Seeing the Trees from Above: A Survey on Real and Synthetic Agroforestry Datasets for Remote Sensing Applications

Highlights

What are the main findings?

• Fields such as remote sensing, statistical methods, and deep learning rely on the availability of high-quality tree datasets for effective analysis.
• Publicly available high-resolution aerial tree datasets, highlighting their characteristics and alignment with FAIR principles.

What is the implication of the main findings?

• Identifying and evaluating these datasets can support future research in tree monitoring and management at scale.
• Both the availability of datasets and their alignment with FAIR principles are crucial for model development and the advancement of analytical methods in agroforestry.

Abstract

Trees are vital to both environmental health and human well-being. They purify the air we breathe, support biodiversity by providing habitats for wildlife, prevent soil erosion to maintain fertile land, and supply wood for construction, fuel, and a multitude of essential products such as fruits, to name a few. Therefore, it is important to monitor and preserve them to protect the natural environment for future generations and ensure the sustainability of our planet. Remote sensing is the rapidly advancing and powerful tool that enables us to monitor and manage trees and forests efficiently and at large scale. Statistical methods, machine learning, and more recently deep learning are essential for analyzing the vast amounts of data collected, making data the fundamental component of these methodologies. The advancement of these methods goes hand in hand with the availability of sample data; therefore, a review study on available high-resolution aerial datasets of trees can help pave the way for further development of analytical methods in this field. This study aims to shed light on publicly available datasets by conducting a systematic search and filter and an in-depth analysis of them, including their alignment with the FAIR—findable, accessible, interoperable, and reusable—principles and the latest trends concerning applications for such datasets.

1. Introduction

Trees are fundamental components of our terrestrial ecosystems, providing a multitude of ecological, economic, and social values. Ecologically, they function as significant carbon sinks, absorbing atmospheric carbon dioxide and thereby mitigating climate change impacts [1]. Trees also have a role in maintaining biodiversity, offering habitats for a vast array of plants and animals, and sustaining complex ecological interactions [2]. From an economic perspective, trees supply renewable resources, including timber, non-timber forest products, fruits, and medicinal compounds, which are used in various industries and local economies [3]. Socially, forests contribute to human well-being by providing recreational spaces and cultural significance [4]. Given these multitude of values, the conservation and management of trees and forest ecosystems and protecting them against natural or human-induced disasters such as wildfires are imperative for maintaining ecological balance and supporting human livelihoods.

Remote sensing has played an important role in forestry applications for several decades, starting from the early aerial photos for forest inventory assessment in the 1950s and 1960s [5] to the multimodal data collected by modern remote sensing platforms [6]. Furthermore, aerial remote sensing has transformed the way we monitor and manage our forests. It offers a bird’s-eye view of vast expanses of woodland that would be impractical to survey on foot. These data, captured from drones, piloted aircraft, or satellites, provide valuable insights into the health, composition, and dynamics of forests; the importance of aerial remote sensing data of trees [7] cannot be overstated. From identifying invasive species and monitoring tree growth to assessing the impact of natural disasters and climate change, aerial data equip forest managers and conservationists with the tools needed to make informed decisions. Moreover, with the integration of statistical methods and machine learning algorithms, these data can be analyzed at scale, automating the detection of forest disturbances, estimating tree biomass, and predicting future forest conditions. As we continue to face challenges such as deforestation and biodiversity loss, the role of data, particularly aerial data, becomes a crucial asset in our efforts to manage and protect our trees for generations to come. Data are also the foundational resource for developing new analytical methods and tools, making the availability of benchmark datasets vital for their continued advancement. Numerous studies have examined various aspects of forestry using remote sensing data such as forest diseases [8], tree species identification [9], and biomass estimation [10], to mention a few. There have also been studies that focus on reviewing available datasets in other scientific fields, such as agriculture [11], medicine [12], and cybersecurity [13]. However, to the best of our knowledge, there has been no comprehensive review study focusing on aerial data used in forestry and agroforestry with an emphasis on individual trees as the objects of interest. Agroforestry, which combines trees into agricultural landscapes to achieve environmental, economic, and social benefits, represents a critical application area where remote sensing data are particularly valuable [14]. While most datasets discussed in this paper are related to trees in general, some specifically focus on fruit trees, which are key components of agroforestry systems.

With the growing importance of data in forestry, and the significance of aerial remote sensing in forest management and conservation of trees, it becomes evident that a review article on tree-related datasets and repositories is not just necessary but essential. Such a review may serve as a resource for researchers and practitioners, offering a compilation of the most relevant and up-to-date sources of aerial data. This paper, with an emphasis on high-resolution data that allow the identification of individual trees, aims to provide insights into the diversity of available data, their alignment with the FAIR—findable, accessible, interoperable, and reusable—principles, and their potential applications, facilitating more efficient and informed decision-making in agroforestry-related works. Furthermore, by highlighting the strengths and limitations of existing repositories and datasets, this paper can help guide future data collection efforts and foster collaborations among the scientific community.

While many other important forestry resources exist, such as long-term ecological inventory networks (e.g., ForestGEO, CTFS, and forestplots.net) and dynamic vegetation models, these fall outside the scope of this review. Our emphasis is strictly on high-resolution aerial visual datasets (RGB, multispectral, hyperspectral, LiDAR, and thermal) that directly support individual-tree analysis and machine learning workflows. These complementary resources are highly important for ecological and carbon stock estimation research, but they were not included in our systematic search for the mentioned reasons.

In Section 2, the methodology of this study is presented followed by a quantitative assessment inspired by FAIR principles of published datasets. Section 3 shows the curation of our search, filter, and analysis of the sources. Lastly, Section 4 outlines this study’s main conclusions and remarks.

2. Methodology

In the context of initiating a review study on data, the first critical step involves defining the eligibility criteria. In this section, we will present our rationale for this matter.

As mentioned before, to the best of our knowledge, there has not been a review study on aerial data with individual trees being visible. Therefore, resolution becomes a crucial factor. While resolution may not directly impact the used methodologies in data analysis, it dictates the level of detail achievable in the analysis. In our study, we establish a criterion based on the visibility of individual trees in the data, resulting in high- and ultra-high resolution (UHR) data [15]:

• Ultra-High Resolution (UHR): 1 cm to 5 cm ground sampling distance (GSD).
  This level of detail allows individual leaves, small branches, and tree crowns to be visible, which is useful for forest monitoring tasks like precise species identification, disease detection, and structural analysis. Drones are typically capable of achieving this GSD due to their ability to fly at low altitudes with high-quality sensors.
• High Resolution: 5 cm to 30 cm GSD.
  This range is common for manned aircraft and some higher-altitude drone operations. Individual tree crowns are distinguishable, but smaller details within the canopy are often less clear.

It is important to note that these classifications are not universally standardized.

Data used in machine learning and deep learning approaches are typically pre-processed, labeled, and organized into a structured format compatible with the specific framework used. With the growing prominence of machine learning, especially deep learning methods in recent years, one might consider including only these labeled datasets. However, it is important to note that unannotated data are still frequently used in forestry applications, depending on the specific task at hand. Therefore, our study considers both annotated and not annotated data sources that we have identified.

We chose the Google Dataset Search engine as our primary data searching tool. However, using this engine for a direct data search is not the only way to find data; peer-reviewed scientific papers that acquired remote sensing data in their study can also be a source of data. We can categorize these papers into two groups. First are the ones that have acquired data and published it or have used already published data. We combine this group with the result of the direct data search from the Google Dataset Search and categorize them as published data. The second category is the papers that have acquired data through the course of their study but have not published them. Consequently, our search procedure contains two separate search queries that merge: one direct data search with the Google Dataset Search engine and the other with the Google Scholar engine for scientific articles. Figure 1 illustrates our search and filtering procedure, explained in the following.

To perform the search, a set of keywords was selected based on four aspects of the subject, namely, Context, Source, Sensor, and Purpose, resulting in the following search query: “(Tree OR Forest OR Forestry) AND (UAV OR “Unmanned Aerial Vehicle” OR UAS OR “Unmanned Aerial System” OR Aerial) AND (RGB OR Multispectral OR Hyperspectral OR LiDAR) AND (“Artificial Intelligence” OR “Machine Learning” OR “Deep Learning” OR Segmentation OR Classification)”.

Table 1. Keywords selection based on different aspects.

Aspect	Keywords
Context	Tree, Forest, Forestry
Source	UAV, Unmanned Aerial Vehicle, UAS, Unmanned Aerial System, Aerial
Sensor	RGB, Multispectral, Hyperspectral, LiDAR
Purpose	Artificial Intelligence, Machine Learning, Deep Learning, Segmentation, Classification

relates the selected keywords with the respective aspect.

The time period of the search was set to 2013 to 2025, and the search was closed on July 2025.

After acquiring the initial result on both search pipelines, we screened the outcome through a manual process by excluding low-resolution datasets, non-tree objects of interest, and resources outside the scope of aerial visual data (such as plot-based forest inventories, e.g., forestplots.net, ForestGEO, and dynamic vegetation simulation models). While these are highly valuable for ecological and carbon stock research, they do not provide the type of high-resolution visual data at the individual-tree level that we focus on in this review. After this exclusion, we further filtered for eligibility based on context derived from title, keywords, abstract, and used methodology. In the case of the published data, these sets are published in online database repositories (DBRs), which were identified and recorded. The 26 identified DBRs are presented in Section 3.3. The final number of total sources found is 4409 entries from the dataset search and 1609 entries from the paper search. The final number of identified eligible sources is 242, including 72 cases for published data and 170 cases for unpublished data. These numbers provide an overview of the identified eligible entities, showcasing the distribution across published data, unpublished data, and data repositories. For a visual representation, refer to Figure 2.

FAIRness Analysis

Furthermore, we adopted a scoring scheme based on the FAIR data principles [16] to evaluate the alignment of the found datasets with them.

The FAIR principles (findable, accessible, interoperable, and reusable) are guidelines for promoting the usability of datasets across diverse domains. They offer a robust framework that promotes enhanced data sharing, comprehension, and utilization. However, amidst the exponential growth in data production, a fundamental challenge persists: the lack of a universal scoring system for quantifying the FAIRness of datasets. This absence of a standardized metric poses a significant hurdle in the assessment and comparison of datasets. Currently, the evaluation of a dataset’s adherence to FAIR principles remains largely subjective, relegated to abstract qualitative judgments. Without a tangible scoring system, researchers, institutions, and industries face a daunting task: the inability to objectively measure, compare, or improve the FAIRness of their datasets. This critical gap in quantifying FAIRness not only obstructs efforts toward efficient data sharing and collaboration but also hinders the establishment of best practices for dataset management and utilization.

While acknowledging the absence of a widely accepted scoring system, this paper does not claim to introduce a universal solution. Rather, it proposes a scoring methodology as a step towards addressing this issue. It does not claim to provide a conclusive solution but an attempt to stimulate dialogue and collaborative efforts towards comprehensive and universally applicable quantification of FAIRness in data.

Definitions for the FAIR principles in this study were sourced from the foundational principles outlined by the Global Open Science and FAIR community, as articulated on the official website [17]. Our interpretation of these definitions and the logic of scoring the datasets are presented in

Table 2. Assessment of FAIR data principles: definitions [17] and interpretations.

Category		Definition	Scoring Logic
F1		(Meta) data are assigned globally unique and persistent identifiers	Score 0: Lacks a DOI or URL, a unique identification Score 0.5: Either a DOI or URL Score 1: Recognized and established source
F2		Data are described with rich metadata (defined by R1 below)	Score 0: When R1 is nonexistent Score 1: When R1 exhibits a non-zero value
F3		Metadata clearly and explicitly include the identifier of the data they describe	Score 0: Metadata lack a distinct reference to the identifier Score 1: Metadata explicitly include the identifier
F4		(Meta) data are registered or indexed in a searchable resource	Score 1: Assigned universally, acknowledging that dataset existence implies indexing
A1	A1.1	The protocol is open, free, and universally implementable	Score 1: Assigned universally due to accessibility via standard internet protocols
	A1.2	The protocol allows for an authentication and authorization procedure	Score 1: Recognizing protocol capability to support authentication and authorization procedures
A2		Metadata are accessible, even when the data are no longer available	Score 0: Uses URL, which may be unreliable Score 1: Uses DOI, ensuring persistent access
I1		(Meta) data use a formal, accessible, shared, and broadly applicable language	Score 0: No identifiable metadata Score 1: Metadata exist, any format (e.g., webpage)
I2		(Meta) data use vocabularies that follow FAIR principles	Score 0: No metadata or vocabularies Score 0.5: Metadata exist but lack standard vocabularies Score 1: Full use of accepted vocabularies
I3		(Meta) data include qualified references to other (meta) data	Score 0: No references to source datasets Score 1: Original work needing no references
R1	R1.1	(Meta) data are released with a clear and accessible data usage license	Score 0: No or unclear license Score 1: Clear and explicit license
	R1.2	(Meta) data are associated with detailed provenance	Score 0: Provenance unclear or missing Score 0.5: Vague/incomplete provenance Score 1: Clear and complete documentation
	R1.3	(Meta) data meet domain-relevant community standards	Score 0: Unclear or non-standard metadata Score 0.5: Weak metadata Score 1: Structured and descriptive metadata

.

We aggregated the assigned scores using the FAIR score proposed Formula (1), averaging them and rounding the numbers to two decimal points, resulting in values ranging from 0 to 4. The compiled results are presented in

Table 3. Aggregate FAIR assessment scores for datasets: general evaluation of data findability, accessibility, interoperability, and reusability.

Dataset	F				A			I			R			FAIR Score
	F1	F2	F3	F4	A1.1	A1.2	A2	I1	I2	I3	R1.1	R1.2	R1.3
[18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52] [53] [54] [55] [56] [57] [58] [59] [60] [61] [62]	1	1	1	1	1	1	1	1	1	1	1	1	1	4.00
[63]	0.5	1	1	1	1	1	1	1	1	1	1	1	1	3.88
[64]	1	1	1	1	1	1	1	1	1	0.5	1	1	1	3.83
[65] [66] [67]	1	1	0	1	1	1	1	1	1	1	1	1	1	3.75
[68]	0.5	1	1	1	1	1	1	1	0.5	1	1	1	1	3.70
[69] [70]	1	1	1	1	1	1	1	1	1	0	1	1	1	3.66
[71]	1	1	1	1	1	1	1	1	0.5	1	1	0	1	3.50
[72]	1	1	1	1	1	1	1	1	1	1	0	0	1	3.33
[73]	1	1	0	1	1	1	1	1	0	1	1	0.5	1	3.25
[74] [75] [76] [77]	0.5	1	0	1	1	1	0	1	1	1	1	1	1	3.13
[78]	0.5	1	0	1	1	1	1	1	0.5	1	0	1	1	3.13
[79]	1	1	0	1	1	1	1	1	0	0	1	1	1	3.08
[80]	0	1	1	1	1	1	1	1	0	1	1	1	0	3.08
[81]	1	1	0	1	1	1	1	1	0	1	1	0.5	0	2.92
[82]	0.5	1	1	1	1	1	1	1	0.5	0	1	0	0.5	2.87
[83]	0.5	1	1	1	1	1	1	1	0.5	0	1	0	0.5	2.87
[84]	1	1	0	0	1	1	1	1	0	1	1	0	1	2.83
[85]	0	1	1	1	1	1	1	1	0	1	0	0.5	0	2.58
[86]	0	1	0	1	1	1	1	1	0	0.5	0	1	0	2.33
[87]	0.5	1	1	0	1	1	1	0	0.5	1	1	0.5	0.5	2.29
[88]	0	1	0	1	1	1	1	1	0	0	0	1	0	2.16
[89]	0.5	1	0	1	1	1	0	0	0.5	1	1	0.5	0.5	2.04

.

$$\mathrm{FAIR}\mathrm{score}={\displaystyle \frac{F1+F2+F3+F4}{4}}+{\displaystyle \frac{\frac{A1.1+A1.2}{2}+A2}{2}}+{\displaystyle \frac{I1+I2+I3}{3}}+{\displaystyle \frac{R1.1+R1.2+R1.3}{3}}$$

(1)

3. Datasets

Among the two categories mentioned, published and unpublished data, the first one stands as the foremost and pivotal source. Unpublished data, though potentially valuable and accessible, remain more prospective than realized in their impact. We start by reviewing the published datasets found.

3.1. Published Datasets

Table 4. Published datasets. (*) CA, GE, N/A, and PA, respectively, stand for Corresponding Author, Geographical Extent, Not Available, and Piloted Aircraft.

Dataset/ Study	Application	Source	Platform	Dataset Size (GB)	Annotation	Open Access	CA (*) Country	GE (*) Country
[43]	ITCD, SI, GM	RGB, LiDAR, PPC	UAV	742.7	Polygons	Yes	Canada	Canada
[83]	ITCD	RGB	UAV	0.32	Polygons	Yes	N/A	N/A
[24]	ITCD	RGB	UAV	10	Semantic labels	Yes	Brazil	Brazil
[82]	ITCD	RGB	UAV	0.15	Semantic labels	Yes	N/A	N/A
[89]	ITCD, SI	RGB, HSI, PPC, LiDAR	UAV	N/A	Polygons, Semantic labels	Yes	USA	China
[31]	ITCD, SS	LiDAR	UAV	1.6	Semantic labels	Yes	Norway	Europe/Australia
[37]	ITCD	RGB	UAV	11	Polygons	Yes	Germany	Colombia
[85]	ITCD	MSI	Satellite	Undisclosed	Semantic labels	No	Australia	Australia
[88]	ITCD	RGB	Satellite	Undisclosed	Semantic labels	Yes	China	USA
[41]	ITCD	RGB, LiDAR	UAV	0.18	Semantic labels	No	China	N/A
[44]	ITCD	RGB	UAV	24.6	Polygons, Bounding boxes	Yes	Canada	Canada
[45]	ITCD	RGB	PA, Satellite	16	Polygons, Bounding boxes	Yes	Germany	Germany
[46]	ITCD	RGB	UAV, PA	4	Polygons	Yes	Switzerland	Global
[47]	ITD, SS, SI	RGB	PA	8.1	Polygons, Bounding boxes	Yes	Switzerland	Switzerland
[48]	ITD, SI	RGB	UAV	150	Polygons, Bounding boxes	Yes	China	Canada
[49]	ITD, GM	RGB	UAV	2.48	Keypoints	Yes	USA	USA
[50]	ITD, Misc.	MSI, LiDAR	UAV	4.5	Semantic labels	Yes	USA	USA
[63]	ITD	RGB	UAV	0.4	Keypoints	Yes	USA	USA
[78]	ITD	RGB, PPC	UAV	Undisclosed	No	Yes	Sweden	Iran
[72]	ITD	RGB	UAV	0.82	Bounding boxes	Yes	Saudi Arabia	Saudi Arabia
[80]	ITD	RGB	PA	Undisclosed	Bounding boxes	No	Belgium	Spain
[71]	ITD	RGB	N/A	0.008	Bounding boxes	Yes	France	N/A
[68]	ITD	RGB	PA	0.075	Bounding boxes	Yes	N/A	N/A
[64]	ITD	RGB, HSI, LiDAR	PA	27.4	Bounding boxes	Yes	USA	USA
[40]	ITD	RGB	PA	0.032	Bounding boxes, Keypoints	Yes	USA	USA
[87]	SS, SI	RGB	UAV	69	Polygons, Semantic labels	Yes	Germany	Germany
[51]	SS, SI	RGB	UAV	0.48	Semantic labels	Yes	Ecuador	Ecuador
[52]	SS, SI	RGB	UAV	7.5	Polygons	Yes	S. Africa	S. Africa
[65]	SS	RGB	PA	1.4	Semantic labels	Yes	Poland	Poland
[69]	SS	RGB, MSI	UAV	2.8	Semantic labels	Yes	USA	USA
[26]	SS	RGB	UAV	3.7	Semantic labels	Yes	Japan	Japan
[38]	SS	RGB	UAV (Synthetic)	0.74	Semantic labels	Yes	Germany	Russia
[39]	SS	LiDAR	PA	10	Semantic labels	Yes	USA	Canada
[53]	SS, Misc.	RGB	UAV	9.3	Polygons	Yes	UK	Spain
[54]	DD, HSA	RGB, MSI	UAV	18	Polygons	Yes	Finland	Finland
[22]	DD	RGB, MSI	UAV, Terrestrial	39.25	Keypoints	No	Greece	Macedonia
[23]	DD	RGB, MSI	UAV, Terrestrial	59.9	Keypoints	No	Greece	Greece
[42]	DD	MSI	UAV	1.14	No	Yes	China	China
[55]	SI	MSI, LiDAR	PA	147.5	Polygons	Yes	France	France
[56]	SI	RGB	UAV	Undisclosed	Bounding Boxes	No	Switzerland	Switzerland
[86]	SI	RGB, MSI	Satellite	Undisclosed	Semantic labels	Yes	China	China
[25]	SI	RGB	UAV	19.3	Semantic labels	Yes	Australia	Australia
[73]	CSE	RGB, PPC	UAV	0.2	No	Yes	Australia	Australia
[84]	CSE	LiDAR	PA	0.52	No	Yes	USA	N/A
[33]	CSE	RGB	UAV	7.5	Bounding boxes	Yes	Germany	Ecuador
[28]	HSA	LiDAR, HSI	UAV	126	No	Yes	UK	Germany
[29]	HSA	RGB	UAV	8.45	No	Yes	UK	Germany
[36]	GM	TLS, PPC	UAV, Terrestrial	11.2	No	Yes	Finland	Finland
[57]	PM	RGB	UAV	5.2	Bounding boxes	Yes	China	Finland
[27]	PM	MSI	UAV	11.3	No	Yes	Netherlands	Netherlands
[58]	Misc.	RGB (Synthetic)	3D model	2000	Semantic labels	Yes	USA	USA
[59]	Misc.	LiDAR	UAV, Terrestrial	10.7	No	Yes	Spain	Germany
[74]	Misc.	RGB	PA	Undisclosed	No	Yes	New Zealand	New Zealand
[75]	Misc.	RGB	PA	Undisclosed	No	Yes	New Zealand	New Zealand
[76]	Misc.	RGB	PA	Undisclosed	No	Yes	New Zealand	New Zealand
[77]	Misc.	RGB	PA	Undisclosed	No	Yes	New Zealand	New Zealand
[81]	Misc.	RGB	UAV	0.25	No	Yes	N/A	N/A
[60]	Misc.	RGB, MSI	UAV	16	No	Yes	France	Côte d’Ivoire
[61]	Misc.	LiDAR	UAV	3.5	No	Yes	Germany	Germany
[62]	Misc.	RGB, LiDAR, PPC	UAV, Helicopter	0.35	No	Yes	Italy	Italy
[18]	Misc.	MSI	UAV	0.212	No	Yes	Sweden	Sweden
[19]	Misc.	PPC	UAV	0.173	No	Yes	Sweden	Sweden
[20]	Misc.	RGB	UAV	1.4	No	Yes	Brazil	Brazil
[21]	Misc.	RGB, MSI	UAV	7.6	No	Yes	Brazil	Brazil
[67]	Misc.	RGB	UAV (Synthetic)	Undisclosed	Semantic labels	Yes	Belgium	N/A
[66]	Misc.	RGB	UAV	1.6	No	Yes	Finland	Finland
[79]	Misc.	LiDAR	PA	0.7	N/A	No	China	China
[30]	Misc.	HSI	PA	28.66	No	No	Finland	Finland
[70]	Misc.	LiDAR	PA	Undisclosed	N/A	No	Switzerland	Switzerland
[32]	Misc.	HSI	PA	4.61	No	No	India	India
[34]	Misc.	RGB, MSI	UAV	3.5	No	Yes	Germany	Germany
[35]	Misc.	RGB, MSI	UAV	1.5	No	Yes	Germany	Germany

lists the published datasets, characterized by applications (possible examples), data source type, platform used in data acquisition, size of the dataset, data annotation availability, if dataset is open access, corresponding author country of origin, and country of geographical extent. The datasets are sorted from most to least frequent application (detailed in the following).

To begin, a set of parameters was extracted from the sources, namely, the following:

• Application: This parameter refers to the specific use or purpose of a dataset according to the dataset author, offering insights for individuals seeking tailored or annotated data suited for particular applications. Notably, this criterion operates as the subsequent categorization factor following the distinction between published and unpublished datasets.

  Table 5. Overview of applications.

  Application	Description
  Individual Tree Crown Delineation (ITCD)	Identifies individual tree crowns, providing detailed tree-level data, but is computationally intensive and depends on image quality.
  Individual Tree Detection (ITD)	Detects individual trees, estimating tree density and distribution, but faces challenges in dense forests and possible counting errors.
  Semantic Segmentation (SS)	Classifies image pixels into categories with high accuracy in tree recognition but requires large labeled datasets and is computationally demanding.
  Disease Detection (DD)	Identifies tree diseases for early detection and timely intervention but may need hyperspectral data, and subtle disease symptoms can be hard to detect.
  Species Identification (SI)	Determines tree species, supporting biodiversity studies, but similar species may be hard to distinguish.
  Carbon Stock Estimation (CSE)	Estimates forest carbon storage, important for climate studies, but requires accurate biomass models and is affected by terrain variability.
  Health Status Analysis (HSA)	Assesses tree health, monitoring forest vitality and aiding conservation, but involves complex health indicators and environmental influences.
  Geometrical Measurement (GM)	Measures tree dimensions, providing quantitative data for growth studies, but needs precise data and can have canopy overlap issues.
  Phenology Monitoring (PM)	Observes seasonal changes, studying climate impacts and aiding agricultural planning, but faces temporal resolution limits and weather interference.
  Miscellaneous (Misc.)	Corresponds to datasets with no clear indication of a specific use case.

  shows the applications of the identified datasets, described ahead in this section.
  Figure 3 shows the distribution of use case applications in the identified datasets. Apart from miscellaneous, the most frequent applications include ITCD, SI, and SS.
• Source: Describing the data’s structure and format. The term “sensor” is intentionally omitted in this context to encompass various data acquisition methods; for instance, point cloud data can be obtained through laser scanning or generated via structure from motion, namely, the photogrammetric point clouds (PPCs). Figure 4 demonstrates the distribution of data source of the identified datasets. The most-used source is the RGB camera, followed by the multispectral camera and LiDAR (Light Detection and Ranging). It must be mentioned that thermal cameras are regularly used in forestry-related studies [7]; however, none of the datasets we found present purely thermal-spectrum data, and any dataset with thermal band modality is included in the multispectral category.
• Platform: Refers to the remote sensing platform that carries the sensor payload. Understanding the platform is crucial as it determines the data perspective, resolution (not definitive), and the feasibility of data collection in diverse environmental conditions. It significantly influences the quality and scope of the acquired data, impacting subsequent analysis and applications. Figure 5 shows distribution of platforms carrying the sensors in selected datasets. The graph highlights the predominant platforms utilized for sensor deployment, with UAVs being the most extensively employed, followed by piloted aircraft and satellites. While our filtering criteria prioritize aerial perspective datasets, the graph also encompasses terrestrial point-of-view (POV) data. This inclusion is prompted by instances where datasets offer both aerial and terrestrial data as complementary sources, thus meeting the eligibility criteria. Moreover, the presence of three instances of synthetic data is worthy of note. Two of these particular datasets offer synthesized data from a UAV’s POV, and the third offers 3D models of trees, which can be used from any POV; therefore, they are presented under the UAV category. Further insights into synthetic datasets are explored in more detail in Section 3.1.11.
• Dataset Size: Indicates the volume or scale of the dataset, which plays a pivotal role in various aspects of analysis and application. Understanding the dataset size is fundamental as it influences computational requirements, model training, and the feasibility of certain analytical approaches.
  As an example, larger datasets may demand advanced computational resources, while smaller ones might be more manageable for specific analyses or applications. Assessing dataset size aids in determining the data’s comprehensiveness and potential limitations, guiding researchers in selecting appropriate methodologies and tools for analysis.
• Annotation: Represents the level of labeling, tagging, or additional information embedded within the dataset. Annotation serves as a crucial aspect, especially in supervised learning or applications requiring labeled data, or even just to assess the developed algorithm performance. Understanding the annotation level informs researchers about the data’s usability for specific tasks such as object detection, classification, or segmentation. Higher levels of annotation often indicate enhanced data richness but may also involve increased resource investment. Assessing annotation levels guides the selection of datasets aligned with the required granularity for precise analysis or model training. Figure 6 shows distribution of different annotation types in the selected datasets across the years they were published. Among the identified datasets, a predominant observation is the prevalence of unannotated data. When annotations are present, semantic-level annotations are the most frequently encountered type, followed by polygons and bounding box annotation. The distinction between polygon and semantic-level annotation arises from the nuanced application of polygons in certain cases, where they are utilized to achieve a higher degree of precision in object detection techniques compared to the more common use of bounding boxes.
• Open Access Status: Indicates whether the dataset is publicly available or restricted in its accessibility. This parameter holds immense significance in fostering collaboration, reproducibility, and the advancement of research. Open access datasets promote transparency and facilitate broader utilization by the scientific community, accelerating innovation and knowledge dissemination. Understanding the open access status aids researchers in identifying available resources for validation, comparison, or extension of existing studies. Moreover, it influences the reproducibility and credibility of research findings, emphasizing the importance of accessible data in driving scientific progress.
• Country: Denotes two key parameters, which are the country of corresponding author’s affiliated institution and the country associated with the dataset’s geographical extent. Understanding the author’s country provides insights into regional focuses, environmental contexts, or unique challenges that might impact the study’s scope or applicability. Simultaneously, recognizing the country representing the geographical extent of the dataset contributes to the broader understanding of global perspectives and diverse approaches within the field, highlighting regional contributions and fostering collaborative opportunities across different geographical areas. Figure 7 depicts the distribution of published datasets by the author’s country, showcasing the prevalence of dataset contributions. Germany leads with 10 published datasets, followed by the United States and China with 9 and 7, respectively. Figure 8 shows the distribution of published datasets by the geographical extent of the dataset. The most frequent countries are the United States with 9, Germany with 8, followed by China and New Zealand each with 5.
  Examining the parameters “Corresponding Author’s Country” (CA) and “Geographical Extent Country” (GE), it is evident that they exhibit congruent coverage, indicating a substantial overlap in the regions they represent.

The following sections describe the different applications considered in Table 4.

3.1.1. Individual Tree Crown Delineation

Individual tree crown delineation (ITCD) is the process of outlining the perimeters of trees within remote sensing imagery or point cloud data, including both two-dimensional (2D) and three-dimensional (3D) spatial representations. This method is widely used in several domains, notably in forest inventory, ecological studies, and advanced land management practices. ITCD helps with the identification and mapping of individual tree crowns within forested areas, facilitating not just accurate quantification, but the holistic characterization of various tree attributes such as crown size, canopy cover, species distribution, and tree health. It allows for the systematic and rapid processing of extensive data, a critical requirement in today’s vast and data-rich domains.

This technique is typically applied to 2D data, but with the advancement in computational capacity and the emergence of sophisticated methods such as CNNs, 3D data are being explored for ITCD. While CNNs have been highly effective in 2D imagery, their extension to 3D point cloud data is less direct. Architectures such as PointNet++ can segment 3D point sets, but they require fixed-size subsampling (e.g., 1024–2048 points) and often depend on a preprocessing step such as clustering or crown isolation to make per-tree delineation feasible. These constraints highlight why 3D ITCD methods remain less mature than 2D approaches. Refs. [24,37,41,44,45,46,82,83,88] are labeled RGB dataset examples, Ref. [31] presents labeled LiDAR-based tree point clouds with detailed annotation, and Ref. [85] offers annotated Multispectral data. Here, ITCD refers to a forestry-specific application that uses data processing techniques such as semantic or instance segmentation. However, these labeled datasets extend beyond this application, serving as valuable resources for AI and image processing development. Further details on semantic segmentation can be found in Section 3.1.3.

Figure 9 illustrates an example of ITCD application with visualized annotation.

3.1.2. Individual Tree Detection

Individual tree detection (ITD) is the identification and localization of individual trees within remote sensing data, utilizing both 2D and 3D spatial representations. Unlike ITCD, which intricately outlines tree crowns, ITD primarily focuses on pinpointing tree positions within a designated area.

This technique is fundamental in forest inventory, ecological monitoring, and land management practices, providing critical insights into tree distribution, density, and spatial arrangement. In 2D applications, such as satellite imagery or aerial photographs, tree detection involves locating tree positions based on spectral signatures and spatial patterns. This aids in broader-scale assessments of forest cover, landscape dynamics, and large-scale forest management. In contrast, 3D data, often derived from technologies like LiDAR, enhance tree detection by providing detailed three-dimensional information. Three-dimensional data facilitate the accurate localization of individual trees within the vertical canopy profile, allowing for precise measurements of tree height, canopy structure, and spatial arrangement in the vertical dimension. In practical applications, individual tree detection from 3D point clouds typically relies on analytical approaches such as local maxima filtering, watershed segmentation, or clustering around canopy apex points. Direct application of deep learning models to large-scale 3D forest scenes remains challenging due to the need for subsampling and preprocessing to isolate individual crowns. As a result, 3D ITD methods are still less mature and less widely adopted than 2D image-based approaches.

Refs. [40,47,48,49,64,68,71,72,80] present RGB data, acquired for ITD application, while [78], in addition to RGB data, also offers a photogrammetric point cloud of the scenery. This point cloud can be utilized in methods such as local maxima filter to identify the tree tops in 3D space. An example of ITD is visualized in Figure 10.

3.1.3. Semantic Segmentation

Semantic segmentation (SS) is a technique in which every pixel in 2D data or point/voxel in 3D space is assigned a label. While ITCD outlines every tree crown within an image and in point cloud data, SS covers broader classifications, e.g., various tree parts including canopy, understory branches, trunk, and other fine details (see Figure 11). This, method prioritizes the precise classification of these specific components present in the scene, leveraging advanced neural network architectures. The advancements in state-of-the-art neural networks, exemplified by architectures like U-Net, have significantly contributed to the refinement and precision of semantic segmentation methodologies. While these advances have strongly benefited 2D imagery, extending semantic segmentation to 3D point clouds remains more challenging. Models such as PointNet++ or KPConv can process 3D data, but they typically rely on fixed-size subsampling of points, which reduces resolution and limits per-class detail when multiple classes are present. As a result, 2D semantic segmentation with CNNs is highly established, whereas 3D methods remain less mature and continue to face scalability and accuracy constraints in complex forest scenes. This focused approach in semantic segmentation complements ITCD by enabling a more versatile analysis across varied data formats and object types within forested landscapes. Moreover, its widespread use in multiple disciplines underscores its adaptability and significance in delineating and understanding specific objects within complex visual scenes.

The datasets by [26,39,52,65,69,87] encompass high-resolution aerial data depicting various landscapes containing objects beyond trees. These datasets are accompanied by corresponding masks, specifically tailored for semantic segmentation purposes. The FOR dataset by [31] offers fine detailed annotation for instance and semantic segmentation of different tree types. Ref. [38] provides a synthetic Siberian Larch tree crown dataset, constructed from drone imagery.

3.1.4. Disease Detection

Disease detection (DD) in trees involves the identification and continual monitoring of various diseases and pathogens that pose threats to them, using remote sensing technologies. This process is instrumental in the early detection, assessment, and effective management of diseases that hold the potential to substantially impair forest health and productivity.

Using 2D data, such as hyperspectral imagery, enables the identification of spectral anomalies and subtle changes in vegetation health, aiding in pinpointing disease hotspots and affected areas across wide forested landscapes. Meanwhile, leveraging 3D data, particularly from LiDAR scans, offers insights into the vertical structure of forests, facilitating the detection of structural changes or deformities within tree canopies caused by diseases or stress factors.

It is important to note, however, that disease detection in 3D data is generally constrained to tree- or scene-level classification. Current point-cloud deep learning methods are not yet capable of robustly segmenting partial diseased structures (e.g., specific branches or sub-crown regions) within complex forest scenes. As a result, while 2D spectral approaches remain the dominant and most effective method for identifying subtle disease symptoms, 3D analyses are typically limited to detecting broader structural changes or canopy-level impacts.

Nevertheless, by combining the strengths of 2D and 3D data, disease detection methodologies enable comprehensive analyses of disease distribution patterns, supporting proactive intervention strategies and informed decision-making in trees’ health management. Ref. [42] (Banana Fusarium wilt), [22] (Cherry Armillaria root rot), [23] (Anarsia lineatella and Grapholita molesta), and [54] (Bark Beetle) are datasets for disease detection, harnessing the significance of multispectral data, which offers an expanded range of spectral bands beyond RGB, thereby enhancing the precision and efficacy of spectral analysis in identifying tree diseases.

3.1.5. Species Identification

Species identification (SI) is defined as discerning and categorizing tree species within forested regions based on the remote sensing data. This process utilizes an array of spectral, textural, and structural information garnered by diverse sensors employed in this field. By analyzing the unique spectral signatures and structural attributes of tree canopies, this method aims to classify the tree species coexisting within a landscape.

Machine learning algorithms facilitate the differentiation of tree species based on their distinct spectral responses and structural attributes captured by the remote sensing data. This automation allows for the precise delineation and classification of diverse tree species across vast forested regions, fostering a detailed understanding of biodiversity, species distribution patterns, and intricate ecosystem dynamics.

See Figure 12 for an example visualization of SI application.

As examples, datasets by [25,45,86] offer labeled RGB and multispectral data for tree SI.

3.1.6. Geometrical Measurement

Geometrical measurement (GM), as the name suggests, is the quantification of diverse spatial and structural attributes defining forest and tree elements, extracted from the remote sensing data. These measurements include parameters like tree height, crown diameter, canopy cover, and tree spacing, among others.

The significance of these measurements extends across various domains including forest inventory, accurate biomass estimation, and continual monitoring of shifts in forest structure over time, ultimately aiding in the formulation of adaptive and sustainable forest management practices. Ref. [36] offers terrestrial LiDAR data along with UAV-born photogrammetric point cloud of Scots pine trees, which they used to measure the geometrical attributes of trees in their study.

3.1.7. Carbon Stock Estimation

Carbon stock estimation (CSE) is the quantification of carbon accumulated within forests, including both the above-ground and below-ground biomass. This process relies on remote sensing data to estimate the volume of carbon stored by trees and vegetation. By analyzing forest parameters such as tree height, canopy structure, and biomass density, this estimation offers valuable insights into the total carbon content stored within forested regions or tree plantations.

This evaluation of carbon stocks is crucial in comprehending the role forests and in general trees play in mitigating climate change and aiding carbon offset initiatives. See Figure 13 for a visualization of the dataset by [33] in which bounding box annotation fused with ground truth data (geometrical measurements along with allometric calculations) is used to pinpoint tree locations with their corresponding carbon stock.

The dataset by [84] offers a relatively small LiDAR-based point cloud dataset for CSE, and [33] presents an RGB dataset for this application. The dataset by [73] offers RGB orthomosaic and photogrammetric point cloud for Mangrove tree biomass estimation (while biomass and carbon stock estimation are distinct yet related concepts in forestry and ecological studies, for this study, we have consolidated them into a single category).

3.1.8. Health Status Analysis

Health status analysis (HSA) goes beyond the singular focus of disease detection. It includes a broader evaluation of the overall condition and vigor of trees. This process aims to identify and assess a spectrum of indicators including factors such as stress, disturbances, and general tree health. Various sensors, capable of detecting anomalies in vegetation health, canopy structure, or even subtle phenological changes, contribute to this analysis. By facilitating the timely implementation of mitigation measures, this assessment contributes significantly to maintaining or restoring trees’ health. The datasets provided by [28,29] collectively comprise RGB, hyperspectral, and LiDAR data obtained from a drought-stressed European beech forest reserve in Germany.

3.1.9. Phenology Monitoring

Phenology monitoring (PM) is tracking and analyzing seasonal changes and patterns in vegetation and forest ecosystems. It utilizes temporal remote sensing data to observe the timing of recurring biological events like leaf emergence, flowering, or leaf senescence within forests. By capturing these temporal dynamics, phenology monitoring provides insights into ecosystem responses to environmental factors, climate variations, and disturbances. It helps us understand the relationships between vegetation phenology and environmental changes, contributing to assessments of ecosystem health, productivity, and biodiversity.

The dataset by [27] offers multispectral data for monitoring individual tree phenology in a multi-species forest. See Figure 14 for a visualization of this dataset.

3.1.10. Miscellaneous

Any entry that did not align with the mentioned applications or did not provide a clear reason for creating the dataset is categorized under miscellaneous.

The work by [79] offers UAV-born LiDAR data of a forest park in China. Datasets by [20,21] offer RGB and multispectral data captured by UAV over a subtropical forest area in Brazil. Datasets by [18,19] provide multispectral point cloud and orthomosiac of Nybygget in Swewden.

The dataset provided by [66] comprises high-resolution georeferenced UAV images featuring trees. While the dataset’s primary objective focuses on localization through image processing, these images also hold potential for various other tasks and applications.

Datasets by [34,35] offer multispectral datasets for mapping of vegetation and hydromorphology along Federal waterways in Germany. Datasets by [30,32] offer high-resolution UAV-born hyperspectral data and the ground truth from tropical forest ecosystems. The dataset by [70] offers an airborn LiDAR dataset for forest research. Lastly, the synthetic dataset by [67] offers several data modalities from the point of view of a low-altitude-flying UAV. More on this dataset and synthetic datasets is presented in the next section.

3.1.11. Synthetic Data

Synthetic data’s emergent trend lies predominantly within academic research rather than in published datasets at this stage. This type of data, while less observable in this dataset compilation, holds significant promise due to the capabilities it offers, such as controlled variability and simulation advantages for testing scenarios and expanding analytical capabilities.

Aside from its influence on model training, one of synthetic data’s most significant advantages is the fact that data labeling is much less labor intensive compared to manual data annotation.

Within the identified datasets, three of them stand out particularly due to their synthetic nature, deviating from the predominant composition of real-world datasets.

First, the MidAir synthetic dataset by [67], which is categorized as miscellaneous. This dataset serves multiple purposes, tailored for low-altitude drone operations. It offers a substantial volume of synchronized data aligned with flight records. This inclusive dataset encompasses information from various vision and navigation sensors installed on a quadcopter in flight. The vision sensors capture a diverse array of data, including RGB images, surface orientation details, depth perception, SS masks, object characteristics, and stereo disparity (Figure 15). Although the dataset’s point of view does not precisely meet our filtering conditions, we have included it in this paper for two main reasons: Firstly, it offers high-resolution data of individual trees; therefore, it is not totally unrelated to the subject. Secondly, it showcases the potential of synthetic data generation.

The second synthetic dataset, by [38], was introduced in Section 3.1.3, being a dataset for SS application. The authors utilized RGB imagery from a UAV platform to create a training set by placing cropped-out Larch trees onto various backgrounds. This set is designed for training models to identify Larch trees in aerial imagery with forest backgrounds. Starting with 117 cropped-out Larch trees as foregrounds and three sets of 35 of backgrounds, by shuffling these foregrounds and backgrounds, this dataset offers 10.000 images for training purposes and 2000 images for validation purposes. See Figure 16 for a visualization of this dataset.

The third work [58] introduces the first large-scale dataset of simulation-ready 3D tree models, featuring 600,000 models reconstructed from single RGB images using diffusion-based priors. The authors propose a novel pipeline that leverages text-prompted, tree-genus-conditioned diffusion models to reconstruct 3D envelopes of trees from single images and then uses a space colonization algorithm to estimate detailed branching structures. The resulting dataset provides richly detailed, environmentally aware 3D tree models suitable for simulations and ecological studies, standardizing benchmarking for forestry applications and enabling tasks such as phenotyping, biomass estimation, species identification, and more. Figure 17 visualizes examples of this dataset.

3.2. Unpublished Data

In academic research, a specific subset of published papers contains valuable yet undisclosed data. These papers, while publicly accessible, withhold the detailed datasets acquired during the course of their investigations. Often authors (for one reason or another) decide to keep the data associated to their published papers undisclosed, impairing further critical discussion on their methods and findings. However, their significance cannot be understated. These datasets, despite being in an inactive state, constitute unexplored potential, presenting opportunities for critical insights and unexplored possibilities in research. Contacting the authors of such papers and asking for their data provides the opportunity to enhance current research paradigms and potentially help with the developments within the academic landscape.

The structure of this section is the same as the previous one, and the same set of parameters were extracted from the entries and analyzed. Since here the data are not published and rather used and explained, the information is extracted from the papers rather than the datasets themselves. Therefore, two parameters, namely, the “Dataset size” and “Open access” statuses were excluded.

Table 6. Unpublished data. (*) CA, GE, N/A, and PA, respectively, stand for Corresponding Author, Geographical Extent, Not Available, and Piloted Aircraft.

Dataset/ Study	Application	Source	Platform	Annotation	CA (*) Country	GE (*) Country
[90]	ITCD, ITD, HSA	RGB	UAV	Polygons, Semantic labels	Malaysia	UAE
[91]	ITCD, ITD	MSI	SAT	Polygons	UAE	UAE
[92]	ITCD, ITD	RGB	UAV	Polygons, Image labels	Malaysia	Malaysia
[93]	ITCD, ITD	MSI	UAV	Polygons	Greece	Greece
[94]	ITCD, ITD	RGB, PPC	UAV	Polygons	China	China
[95]	ITCD, ITD	RGB	UAV	Polygons, Bounding boxes	Mexico	Canada
[96]	ITCD, SI, GM	HSI	UAV	Polygons	China	China
[97]	ITCD, SI	RGB	UAV	Polygons, Semantic labels	Japan	Japan
[98]	ITCD, SI	HSI	UAV	Polygons, Semantic labels	China	China
[99]	ITCD, SI	RGB, MSI	UAV	Polygons, Semantic labels	Australia	Australia
[100]	ITCD, SI	RGB, LiDAR	UAV, PA	Semantic labels	Japan	Japan
[101]	ITCD, SI	RGB	UAV	Semantic labels	Brazil	Brazil
[102]	ITCD, HSA	MSI	UAV	Polygons, Semantic labels	Italy	Greece
[103]	ITCD, SS	RGB	UAV, SAT	Polygons, Semantic labels	China	China
[104]	ITCD, PM	MSI	UAV	Bounding boxes, Semantic labels	USA	USA
[105]	ITCD	RGB, PPC	UAV	Polygons, Semantic labels	Poland	Iran
[106]	ITCD	RGB	UAV	Polygons	China	China
[107]	ITCD	MSI	UAV	Polygons	Mexico	Peru
[108]	ITCD	RGB	UAV	Polygons	China	China
[109]	ITCD	RGB, PPC	PA	Polygons	China	China
[110]	ITCD	RGB	UAV	Polygons	Turkey	Turkey
[111]	ITCD	RGB, LiDAR	PA	Polygons	China	China
[112]	ITCD	MSI. LiDAR	UAV	Polygons	Germany	Germany
[113]	ITCD	RGB	UAV	Polygons	China	China
[114]	ITCD	RGB	UAV	Polygons	China	China
[115]	ITCD	MSI	PA	Polygons	Australia	Australia
[116]	ITCD	RGB	UAV	Semantic labels	China	China
[117]	ITCD	RGB	UAV	Semantic labels	Brazil	Brazil
[118]	ITCD	RGB, PPC	UAV	No	China	China
[119]	ITCD	MSI	UAV	No	Italy	Italy
[120]	ITCD	RGB, PPC	UAV	No	UK	Malaysia
[121]	ITCD	RGB	UAV	Bounding boxes, Polygons	Japan	Japan
[122]	ITCD	RGB	UAV	No	Iran	Iran
[123]	ITCD	MSI	UAV	Semantic labels	Spain	Spain
[124]	ITCD	RGB	UAV	Semantic labels	China	China
[125]	ITCD	MSI	UAV	Semantic labels	India	India
[126]	ITCD	RGB	UAV	Semantic labels	China	China
[127]	ITCD	LiDAR	UAV	No	China	China
[128]	ITCD	RGB, Thermal	UAV	Semantic labels	Netherlandsbn	Iran
[129]	ITCD	HSI, LiDAR	PA	Semantic labels	Switzerland	Switzerland
[130]	ITCD	MSI	UAV	Semantic labels	Spain	Spain
[131]	ITCD, GM	RGB, PPC	UAV	Semantic labels	China	China
[132]	ITCD	RGB	UAV	Bounding boxes	Malaysia	Malaysia
[133]	ITCD, HSA	RGB	UAV	N/A	Romania	N/A
[134]	ITCD, SI	RGB	UAV	Semantic labels	Brazil	Brazil
[135]	ITCD	RGB	UAV	Bounding boxes, Semantic labels	China	China
[136]	ITCD	LiDAR	UAV	No	UK	UK
[137]	ITCD	RGB	UAV	Bounding boxes, Semantic labels	China	China
[138]	ITCD	RGB	Satellite	Keypoints	China	French Polynesia
[139]	DD, SS	RGB	UAV	Image labels	S. Korea	S. Korea
[140]	DD, HSA	RGB, Thermal, LiDAR	UAV	Semantic labels	China	China
[141]	DD, HSA	MSI, LiDAR	UAV	Polygons, Semantic labels	China	China
[142]	DD	RGB	UAV	Bounding boxes	China	China
[143]	DD	RGB	UAV	Bounding boxes	India	India
[144]	DD	MSI	UAV	Polygons	S. Africa	S. Africa
[145]	DD	RGB	UAV	Polygons	China	China
[146]	DD	N/A	UAV	Bounding boxes	China	China
[147]	DD	RGB	UAV	Image labels	China	China
[148]	DD	RGB	UAV	Semantic labels	China	China
[149]	DD	RGB, MSI	UAV	Image labels	China	China
[150]	DD	RGB	UAV	Image labels	China	China
[151]	DD	HSI	UAV	Polygons, Bounding boxes	China	China
[152]	DD	MSI	UAV	Polygons, Semantic labels	Italy	Italy
[153]	DD	MSI	UAV	Semantic labels	Greece	Greece
[154]	DD	RGB	UAV	Bounding boxes	China	China
[155]	DD	HSI	UAV	Bounding boxes	China	China
[156]	DD	RGB	UAV	Bounding Boxes, Semantic labels	China	China
[157]	DD	MSI	UAV	No	Malaysia	Malaysia
[158]	DD	RGB, MSI	UAV	N/A	Finland	Finland
[159]	DD	RGB	UAV	Image labels	S. Korea	S. Korea
[160]	DD	MSI	UAV	N/A	China	Mongolia
[161]	DD	MSI	UAV	Bounding boxes	Czech R.	Czech R.
[162]	DD	HSI	UAV	N/A	China	China
[163]	DD	RGB	UAV	No	USA	USA
[164]	DD	HSI	UAV	Semantic labels	Australia	Australia
[165]	DD	RGB	UAV	Keypoints	Japan	Japan
[166]	DD	RGB, MSI, HSI	UAV	Image labels	Finland	Finland
[167]	DD	RGB, MSI	UAV	No	China	China
[168]	DD	RGB	UAV	Semantic labels	China	China
[169]	DD	HIS, LiDAR	UAV	No	China	China
[170]	DD	HSI	UAV	No	China	China
[171]	DD	HSI, LiDAR	UAV	Bounding Boxes, Semantic labels	China	CHina
[172]	DD	MSI	UAV	Bounding boxes	China	China
[173]	DD	RGB, HSI	UAV	N/A	China	China
[174]	DD	MSI	UAV	Bounding boxes	China	China
[175]	DD	RGB	PA	Bounding boxes	China	China
[176]	ITD, SS, GM	RGB, MSI, LiDAR	UAV	Polygons, Semantic labels	Russia	Russia
[176]	ITD, GM, Misc.	RGB	UAV	Keypoints	Turkey	Turkey
[177]	ITD, SI	MSI	UAV	Polygons	Brazil	Brazil
[178]	ITD, SI	RGB, LiDAR	UAV	Bounding boxes, Semantic labels	China	China
[179]	ITD, SI	RGB, MSI	UAV, Satellite	Polygons	China	China
[180]	ITD, SI	HSI, LiDAR	UAV	Keypoints, Semantic labels	China	China
[181]	ITD, SI	RGB	UAV	Polygons	China	China
[182]	ITD, HSA	RGB	UAV	Image labels	USA	USA
[183]	ITD, Misc.	RGB, PPC	UAV	Polygons	Germany	Germany
[184]	ITD, Misc.	RGB, MSI	UAV	Polygons, Bounding boxes	China	China
[185]	ITD	RGB	UAV	Bounding boxes	Peru	Peru
[186]	ITD	LiDAR	UAV	Polygons, Bounding boxes	USA	USA
[187]	ITD	RGB	UAV	Image labels	Morocco	Morocco
[188]	ITD	RGB	UAV	Keypoints	China	China
[189]	ITD	RGB	UAV	Bounding boxes	China	China
[190]	ITD	RGB	UAV	Bounding boxes	Sweden	Sweden
[191]	ITD	RGB	Satellite	Keypoints	China	China
[192]	ITD	RGB	UAV	Keypoints	Brazil	Brazil
[193]	ITD	RGB	UAV	Bounding boxes	Canada	Canada
[194]	ITD	RGB, PPC	UAV	N/A	Chile	Chile
[195]	ITD	MSI	UAV	Semantic labels	USA	USA
[196]	ITD	RGB	UAV	Keypoints	China	China
[197]	ITD	RGB	UAV	Image labels	India	India
[198]	ITD	RGB	UAV	N/A	Brazil	Brazil
[199]	ITD	RGB, MSI	UAV	Bounding boxes	China	China
[200]	ITD, GM	RGB	UAV	Bounding boxes	USA	USA
[201]	ITD	RGB	UAV	No	Iran	Iran
[202]	ITD, SI	HSI	UAV	Keypoints	Brazil	Brazil
[203]	ITD	PPC, LiDAR	UAV	No	Brazil	Brazil
[204]	ITD	MSI	UAV	Keypoints	Brazil	Brazil
[205]	ITD	RGB	UAV	No	Spain	Spain
[206]	ITD, BE	RGB	UAV	Bounding boxes	China	China
[207]	ITD	RGB	UAV	Bounding boxes	China	China
[208]	ITD	RGB	UAV	Bounding boxes	China	Brazil
[209]	SI, CSE	RGB	UAV	Semantic labels	Nigeria	Nigeria
[210]	SI, Misc.	RGB, LiDAR	UAV	Semantic labels	China	China
[211]	SI	RGB, MSI	UAV	Bounding boxes	Finland	Finland
[212]	SI	RGB	UAV	Polygons	S.Korea	S.Korea
[213]	SI	MSI, LiDAR	PA	Polygons, Semantic labels	Canada	Canada
[214]	SI	RGB, MSI	UAV	Semantic labels	Taiwan	Taiwan
[215]	SI	RGB	UAV	Polygons	China	China
[216]	SI	RGB, MSI, HSI	UAV	Polygons	China	China
[217]	SI	RGB	UAV	Image labels	Italy	Italy
[218]	SI	MSI	UAV	Polygons	China	China
[219]	SI	RGB, MSI	UAV, Satellite	Polygons	China	China
[220]	SI	HSI	UAV	Semantic labels	China	China
[221]	SI	RGB, MSI, LiDAR	UAV	Keypoints	Finland	Finland
[222]	SI	RGB	UAV	Image labels	Germany	Germany
[223]	SI	RGB	UAV	Image labels	China	China
[224]	SI	RGB	PA	Semantic labels	Brazil	Brazil
[225]	SI	RGB	UAV	Semantic labels	Brazil	Brazil
[226]	SI	LiDAR, HSI	UAV	Semantic label, Polygons	China	China
[227]	SI	RGB, MSI	UAV	Semantic labels	China	China
[228]	GM	RGB	UAV	Polygons	Slovakia	Slovakia
[229]	GM	RGB	UAV	No	Brazil	Brazil
[230]	GM	RGB	UAV	Semantic labels	Romania	Iran
[231]	GM, PM	RGB	UAV	No	Norway	Norway
[232]	GM	RGB, MSI	UAV	No	Mexico	Mexico
[233]	HSA, Misc.	RGB, MSI, LiDAR	UAV	Polygons, Image labels	Greece	Greece
[234]	HSA	RGB	UAV	Image labels	Japan	Japan
[235]	HSA	MSI	PA	Polygons	USA	USA
[236]	HSA	MSI	UAV	Polygons	Sri Lanka	Sri Lanka
[237]	HSA	HSI	UAV	Polygons, Semantic labels	Malaysia	Malaysia
[238]	HSA	RGB, MSI	UAV	Polygons, Image labels	Pakistan	Pakistan
[239]	HSA	MSI	UAV	Polygons	China	China
[240]	HSA	RGB, MSI	UAV	Image labels	Malaysia	Malaysia
[241]	HSA	MSI	UAV	Polygons, Semantic labels	Germany	Germany
[242]	HSA	MSI	UAV	Semantic labels	Australia	Australia
[243]	HSA	RGB, MSI, Thermal	UAV	No	Netherlands	Netherlands
[244]	PM	RGB	UAV	Bounding boxes	China	China
[245]	CSE, HSA	RGB, PPC	UAV	Polygons, Semantic labels	China	China
[246]	CSE	RGB	UAV	Polygons	Spain	Spain
[247]	CSE	RGB	UAV	Polygons, Semantic labels	India	India
[248]	CSE	RGB, MSI	UAV	Semantic labels	Russia	Russia
[249]	SS, Misc.	RGB	UAV	Polygons	Vietnam	Vietnam
[250]	SS, Misc.	RGB	UAV	Polygons, Semantic labels	Portugal	Portugal
[251]	SS	MSI, LiDAR	UAV	Polygons, Semantic labels	China	China
[252]	Misc.	RGB	UAV	No	New Zealand	New Zealand
[253]	Misc.	RGB	UAV	Polygons, Semantic labels	Australia	Australia
[254]	Misc.	RGB	UAV	No	Germany	Germany
[255]	Misc.	HSI	UAV	Polygons, Semantic labels	USA	USA
[256]	Misc.	RGB	UAV	Semantic labels	China	China
[257]	Misc.	RGB	UAV	Bounding boxes	China	China
[258]	Misc.	HSI	UAV	No	India	Belgium

displays the selected papers that have acquired data but have not published them. The entries are organized based on the frequency of application and then arranged alphabetically.

Figure 18 demonstrates distribution of application of acquired data in the selected papers. The most frequently observed applications are ITCD, followed by ITD, DD, SI, HSA, Misc., GM, CSE, SS, and PM. This prevalence underscores the significance of ITCD and its versatile applications within forestry, highlighting its prominence in the field.

Figure 19 shows the distribution of data sources across the publication year of selected papers. The prevalence of the RGB sensor use case is clearly visible. The next most used sensors are MSI, LiDAR, HSI, PPC, and Thermal.

Figure 20 shows the distribution of platforms carrying the sensors across the number of publications of selected papers. The substantial prevalence of UAVs compared to piloted aircraft and satellites is clearly evident in the data. This dominance signifies a notable shift in sensor deployment methods, showcasing UAVs as the primary choice for data acquisition. Factors such as cost-effectiveness, accessibility, and maneuverability contribute to this marked preference, highlighting the evolving landscape of aerial data collection in forestry-related applications.

Figure 21 presents the distribution of annotation levels across the publication years of selected papers. Notably, the most prevalent annotation type observed is polygons, followed by semantic labels, bounding box, image level, and keypoint annotation. Semantic labels extend beyond the traditional 2D pixel-level annotations, often relating to semantic segmentation tasks that encapsulate not only 2D visual data but also pertain to semantic understanding within 3D data. This emphasizes the multifaceted nature of semantic labeling, addressing both 2D and 3D semantic contexts within the annotation spectrum.

Figure 22 illustrates the frequency of publications by the corresponding author’s affiliated institution country, indicating potential data sources for undisclosed datasets. Notably, it showcases China with 71 instances, followed by Brazil with 12 and United States with 8, signifying the prevalence of potential data reservoirs within these countries.

Figure 23 shows the distribution of countries to which the geographical extent of acquired datasets belong. China with 68 cases is first, followed by Brazil with 12 and United States with 9.

3.3. Repositories

The identification and utilization of data repositories has a crucial role in scientific research by supporting the integrity and accessibility of scholarly information. These repositories, often diverse and distributed across various platforms, hold valuable information generated through thorough research efforts. Their significance is not only in providing a centralized reservoir of data but also in fostering transparency, reproducibility, and the potential for collaborative exploration. The identification and management of these repositories is important to fortify the foundations of academic inquiry, ensuring that valuable datasets are not only preserved but also made readily available for further analysis, verification, and innovation. Within this context,

Table 7. Identified repositories. The “$” sign in front of the repository names indicates the existence of paid-access or restricted-access datasets in that repository. N/A stands for Not Available.

Name	Operator/ Country	Size Limit	Number of Datasets	Establishment Date	Geographical Extent
data.gov.uk	UK	N/A	N/A	2010	UK
Datarade ($)	Germany	N/A	N/A	2018	Worldwide
Direção-Geral do Território ($)	Portugal	N/A	378	N/A	Portugal
DRYAD ($)	Germany	Free individual files: 300 MB	+50,000	2007	Worldwide
EDI Data Portal	United States	N/A	+85,000	N/A	Worldwide
Github $)	United States	Free: 500MB, Paid: 50GB	N/A	2008	Worldwide
Global Forest Watch	WRI	N/A	142	1997	Worldwide
Hugging Face	USA	No limit	70,735	2016	Worldwide
IEEE Dataport ($)	USA	2/10 TB	+8000	2018	Worldwide
Kaggle	USA	100 GB/dataset	260,779	2010	Worldwide
Land use Opportunities-Data Supermarket	New Zealand	N/A	74	2020	New Zealand
LILA BC	USA	N/A	39	N/A	Worldwide
Mendeley Data ($)	Netherlands	Free: 10 GB, Paid: 100 GB	26.9 million	2015	Worldwide
NEON Data portal	United States	N/A	20 regions	2016	USA
ORNL DAAC	United States	N/A	N/A	1994	Worldwide
Official portal for European data ($)	European Union	N/A	1,544,800	2015	Europe
OAM	Community driven	N/A	14,181 images	2007	Worldwide
OpenTopography ($)	United States	N/A	3532	2009	Worldwide
PACIFIC DATA HUB	SPC	N/A	12,333	N/A	Pacific Islands and countries
Pacific Environment Data Portal	SPREP	N/A	18,994	N/A	Pacific Islands and countries
PANGAEA	Germany	15 GB/file	422,680	1987, online 1995	Worldwide
Projeto áGIL—Dados LiDAR	ICNF/Portugal	N/A	7 datasets	2020	Portugal
RNDT	Italy	N/A	20,222	2012	Italy
Roboflow ($)	United States	Free: 10,000 images	+200,000	2019	Worldwide
UC Irvine Machine Learning Repository	United States	N/A	657	1987, 2023	Worldwide
Zenodo ($)	CERN/Europe	50 GB/file	+3,000,000	2013	Worldwide

delineates a compilation of identified repositories, showing their distinctive attributes. A set of parameters is extracted from the identified repositories, namely, the following key aspects:

• Operating Country: Understanding the operational location of a repository sheds light on regional focus, legal considerations, and potential data limitations based on geographical boundaries or regulations.
• Size Limit: This parameter delineates the maximum capacity for data storage within a repository, crucial for users assessing data contribution or access possibilities.
• Number of Datasets: Reflecting the repository’s scale and diversity, a higher count often indicates a broader range of available datasets, appealing to a larger user base.
• Establishment Date: Signifying the repository’s maturity and experience, the creation date offers insights into its reliability and history of dataset curation.
• Geographical Extent: Describing the coverage area of datasets, this parameter assists researchers in evaluating the data’s relevance to specific regions or global applicability.

4. Conclusions

High-resolution aerial datasets are essential in advancing agroforestry research by providing detailed insights into these complex ecosystems. This study conducts an exploration of the availability of such datasets, examining existing and unpublished sources as well as data repositories to show the current landscape of data publication and accessibility. Our analysis reveals a significant increase in dataset availability over the past five years, which correlates with the advancements in machine learning and deep learning methods that require diverse and high-quality data for training and validation. This shows the dynamic interplay between technological development and data availability in research.

Other than the vivid growth in data publication and availability, following is the summary of the most important insights of this study:

• UAVs are the dominant platform when it comes to HR and UHR data: The deployment of UAVs has witnessed a remarkable and ever-growing prominence in the data acquisition phase for forestry-related studies. These aerial platforms offer a versatile and efficient means to gather high-resolution and comprehensive datasets, pushing the envelope of conventional approaches to forest monitoring. UAVs equipped with various sensors, such as RGB, multispectral, hyperspectral cameras, LiDAR, and thermal sensors, provide researchers with a diverse array of data modalities. The ability to navigate through challenging terrains and rapidly capture detailed information about forest ecosystems makes drones indispensable tools in forestry research. Their cost-effectiveness, agility, and capacity to cover vast areas make them ideal for tasks ranging from tree species classification to monitoring forest health and assessing environmental changes over time. As technology continues to advance, the utilization of UAVs in forestry studies is poised to expand further, fostering more precise and comprehensive insights into the dynamics of forest ecosystems.
• There are more unpublished data than published: The visible imbalance between the amount of published and unpublished data uncovered through our study underscores a challenge in forestry-related remote sensing datasets. It becomes evident that a substantial amount of valuable data remains unshared, inaccessible to the wider research community. While the past years have witnessed a growth in the publication of forestry-related remote sensing data, there remains a pressing need for increased efforts to motivate researchers to share their data openly. The prevalence of unpublished data suggests untapped potential and unexplored possibilities for advancing forestry, robotics, and computer vision research. Initiatives promoting data sharing, fostering collaborative platforms, and highlighting the benefits of open data practices are essential to bridge this gap. Encouraging a culture of data transparency and openness is crucial to harness the full potential of available datasets, fostering innovation and contributing to a more comprehensive understanding of forest ecosystems.
• The scarcity of co-aligned datasets: The scarcity of co-aligned data with different modalities is a notable gap in forestry-related datasets. Pursuing a comprehensive understanding of forest ecosystems requires the fusion of data from various sources and modalities. Integrated datasets, encompassing information captured through diverse sensors like RGB, multispectral and hypespectral cameras, and LiDAR sensors, offer a holistic perspective on forest characteristics. Data fusion not only enhances the accuracy of forestry analyses but also enables the extraction of richer insights through the collaborative utilization of complementary information. In machine learning, particularly deep learning and neural networks, the integration of different data modalities becomes imperative. Techniques such as attention mechanisms benefit significantly from diverse data sources, allowing models to focus on relevant features and relationships. Addressing the scarcity of co-aligned data can help advancing research methodologies, optimizing the performance of machine learning algorithms, and unlocking the full potential of data-driven approaches in forestry studies and more.
• Synthetic data have immense potential: Synthetic data emerge as a transformative solution in addressing the complexities and limitations associated with current data acquisition methods in forestry research. By offering a controlled and versatile environment for generating datasets that mirror real-world scenarios, its role extends beyond mere supplementation. The controlled nature of synthetic data allows researchers to manipulate variables such as environmental conditions, tree species distribution, and terrain characteristics, providing a tailored and efficient means to curate datasets. One of its notable contributions is in the generation of annotation masks, which is essential for training machine learning models. Synthetic datasets facilitate the creation of diverse and precisely labeled datasets, streamlining the process of developing robust algorithms for tasks like object detection and classification. This comprehensive simulation extends to various sensors, enhancing the applicability of synthetic data in different modalities. As technology continues to advance, the integration of synthetic data into forestry research methodologies seems very promising in propelling the field toward more accurate and comprehensive insights into forestry. However, it is crucial to acknowledge that synthetic data are still in their early stages, requiring substantial further development and refinement for them to become a comprehensive and widely adopted solution in forestry research.