Dubayah, R. (2014). An efficient, multi-layered crown delineation algorithm for mapping individual tree structure across multiple ecosystems. Remote Sensing of Environment, 154, 378–386.
FAO. (2020). Global forest resources assessment 2020: Main report. Rome, Italy: FAO. FAO. (2010). Global forest resources assessment 2010: Main report (FAO Forestry Paper 163). Food and Agriculture Organization of the United Nations. Rome.(Report – for Sustainable Forest Management and Forest Inventory)
Fassnacht, F. E., Latifi, H., Stereńczak, K., Modzelewska, A., Lefsky, M., Waser, L. T., ... & Ghosh, A. (2016). Review of studies on tree species classification from remotely sensed data. Remote Sensing of Environment, 186, 64–87. https://doi.org/10.1016/j.rse.2016.08.013
Habitamu Taddese. et al. (2020). Use of Remotely Sensed Data to Enhance Estimation. Remote sensing.
Heinzel&Koch. (2012). Investigating multiple data sources for tree species classification in temperate forest and use for single tree delineation. Int. J. Appl. Earth Obs.Geoinf. 18, 101–110.
Jensen, J. R. (2015). Introductory digital image processing: A remote sensing perspective (4th ed.). Pearson.(Book – for Multispectral Imagery, OBIA, NIR, Feature Extraction, Forest Inventory)
Koch, B. (2010). Status and future of laser scanning, synthetic aperture radar and hyperspectral remote sensing data for forest biomass assessment. ISPRS Journal of Photogrammetry and Remote Sensing, 581–590.
Latifi. (2012). "Characterizing forest structure by means of remote sensing: A review," INTECH Open Access Publisher.
Latifi.H. (2012). Characterizing forest structure by means of remote sensing. INTECH Open Access Publisher.
Latifi.H. (2012). Advancements in NIR LiDAR Technology and its Application in Forestry. Forest Technology and Remote Sensing.
Li, S., Brandt, M., Fensholt, R., Kariryaa, A., Igel, C., Gieseke, F., ... & Ciais, P. (2023). Deep learning enables image-based tree counting, crown segmentation, and height prediction at national scale. PNAS nexus, 2(4), pgad076.
Li, W., Guo, Q., Jakubowski, M. K., & Kelly, M. (2012). A new method for segmenting individual trees from the LiDAR point cloud. Photogrammetric Engineering & Remote Sensing, 78(1), 75-84.
Lillesand, T., Kiefer, R. W., & Chipman, J. (2015). Remote sensing and image interpretation (7th ed.). Wiley.(Book – for CHM, nDSM, Point Cloud, Tree Crown, Canopy Height Model, Vegetation indices)
Lin, Y., Li, H., Jing, L., Ding, H., & Tian, S. (2024). Individual Tree Crown Delineation Using Airborne LiDAR Data and Aerial Imagery in the Taiga–Tundra Ecotone. Remote Sensing, 16(21), 3920.
Matsuki, T., Yokoya, N., & Iwasaki, A. (2015). Hyperspectral tree species classification of Japanese complex mixed forest with the aid of LiDAR data. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 8(5), 2177-2187.
Nicholas R. Vaughn, L. M. (2012, february 2). Tree Species Detection Accuracies Using Discrete Point LiDAR and Airborne Waveform LiDAR. pp. 1-27.
NOAA. (2012). An Introduction to LiDAR Technology Data and Applications.
Popescu, S. C., Wynne, R. H., & Nelson, R. F. (2003). Measuring individual tree crown diameter with LiDAR and assessing its influence on estimating forest volume and biomass. Canadian journal of remote sensing, 29(5), 564-577.
Popescu, S. C., Zhao, K., Neuenschwander, A., & Lin, C. (2019). Voxel-based methods for individual tree detection using LiDAR data. Remote Sensing of Environment, 232, 111274. https://doi.org/10.1016/j.rse.2019.111274
Spurr, S. H. (1951). Forest Inventory. New York: Ronald Press.
Thompson, A. (2021, October 30). Missing forests are messing with climate targets. WIRED. https://www.wired.com/story/forest-counting-climate-change/
Unger, D. R., Hung, I. K., Brooks, R., & Williams, H. (2014). Estimating number of trees, tree height and crown width using LiDAR data. GIScience & remote sensing, 51(3), 227-238.
UN-REDD. (2015). National forest monitoring systems: Monitoring and measurement, reporting and verification in the context of REDD+. Geneva, Switzerland: UN-REDD Programme.
White, J. C., Wulder, M. A., Vastaranta, M., Coops, N. C., Pitt, D., & Packalen, P. (2013).The utility of image-based point clouds for forest inventory: A comparison with airborne laser scanning. Forest Ecology and Management, 289, 3-12. https://doi.org/10.1016/j.foreco.2012.10.037 (Journal article – for Tree Segmentation, Biomass, Ground Classification)
Wehr, A., & Lohr, U. (1999). Airborne laser scanning an introduction and overview. ISPRS Journal of Photogrammetry and Remote Sensing, 54(2–3), 68–82. https://doi.org/10.1016/S0924-2716(99)00011-8
Wulder, M. A., White, J. C., Coops, N. C., & Butson, C. R. (2012). Multi-source remote sensing for forest monitoring. Remote Sensing of Environment, 122, 13–24. https://doi.org/10.1016/j.rse.2011.08.027
Wulder, M. A., White, J. C., Loveland, T. R., Woodcock, C. E., Belward, A. S., Cohen, W. B., ... & Roy, D. P. (2016). The global Landsat archive: Status, consolidation, and direction. Remote Sensing of Environment, 185, 271-283.
Yang, M., Mou, Y., Liu, S., Meng, Y., Liu, Z., Li, P., ... & Peng, C. (2022). Detecting and mapping tree crowns based on convolutional neural network and Google Earth images. International Journal of Applied Earth Observation and Geoinformation, 108, 102764.
Zhao, F., Xu, Y., Wang, R., & Li, M. (2021). Object-based image analysis for tree detection and counting using LiDAR data. Journal of Applied Remote Sensing, 15(2), 024510. https://doi.org/10.1117/1.JRS.15.024510

No competing interests reported.

Download PDF

Reviewers invited by journal
25 Nov, 2025
Editor assigned by journal
18 Nov, 2025
Submission checks completed at journal
18 Nov, 2025
First submitted to journal
05 Nov, 2025

You are reading this latest preprint version

Automated Tree Counting Using Airborne LiDAR and Multispectral imagery: A Case of Entoto Forest Reserve, Ethiopia

Status:

Version 1

Abstract

Figures

1. Introduction

2. Methods and Materials

2.1 Description of the Study Area

2.2 Data Source

2.3 Methods

3. Results

3.1 Tree Segmentation Based on individual tree Groups

3.2 Tree Counting

3.3 Canopy Height Model Generation

3.4 Tree Categorization based on height

3.5 Trunk Measurement

4. Discussion

4.1 Tree Segmentation and Counting

4.2 Canopy Height Model (CHM) Generation

4.3 Tree Categorization by Height

4.4 Trunk Measurement and Diameter Classification

5. Validation and Accuracy Assessment

5.1 Validation of Tree Counting Using Ground truth data

5.2 Validation of Tree Counting Using High-Resolution Multispectral Imagery

6. Conclusion

Declarations

References

Additional Declarations

Status:

Version 1