Deep Learning (DL) has fundamentally changed Computer Vision (CV), enabling machines to interpret visual data with precision that often equals or surpasses human performance (Abd Zaid et al. 2025; Agduma and Altarez 2025; Ahmed et al. 2023). The rapid development of DL frameworks—from Convolutional Neural Networks (CNNs) (Ali and Khati 2024; Alzubaidi et al. 2021) to Vision Transformers (ViTs) (Dosovitskiy et al. 2021; Khan et al. 2025), object-detection transformers (Amitrano et al. 2023; Arkin et al. 2022; Arun et al. 2025; Ayrey and Hayes 2018), and generative models (Ball et al. 2017; Beal et al. 2020)—has greatly expanded CV’s reach across scientific domains, particularly forestry. This proliferation is underpinned by broader reviews of deep learning theory (Canu 2017; Zhu et al. 2017) and by architectural innovations such as graph attention networks (Veličković et al. 2018) and end-to-end set-based detectors (Carion et al. 2020; Zhu et al. 2020). Dense-prediction and deep image transformers (Ranftl et al. 2021; Touvron et al. 2021; Peng et al. 2021), transformer-based detection frameworks (Misra et al. 2021; Vaidwan et al. 2021; Ma et al. 2021), and classical anchor-based detectors (Redmon and Farhadi 2018) have been comprehensively benchmarked in recent surveys of object-detection evolution (Sun et al. 2024). Generative adversarial networks (Karras et al. 2020) and transformer-based image restoration (Song et al. 2023) further extend this methodological toolkit.

Effective forest management relies on accurate, timely inventory data for sustainable stewardship (Bertrand et al. 2017; FAO 2020), building on a long tradition of remote-sensing applications to forest ecology and LiDAR-based sampling reviews (Lechner et al. 2020; Wulder et al. 2012). Traditional manual methods are labour-intensive, costly, and subjective (Aijazi et al. 2017; Beyene 2020). Consequently, automated systems are urgently needed to scale to complex ecological questions—such as monitoring climate-induced canopy defoliation, mapping invasive species, and quantifying aboveground biomass for global carbon cycle modeling (Zhang et al. 2015).

Despite significant progress, three fundamental challenges remain unaddressed. First, the lack of standardized evaluation protocols prevents meaningful cross-study comparisons; only 23% of reviewed studies report consistent train-test splits or confidence intervals. Second, the ‘site-specificity problem’ persists: models trained on temperate forests show accuracy degradation of 20–45% when applied to tropical ecosystems without fine-tuning (Narine et al. 2019; Weinstein et al. 2019). Third, the computational-accuracy paradox creates deployment barriers. State-of-the-art models require 16–24 GB of GPU memory, prohibiting edge deployment on UAVs or field devices (El Asmar and Felfla 2025).

Prior reviews surveyed tree species classification (Fassnacht et al. 2016), CNNs for plant ecology (Kattenborn et al. 2021), biomass estimation (Li et al. 2022a, b, c), and hyperspectral classification (Gong et al. 2025). The present review extends this body of work by: (1) covering a broader temporal arc (2011–2026) and architectures including ViTs and Graph Neural Networks (GANs); (2) quantifying cross-biome model transferability; (3) profiling computational complexity; and (4) proposing an evidence-informed practitioner decision framework for hardware-constrained field deployment.

Evolution of tree monitoring methods: from manual to automated approaches

Historically, tree monitoring relied on labour-intensive field surveys susceptible to human variability. The advent of remote sensing initiated a shift towards automation (Zhang et al. 2015). Early automation utilized rule-based algorithms or traditional machine learning, which struggled with complex forest structures and occlusion. The emergence of DL has substantially advanced automation, enabling highly accurate, scalable Individual Tree Detection (ITD) and characterisation across diverse landscapes (Hu and Li 2020; Weinstein et al. 2019).

Deep learning architectures for tree detection and classification

DL models process complex visual data by learning hierarchical features directly from raw imagery or point clouds. CNNs and their variants are fundamental for object detection and image classification (Voulodimos et al. 2018). For example, RetinaNet has been adapted for palm tree detection (Culman et al. 2020), while ResNet detects diseased trees from UAV imagery (Deng et al. 2020) and DNN autoencoders monitor environmental anomalies (Abdurakipov and Butakov 2019). Semantic segmentation frameworks like U-Net delineate tree crowns, while object detection frameworks like Faster R-CNN and YOLO identify and track specific targets, from landmines to diseased pine, in complex aerial scenes (Baur et al. 2020; Lai and Huang 2020; Opromolla et al. 2019). CNN backbones validated on visually analogous classification tasks such as animal recognition (Chen et al. 2014; Kuppusamy and Naga Chaitanya 2025) and texture-based species discrimination (Sharma and Krishna 2019) demonstrate the transferability of these architectures to forestry problems, as do flower and fruit classification pipelines (Liu 2024; Xianchong and Yuanyuan 2021; Noris and Waluyo 2023). Within forestry specifically, CNNs have been applied to wood- and tree-species identification from macroscopic and leaf imagery (Minowa and Nagasaki 2020; Oktaria et al. 2019) and generic forest-scene classification (Mardiyah and Purwaningsih 2020), while hierarchical tree-structured learning (Roy et al. 2020; Liu et al. 2018), asymmetric transfer-learning networks (Shi et al. 2021a, b), and hyperspectral-specific CNNs (Paoletti et al. 2018) extend fine-grained discrimination capacity. UAV-based computer vision has also supported single-tree detection from LiDAR (Balsi et al. 2018) and RGB imagery (Santos et al. 2019), pest surveillance (Roosjen et al. 2020), structural defect identification (Bhowmick et al. 2020), landing-zone human detection (Safadinho et al. 2020), and integrated anti-poaching/fire-monitoring systems (Pal et al. 2024).

Recent benchmark datasets are transforming evaluation by enabling consistent, cross-biome, and cross-sensor comparisons (Table 1), demonstrating that cross-sensor pre-training improves out-of-distribution generalisation (Mowla et al. 2025, 2024).

Multi-modal data fusion techniques

Fusing 3D LiDAR point clouds with optical imagery enriches structural data with spectral characteristics. Architectures like 3D CNNs and PointNet +  + effectively process these multi-modal inputs for tasks like individual tree detection and species classification (Ayrey et al. 2019; Briechle et al. 2020). Combining LiDAR with Synthetic Aperture Radar (SAR) or hyperspectral imagery maintains robustness even under dense canopy cover or adverse atmospheric conditions (Zhang et al. 2015). Incorporating near-infrared (NIR) and thermal infrared (TIR) channels offers distinct advantages: NIR assesses canopy water status and improves species classification accuracy by 3–8 percentage points over visible-spectrum approaches (Dalponte and Coomes 2016; Farella et al. 2022; Seely et al. 2025), while TIR maps heat stress and drought-induced decline before visible symptoms emerge (Farella et al. 2022; Zakrzewska et al. 2022). Despite these benefits, NIR and TIR remain underrepresented in the literature. Multi-sensor biomass retrieval has additionally been pursued through PolSAR water-cloud modelling (Kumar et al. 2019), L- and P-band SAR fusion (Schlund and Davidson 2018), airborne L-band SAR machine learning (Ramachandran and Dikshit 2022), and multifrequency SAR-optical fusion (Ozdemir and Abdikan 2025). Optical and multispectral pipelines include Sentinel-2 bamboo biomass mapping (Chen et al. 2018), Landsat feature-engineering comparisons (Kilbride and Kennedy 2024), national-scale multi-source mapping in China (Tang et al. 2022), and deep-learning estimation in the Hangzhou region (Tian et al. 2024). Further contributions span ICESat/Landsat fusion (Chi et al. 2017), global LiDAR-optical fusion (Hu et al. 2016), LiDAR-based attribute imputation (Silva et al. 2016; Li et al. 2018), and TLS-PALSAR integration (Singh et al. 2023). Comprehensive machine-learning and deep-learning biomass syntheses (Talebiesfandarani and Shamsoddini 2022; Utla et al. 2025) and model-based inference frameworks (Chen et al. 2016) contextualise these advances, alongside artificial neural network estimation over alpine and karst landscapes (Wang et al. 2018; Qian et al. 2021) and RADAR-based retrieval in tropical forests (Espinoza-Mendoza 2018). Additional regional studies cover pine biomass ANN modelling (Özçelik et al. 2017), larch plantation upscaling (Zhen et al. 2022), GNSS-reflectometry retrieval (Pilikos et al. 2024), NEON/GEDI density estimation (Mahaur 2024), GEDI multi-source fusion (Mohite et al. 2024), and Khingan Mountains machine-learning fusion (Wang et al. 2022). Recent contributions further include Myanmar Sentinel-2 geostatistical modelling (Wai et al. 2022), subtropical canopy-height fusion (Wu et al. 2025), mixed-temperate multimodal estimation (Lamahewage et al. 2025), Sentinel-1 backscatter retrieval (Leena et al. 2024), atom-search-optimised neural estimation (Chen et al. 2023), and machine-learning carbon-stock mapping (Cheng et al. 2024). Global gridded biomass fusion (Zhang and Liang 2020), eight-algorithm regression benchmarking (Zhang et al. 2020a, b), tropical RNN-based estimation (Zhang and Zhuo 2024), and UAV-to-satellite biomass integration (Ma et al. 2025) extend this evidence base across biomes and sensor combinations. Pasture-specific estimation has likewise advanced through UAV RGB deep learning (Vahidi et al. 2023) and multimodal computer vision frameworks (Ugwumba 2025), while mangrove species classification (Vellasamy et al. 2025), wireless-sensor forest ecological monitoring (Ying 2020), and machine-learning biomass retrieval pipelines (Kumari and Kumar 2023) round out the diversity of fusion-based applications.

Aim and objectives

The aim of this systematic review was to critically evaluate the evolution, performance, and operational constraints of deep learning and computer vision technologies in forest monitoring, providing a translational framework that bridges the "deployment gap" between computational research and applied ecological engineering for sustainable ecosystem management.

To achieve this aim, the study addressed the following specific objectives:

• To synthesize the temporal and methodological evolution of deep learning applications in forestry (2011–2026), mapping the structural dependencies between specific remote sensing platforms (e.g., UAVs, LiDAR, hyperspectral, satellites) and their downstream ecological tasks.

• To quantify model performance through statistical meta-analysis, comparing the efficacy of traditional Convolutional Neural Networks (CNNs) against emerging architectures like Vision Transformers (ViTs) and GNNs across critical conservation tasks, including species classification, biomass estimation, and early-stage disease detection.

• To evaluate the cross-biome transferability and operational limitations of current models, specifically measuring the transferability paradox (performance degradation across distinct ecological zones) and assessing the computational trade-offs that restrict real-time edge deployment.

• To formulate actionable decision frameworks for ecological practitioners, including an interpretability-performance-efficiency ternary model, a research prioritization matrix, and a structured deployment protocol that aligns hardware constraints with the logistical realities of field-based habitat reconstruction and monitoring.

Protocol and registration

This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement (Page et al. 2021; Moher et al. 2015). The review protocol was developed and registered a priori (available at: https://osf.io/zb8eh), specifying eligibility criteria, search strategy, data extraction fields, and quality-assessment domains. The Population, Intervention, Comparison, Outcomes (PICO) framework was adapted as: Population = forest ecosystems monitored by remote sensing; Intervention = deep learning (DL) or computer vision (CV) model; Comparison = across DL architectures, sensor platforms, and biomes; Outcomes = quantitative performance metrics (accuracy, F1-score, Root Mean Square Error (RMSE), coefficient of determination (R²), mean average precision (mAP)) and computational efficiency indicators (FLOPs, inference latency, parameter count). The scope encompasses peer-reviewed articles, conference proceedings, and verified early-access publications issued between January 2011 and December 2025, inclusive of 2026-dated studies with stable digital object identifiers (DOIs) or arXiv identifiers captured prior to database freeze.

Eligibility criteria

Studies were included if they: (1) applied DL/CV to forestry; (2) utilized remote sensing data; and (3) reported quantitative performance metrics. Criteria were operationalised using the Population–Concept–Context framework (Aromataris & Munn 2020). Studies relying solely on traditional machine learning, lacking methodological detail, or reporting only simulated outputs without empirical validation were excluded. Review articles (n = 7) were retained for qualitative synthesis only.

Information sources and search strategy

Four electronic databases were searched: Scopus, Web of Science (Core Collection), IEEE Xplore, and Google Scholar. Searches were executed in January 2026 using combinations of controlled vocabulary and free-text keywords covering three thematic pillars: algorithmic frameworks, application domains, and sensor technologies. The primary Boolean search string was: (“deep learning” OR “convolutional neural network” OR “computer vision”) AND (“forestry” OR “forest management” OR “tree monitoring”) AND (“remote sensing” OR “LiDAR” OR “UAV” OR “hyperspectral”). Secondary iterative searches and citation snowballing employed supplementary terms such as “crown segmentation,” “sparse convolution,” and “foundation models” to capture specialised sub-fields not covered by the primary string (Fassnacht et al. 2016; Gong et al. 2025). No language or date restrictions were imposed at the search stage; temporal and language filters were applied at the screening stage in accordance with the eligibility criteria.

Study selection and deduplication

A systematic search identified 238 records. After deduplication and screening, 190 studies (183 primary, 7 reviews) were included in the final synthesis. The PRISMA flow diagram illustrating selection is provided in supplementary Figure S1.

Data extraction

Data extracted into standardized spreadsheets using Python 3.12 included: study ID, sensor platform, data modality, DL architecture family, primary application domain, quantitative performance outcomes, training dataset size, computational characteristics, and biome context.

Quality assessment and risk-of-bias evaluation

Methodological quality was evaluated across seven domains adapted from QUADAS-2 (Whiting et al. 2011): (1) dataset representativeness; (2) data leakage risk; (3) hyperparameter transparency; (4) validation rigour; (5) reporting completeness; (6) class-imbalance handling; and (7) interpretability provision. Domains were rated low risk, some concerns, or high risk.

Statistical synthesis and meta-analysis

Quantitative synthesis was conducted using random-effects meta-analysis models with Restricted Maximum Likelihood (REML) and DerSimonian–Laird (DL) estimators to account for between-study heterogeneity (DerSimonian & Laird 1986; Higgins et al. 2003). Weighted mean performance metrics were calculated using inverse-variance weighting:

Prior to pooling, bounded metrics (accuracy, F1-score) underwent logit transformation to stabilise variance, and were back-transformed for reporting. Disparate metric types were not averaged together; meta-analyses were stratified by metric family (accuracy for classification; mAP@IoU for detection; R² for regression) as illustrated in the stratified forest plots (supplementary Figure S2). Between-study heterogeneity was assessed using I² (the proportion of total variance due to between-study variation) and τ² (the estimated between-study variance) statistics. For the 29 studies lacking reported per-study variance, standard errors were imputed via bootstrap resampling (1,000 iterations) from cross-validation folds or confusion matrices (supplementary Figure S3). A trim-and-fill sensitivity analysis (Duval & Tweedie 2000) was applied because Egger’s regression test (Egger et al. 1997) detected significant funnel-plot asymmetry (p = 0.0087), indicating the presence of a small-study effect (supplementary Figure S4). All procedures were implemented in Python 3.12 using the meta-analysis and scipy libraries. Cumulative meta-analysis was performed by iteratively pooling studies in chronological order to evaluate the temporal stability of pooled estimates. To improve transparency of the quantitative synthesis, metric harmonization proceeded in three steps: (1) all bounded metrics (accuracy, F1-score, sensitivity, specificity) underwent logit transformation prior to pooling and were back-transformed for reporting; (2) analyses were stratified strictly by metric family—classification accuracy, detection mAP@IoU, and regression R²—and these families were never averaged together; and (3) subgroup definitions (architecture family, biome, sensor modality) were pre-specified in the registered OSF protocol and applied without post-hoc modification. The number of contributing studies, heterogeneity statistics (I², τ²), and imputation approach for each pooled analysis are summarized in Sect. "Meta-analysis of model performance across applications".

Cross-biome transferability quantification

To evaluate model generalisation beyond the biome providing the training data, a symmetric cross-biome transferability score (T) was computed following the Symmetric Percentage Difference metric (Makridakis 1993). This formulation avoids the artificial exaggeration of domain shift that occurs with asymmetric normalisation when source-domain accuracy is low:

A score of T = 1.0 indicates perfect transferability; values approaching zero reflect severe domain-shift degradation. Source–target asymmetry was explicitly isolated to compare degradation directionality (e.g., temperate → tropical vs. tropical → temperate). Effect sizes with 95% confidence intervals are presented in supplementary Figure S5, alongside a domain-adaptation benefit curve quantifying the marginal accuracy recovery achieved by feature-level and full fine-tuning strategies.

Computational complexity and data visualisation

A computational complexity–performance trade-off analysis mapped empirical FLOPs, peak GPU memory, and inference latency against accuracy across 186 studies. Network co-occurrence graphs were constructed using sensor–task adjacency matrices. All visualisations were produced in Python 3.12.

Trends in deep learning applications

The 15-year review period (2011–2026, n = 190) reveals a rapid increase in DL forestry publications (Li et al. 2022a, b, c; Li 2025; Li et al. 2019), delineated into three phases: Early Adoption (2011–2018), Consolidation (2019–2021), and Diversification (2022–2026) (Fig. 1a & b). Early output was incremental, whereas 2020 marked a pronounced surge (n = 34) driven by foundational architectures (ViT, DETR, Mask R-CNN). Biomass & Carbon Estimation became the dominant task post-2022. Conversely, Forest Inventory & Tree Detection declined after 2020, suggesting robust computational baselines have been reached. Species & Biodiversity Classification rebounded in 2025, driven by transfer-learning (Gregoire et al. 2016; He et al. 2017; Heinzel and Koch 2011; Hidayat et al. 2022) (Fig. 1c).

The sensor–task heatmap (Fig. 1d) reveals distinct specialisations. Airborne LiDAR (ALS) dominates (Σ = 123), pairing strongly with Biomass Estimation (n = 28). Multi-source Fusion (Σ = 105) and UAV RGB/Multispectral (Σ = 104) follow. Conversely, SAR/InSAR is highly task-specific (Σ = 56) for Biomass and Change Detection, whereas Airborne Hyperspectral leads Species Classification. Notably, Disease & Pest Detection remains under-served by sensors like ALS and SAR, highlighting a critical operational gap.

Sensor breadth and dominance

Airborne LiDAR (ALS) recorded the highest row total (Σ = 123), confirming its status as the most extensively studied platform in the corpus. Its strongest associations are with Biomass & Carbon Estimation (n = 28, the single highest value in the entire matrix, highlighted in the figure) and 3D Structure & Volume (n = 22), consistent with ALS’s well-established capacity for canopy height and structural characterisation. Multi-source Fusion (Σ = 105) and UAV RGB/Multispectral (Σ = 104) rank second and third overall, with Fusion notable for its relatively even distribution across all eight tasks — no cell falling below 8 — reflecting its role as a complementary strategy deployed wherever a single sensor proves insufficient. Terrestrial laser scanning (TLS) recorded the lowest sensor total (Σ = 51), concentrated almost exclusively in 3D Structure & Volume (n = 15) and Crown Segmentation (n = 8), confirming its niche as a ground-truth and structural measurement instrument rather than a scalable operational sensor.

Task demand and sensor preference

Biomass & Carbon Estimation attracted the highest task-level demand across the corpus (Σ = 124), with ALS (n = 28), Optical Satellite/Landsat/Sentinel (n = 20), and Multi-source Fusion (n = 18) as its three dominant sensor pairings. Individual Tree Detection ranked second (Σ = 103), uniquely distinguished by the joint leadership of UAV RGB/Multispectral (n = 24) and UAV LiDAR (n = 23) — the two orange-highlighted cells in the figure — a pattern that reflects the fine spatial resolution and point-cloud density required for individual-crown delineation. Disease & Pest Detection registered the lowest task total (Σ = 57), with UAV RGB/Multispectral as its clear dominant sensor (n = 18, more than double any other sensor for this task), which is consistent with the use of high-resolution visible-band imagery for canopy discolouration and symptom detection at the individual-tree scale.

Sensor specialisation versus generalism

SAR/ Interferometric Synthetic Aperture Radar (InSAR) displays the most concentrated and task-specific profile of any sensor (Σ = 56): nearly two-thirds of its co-occurrences fall across just two tasks — Biomass & Carbon Estimation (n = 16) and Change Detection (n = 14) — while recording near-zero contributions to Individual Tree Detection (n = 2), Crown Segmentation (n = 2), and Disease Detection (n = 2), tasks that demand spatial resolutions below what spaceborne SAR can reliably provide. Airborne Hyperspectral, by contrast, is the sole sensor for which Species Classification constitutes the primary task (n = 18, the column leader), aligning with its unique capacity to discriminate species through fine spectral resolution, while also contributing meaningfully to Biomass (n = 14) and Disease Detection (n = 10). Optical Satellite imagery occupies a complementary specialisation, co-leading Change Detection alongside SAR (n = 20 each) and ranking second in Biomass Estimation (n = 20), reflecting its temporal revisit frequency and synoptic coverage — advantages that are structurally suited to landscape-scale monitoring rather than fine-grained structural tasks, as evidenced by its near-absent contribution to Crown Segmentation (n = 4).

Structural gap

The most conspicuous structural absence in the matrix is the consistently low engagement of all sensors with Disease & Pest Detection (column total Σ = 57, the lowest of any task), and particularly the near-zero values for ALS (n = 3), TLS (n = 2), and SAR (n = 2) in this column. This pattern suggests that automated disease and pest surveillance remains under-served by the full range of available sensor technologies, and represents a methodological gap with direct implications for operational forest health monitoring systems.

The map in supplementary Figure S6 illustrates the geographic concentration, showing the highest densities in East Asia, North America, and Western Europe. In contrast, Africa and most of Central and South America are markedly underrepresented, despite hosting the majority of the world’s tropical forest biomes. This mismatch not only limits the generalisability of models but highlights an ethical imperative for capacity-building, open data governance, and equitable co-authorship with institutions managing these biodiversity hotspots.

Data types and sensors

The Sankey diagram (Fig. 2) maps the flow between data sources, architectures, and applications. CNNs are overwhelmingly dominant, feeding into all major domains. Emerging architectures (Transformers, GNNs) show thinner but growing flows for complex dependencies. Structurally, LiDAR flows heavily into Biomass Estimation for 3D vertical data, while Hyperspectral and UAV RGB couple with Species Classification and Disease Detection for fine spectral detail. Satellite imagery bridges the scale gap for large-area monitoring.

UAVs provide high-resolution, flexible data for Individual Tree Detection (ITD) and structural measurements (Bayat et al. 2019; Wallace et al. 2012; Yan et al. 2020; Picos et al. 2020; Barbedo et al. 2019; Ammour et al. 2017). Conversely, satellite and airborne LiDAR offer broad-area coverage for regional inventories (Ayrey et al. 2019; Shinohara et al. 2020). LiDAR remains invaluable for 3D segmentation and height measurement (Hatamizadeh et al. 2023; Huy et al. 2022; Immitzer et al. 2012; Mukhopadhyay et al. 2023; Naderpour et al. 2021). UAV RGB/multispectral cameras democratize local precision forestry (Kamilaris and Prenafeta-Boldú, 2018; Naik et al. 2023, 2021; Narine et al. 2019), while hyperspectral sensors capture subtle biochemical differences for fine-grained classification (Kangas et al. 2018; Narine et al. 2019; Nezami et al. 2020; Ningthoujam et al. 2017). Tables 2 and 3 summarize these relationships.

Meta-analysis of model performance across applications

Performance metrics (Fig. 3, supplementary Figure S2) were harmonized prior to pooling. For tree species classification (n = 44), the pooled mean accuracy is 91.2% (95% CI: 89.7–92.7%; I² = 76%, p < 0.001). Subgroup analysis shows 3D-CNNs (94.3%) significantly outperform 2D-CNNs (88.7%, p = 0.002) on fused LiDAR-hyperspectral data. ViTs achieved the highest pooled estimate (96.1%), though wider confidence intervals from a smaller sample (n = 8) necessitate independent replication.

For biomass estimation (n = 34), meta-regression yields a pooled R² = 0.80. Sensor modality explains 43% of between-study variance (p < 0.001), with LiDAR + hyperspectral fusion achieving R² = 0.90. RMSE decreases logarithmically with training sample size (R² = 0.67, p < 0.001):

where n represents the number of training samples, suggesting diminishing returns beyond ~ 5,000 training samples.

Disease detection (n = 34) shows a pooled sensitivity of 87.4% and specificity of 94.6%. Early-stage detection yields significantly lower sensitivity than late-stage (78.3% vs. 94.6%, p < 0.001) (Deng et al. 2020; You et al. 2021). Sensitivity analyses confirmed the stability of these estimates (supplementary Figures S3 and S4).

Deep learning architectures

Across 148 architecture-reporting studies (Fig. 4), ViTs achieve the highest weighted mean accuracy (μ = 94.1%), suitable for large-scale, multi-modal mapping (Amitrano et al. 2023; Arkin et al. 2022; Arun et al. 2025; Chiang et al. 2020; Clark et al. 2011). 3D-CNNs rank second (μ = 92.3%), followed by ResNet/EfficientNet (μ = 89.8%) and U-Net variants (μ = 88.8%). CNNs remain the dominant architecture (52.7%), processing 2D imagery and voxelised LiDAR (Ding et al. 2025; He et al. 2017; Kamilaris and Prenafeta-Boldú, 2018; Naik et al. 2021; Narine et al. 2019). The remainder includes ViTs (12.2%), Generative Adversarial Networks (GANs, 4.1%) for synthetic data (Ball et al. 2017; Beal et al. 2020), Graph Neural Networks (GNNs, 1.4%) (Dosovitskiy et al. 2021), and other variants (29.7%). Fig. 6 annotates GANs (Castonena 2025; Liu 2025) achieving μ = 85.7% for data scarcity mitigation.

Architecture evolution (CNN vs. Transformers)

The shift from CNNs to ViTs marks a significant paradigm change (Table 4). CNNs utilize local convolution kernels for spatial hierarchy (Krizhevsky et al. 2012; LeCun et al. 2015), ideal for localized pattern recognition like leaf diseases (Arun et al. 2025; Saengpetch et al. 2025; Hafemann et al. 2014; Shorten and Khoshgoftaar 2019). Conversely, Transformers use self-attention for global context (Dosovitskiy et al. 2021; Khan et al. 2025), excelling in complex scene integration (Ding et al. 2025; Weber et al. 2025) but requiring massive datasets due to a lack of inductive bias (Li et al. 2023; Tu et al. 2022) and higher computational costs (Alzubaidi et al. 2021; Strudel et al. 2021; Ma et al. 2019; Meng et al. 2024).

Figure 5 illustrates the three-way trade-off between accuracy, efficiency, and interpretability across 17 architectures. Tier 1 comprises research-oriented transformers (ViT-Large, EarthView, MAESTRO). Tier 2 includes operationally balanced CNNs (EfficientNet, ResNet-50). Tier 3 occupies the high-efficiency end and contains edge-optimised architectures—MobileNet-V3, MobileNet-V3 (INT8 quantized), and GNN (PointNet + +)—that sacrifice peak accuracy in favour of real-time inference on resource-constrained UAV and field hardware (see supplementary Figure S7 for empirical FLOPs and latency measurements across named architectures).

Performance benchmarking

DL models frequently exceed traditional statistical methods (Table 5). For species classification, fusing hyperspectral and RGB imagery with 3D-CNNs yields accuracies up to 99.6% (Nezami et al. 2020), as 3D-CNNs preserve volumetric features better than 2D-CNNs (Maschler et al. 2018; Li et al. 2013). DL also excels in single-tree segmentation (Hu and Li 2020), palm detection (Culman et al. 2020; Miyoshi et al. 2020), and stem reconstruction from LiDAR (Windrim and Bryson 2020). Biomass estimation using fused Sentinel-2 and LiDAR achieves high R² values (Zhu and Liu 2019; Li et al. 2022a, b, c; Agduma and Altarez 2025; Ding et al. 2025) with low RMSE (Ayrey and Hayes 2018). Forest health monitoring has seen similar gains, with models achieving high precision in disease and smoke detection (You et al. 2021; Kim and Muminov 2023; Liu et al. 2021; Chiang et al. 2020; Mohanty et al. 2016).

Table 6 highlights architectural trade-offs. ViT-Large achieves highest accuracy but requires > 30 TFLOPs and 16–24 GB VRAM. EfficientNet-B5 provides optimal operational balance, while MobileNet variants operate within < 100 ms latency for edge computing. The full FLOPs-latency-accuracy trade-off surface across 15 model families is provided in supplementary Figure S6.

Model transferability across ecological zones

Table 7 quantifies the "transferability paradox" across 18 studies (Supplementary Figure S4). Temperate-to-tropical transfers exhibit a − 38.4% (± 9.2) accuracy degradation—3.1 times greater than within-climate transfers—due to species diversity, canopy complexity, and phenological differences (FAO 2020). Domain adaptation (DA) mitigates this: feature-level adaptation reduced the accuracy drop to 22.1%, and full fine-tuning on just 10% of target data limited the deficit to 14.7%.

Temporal analysis of methodological evolution

Figure 6 tracks technological adoption. Early Adoption (2011–2018) relied on CNNs with manual feature engineering. Consolidation (2019–2021) saw widespread transfer learning (van Geffen et al. 2021). Diversification (2022–2026) features ViTs and self-supervised foundation models like MAESTRO (Labatie et al. 2025), EarthView (Velazquez et al. 2025), and WHU-STree (Ding et al. 2026). Accuracy improved by 1.8% annually (p < 0.001) alongside dataset and complexity growth. However, accuracy gains decelerated to + 0.7% annually in 2019–2025, suggesting diminishing returns against sensor limits. Inference times highlight the deployment challenge, particularly for generative models (Castorena et al. 2025).

Challenges, impacts, and mitigation strategies

Several challenges hinder operational deployment (Table 8). Model generalisability across ecosystems requires extensive data (Weinstein et al. 2019; Jiang et al. 2020; Høye et al. 2020). Synthetic data (van Geffen et al. 2021; Liu et al. 2025; Castorena et al. 2025; Tao et al. 2022; Ball et al. 2017) and weighted loss functions (Lin et al. 2020; Shorten and Khoshgoftaar 2019) help mitigate data scarcity and class imbalance. Model interpretability remains poor; only 11% of studies used Explainable AI (XAI) like Grad-CAM (Ahmed et al. 2023; Alzubaidi et al. 2021; Mienye and Swart 2024). High computational costs restrict edge deployment, necessitating model pruning and TinyML (El Asmar and Felfla 2025; Noor and Ige 2024; Abd Zaid et al. 2025; Maschler et al. 2018).

Figure 7 maps the identified research gaps onto a practical priority framework. The left panel evaluates five key research directions using grouped horizontal bars to represent Urgency, Feasibility, and Field Impact on a 10-point scale. A composite priority score (amber diamond) determines the recommended implementation sequence (#1 to #5), with a dashed vertical line at a score of 7 marking the threshold for immediate operational investment.

Uncertainty Quantification emerges as the highest-ranked immediate priority (#1). While its urgency and impact are moderate (scored at 6), its exceptionally high feasibility (9) makes it an ideal initial priority. As shown in the right panel, it requires only an estimated 6 months and a $0.1 million investment to address the 92% of studies that currently provide point estimates without confidence intervals (e.g., via Monte Carlo Dropout implementation). These timeline and investment estimates are informed by comparable research infrastructure costs reported in the domain-adaptation and benchmark-development literature, and should be treated as indicative figures rather than precise projections.

Standardised Benchmarking (2) and Domain Adaptation (#3) form the core structural priorities. Benchmarking scores highest in both urgency (9) and impact (9), demanding immediate action to resolve the 77% of studies relying on incomparable proprietary datasets; this requires a sustained 24-month timeline and $0.8 million investment. Domain Adaptation closely follows, specifically targeting the 20–45% cross-biome accuracy degradation with an 18-month, $0.7 million research requirement.

Explainable AI Integration (#4; 12 months, $0.5 M) and Multi-Task Learning Frameworks (#5; 30 months, $0.8 M) round out the portfolio. Although Multi-Task Learning offers high field impact and urgency (scored at 8), its lower feasibility (5) pushes it to the end of the implementation sequence, framing it as a longer-term strategic endeavour rather than an immediate fix. The right panel demonstrates that this comprehensive, five-step portfolio requires a cumulative investment of $2.9 million over a 30-month horizon. Executing this sequence offers a highly structured, cost-effective pathway to transition deep learning in forest monitoring from its current 1.7% operational deployment rate to widespread, production-ready use.

Risk of bias

Risk-of-bias summaries are presented in Fig. 8. Across all 190 studies, only 27% (n = 50) were rated low risk across all evaluated domains. This finding is consistent with the benchmarking deficit identified earlier and suggests that reported accuracy values should be interpreted cautiously rather than as definitive operational benchmarks. The modified QUADAS-2 assessment reveals distinct vulnerability patterns across the different architecture groups. Statistical validation rigor consistently emerges as the most problematic domain across the entire corpus.

For instance, 3D-CNN studies exhibit a severe 22 percent high-risk rate in this specific category. Similarly, hyperparameter transparency remains broadly categorized as unclear or high-risk across most models. Convolutional Neural Networks and Vision Transformers demonstrate relatively better dataset representativeness, with approximately half of those studies scoring low-risk in that domain. Nevertheless, the pervasive methodological opacity regarding train-test contamination and statistical validation significantly undermines the real-world reliability of the reported accuracy metrics.

Through a critical evaluation of the evolution, performance, and operational constraints of deep learning and computer vision in forest monitoring, this systematic review establishes a translational framework aimed at bridging the deployment gap between computational research and applied ecological engineering for sustainable ecosystem management.

Theoretical implications: from feature engineering to end‑to‑end learning

The transition from hand‑crafted features to fully end‑to‑end deep learning (DL) architectures represents a substantial change in how forestry remote sensing is conducted. This review quantifies this change: early studies (2011–2014) that relied on manual feature extraction and rule‑based algorithms (Heinzel and Koch 2011; Gleason and Im 2012) achieved lower baseline accuracies than modern approaches. In contrast, end‑to‑end architectures (2015–2025), such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), have shown a significant performance leap (LeCun et al. 2015; Krizhevsky et al. 2012; Dosovitskiy et al. 2021). For example, where traditional methods struggled with spectral‑spatial decoupling, 3D‑CNNs effectively learn volumetric feature representations directly from hyperspectral cubes and LiDAR point clouds (Nezami et al. 2020; Ayrey and Hayes 2018). Connecting these empirical findings to ecological modelling theory reveals an important caveat. Spatial autocorrelation among training and test samples—a structural feature of geospatial data known to inflate apparent model accuracy in geographic machine learning (Roberts et al. 2017)—is rarely accounted for in the reviewed studies. Standard random train‑test splits do not respect spatial independence, meaning that pooled accuracy estimates may partly reflect proximity-based interpolation rather than true ecological generalisation. The random-effects model applied in this meta-analysis accounts for between-study heterogeneity but cannot correct for within-study spatial dependence where this is unreported. Future studies should employ spatially blocked cross-validation to produce accuracy estimates better aligned with the principles of spatially explicit ecological modelling and species distribution forecasting.

Methodological advances and persistent limitations

The pooled performance estimates presented in Sect. "Results" must be contextualised against the widespread risk-of-bias. With a substantial proportion of studies exhibiting high risk for statistical validation rigour and lacking hyperparameter transparency, many high-performing results cannot be independently reproduced. These quality concerns do not invalidate the overall findings, but they underscore that reported accuracies represent an upper-bound estimate of true operational performance.

Training data requirements

The analysis reveals a logarithmic relationship between training sample size and performance saturation. While deep learning is inherently data-hungry (LeCun et al. 2015), massive datasets are typically required for ViTs to outperform CNNs due to their lack of inductive bias (Tu et al. 2022; Li et al. 2023). Because annotation costs for complex forest scenes remain prohibitive (Ball et al. 2017), Generative Adversarial Networks (GANs) and self-supervised learning (SSL) have emerged as viable solutions for synthetic data augmentation, creating realistic imagery to balance rare classes such as diseased trees with fewer labeled samples (Tao et al. 2022; Ball et al. 2017; Shorten and Khoshgoftaar 2019).

The transferability paradox

A critical finding is the asymmetry in model transferability across ecological zones. Studies indicate that models trained on high-diversity tropical datasets transfer more effectively to low-diversity temperate forests than vice versa (Narine et al. 2019; Weinstein et al. 2019). We hypothesize that the structural and phenological complexity inherent to tropical biomes may encourage neural networks to learn more generalizable representations of tree architecture. Conversely, applying temperate-trained models to tropical environments results in significant accuracy degradation (Narine et al. 2019). This highlights that a “universal model” remains elusive, requiring training strategies that account for diverse spectral and structural phenology (Weinstein et al. 2019; Abd Zaid et al. 2025).

Operational deployment framework: bridging research and practice

Despite high quantitative performance on retrospective test sets, a significant deployment gap persists. This disconnect largely stems from the prohibitive computational requirements of state-of-the-art architectures (ViTs, 3D-CNNs), which demand substantial GPU memory unsuitable for real-time edge deployment on UAVs or handheld devices (El Asmar and Felfla 2025; Noor and Ige 2024).

Beyond computational constraints, the high cost of acquiring coincident multi-sensor data limits large-scale, continuous monitoring (Kangas et al. 2018; Vafaei et al. 2018). Furthermore, models trained on pristine imagery often suffer performance degradation under variable lighting and atmospheric occlusions (Maschler et al. 2018; Weinstein et al. 2019). Addressing these challenges requires a shift toward “TinyML” architectures (e.g., MobileNet, quantized CNNs), optimizing trade-offs to enable viable edge computing solutions (Noor and Ige 2024). Operationally, DL outputs can be projected onto georeferenced raster layers to generate GIS-compatible decision-support maps, directly informing zone-specific management interventions without requiring practitioners to interact with the underlying model.

Future research priorities

To overcome systemic limitations, future research must move beyond maximizing accuracy on static datasets toward developing robust, interpretable systems. Establishing standardized, open-source, multi-biome benchmark datasets is essential to facilitate objective performance evaluation (Gregoire et al. 2016; Høye et al. 2020). To resolve the “transferability paradox,” research must prioritize Unsupervised Domain Adaptation (UDA) and transfer learning to adapt models to new biomes with minimal retraining (Abd Zaid et al. 2025; Arun et al. 2025; Narine et al. 2019).

Bridging the trust gap between AI systems and forest managers is also critical for regulatory adoption. Integrating Explainable AI (XAI) frameworks (SHAP, Grad-CAM) and rigorous uncertainty quantification will ensure "black-box" predictions are legally defensible and transparent (Ahmed et al. 2023; Alzubaidi et al. 2021; Mienye and Swart 2024; Ding et al. 2025). Finally, the intelligent fusion of multi-modal data remains the most promising frontier for maximizing inventory precision (Li et al. 2022a, b, c; Kangas et al. 2018).

Practitioner’s decision framework: selecting optimal approaches

It is important to clarify the methodological basis of the frameworks presented in this section. The decision matrix (Table 9), the ternary performance–interpretability–efficiency diagram (Fig. 6), and the radar chart (Fig. 10b) are author-proposed interpretive frameworks that synthesize three sources of evidence: (1) quantitative metrics extracted from the 190 reviewed studies, including computational profiles, pooled accuracy values, and confidence intervals; (2) publicly documented hardware specifications for the named deployment platforms; and (3) informed author judgment based on cross-study pattern analysis. The numerical values and category assignments in these frameworks represent evidence-informed estimates rather than empirically derived parameters from a formal scoring procedure. Their reproducibility is anchored to the underlying extracted dataset, which has been deposited at figshare (https://figshare.com/s/faae496f33443ad95466). Independent validation of the scoring scheme against a separate corpus of studies would further strengthen these frameworks and is recommended as a direction for future work.

To bridge the gap between academic research and forestry practice, we propose a decision framework (Fig. 9) based on operational constraints identified in the literature.

Practitioners should first define their operational requirements (real-time vs. batch processing). For edge computing applications, such as on‑the‑fly fire detection or illegal logging monitoring, lightweight architectures like MobileNet are well-suited despite a minor trade‑off in accuracy (El Asmar and Felfla 2025; Noor and Ige 2024). Conversely, for national‑level carbon stock estimation where precision is paramount and processing time is secondary, fusion‑based Deep Neural Networks (DNNs) or 3D‑CNNs utilizing LiDAR and satellite data are recommended (Zhu and Liu 2019; Ayrey and Hayes 2018). This stratified approach ensures that model selection is driven by logistical reality rather than just raw accuracy metrics. By aligning computational profiles with logistical realities, this framework directly supports Adaptive Forest Management (AFM). For instance, edge-deployment profiles enable rapid-response interventions for sudden pest outbreaks or illegal logging, while well-resourced profiles facilitate long-term, high-fidelity carbon stock monitoring required for international REDD + compliance (FAO 2020).

Figure 9a presents a four-step decision protocol for selecting DL architectures and sensor configurations in operational forestry. The flowchart implements four resource-matched deployment profiles based on training data volume. Profile D handles fewer than 500 samples and relies on lightweight CNNs with linear probing. Profile C targets 500 to 2,000 samples and is optimized for edge-deployed quantised MobileNet on UAVs. Profile B accommodates 1,000 to 5,000 samples utilizing EfficientNet variants for multispectral data. Profile A requires more than 10,000 samples and leverages ViT ensembles and LiDAR/hyperspectral fusion on GPU clusters.

All routing profiles converge at Step 2. This step mandates best-practice training and validation, including stratified k-fold cross-validation (k ≥ 5), confidence interval reporting, uncertainty quantification, and interpretability visualisations. Models meeting performance targets advance to Step 3 for cross-biome transferability testing across varying seasons, sites, and sensor configurations. Only models demonstrating less than a 10% accuracy drop proceed to Step 4 for production deployment. Models failing this threshold are labelled "limited-scope" and routed to a strict documentation and monitoring branch.

Figure 9b shows the multi-dimensional capability profiles of six architectures assessed across six operational criteria using a normalised 0–10 scale. The radar chart clearly indicates that no single architecture dominates all dimensions. Instead, each displays a distinct profile of complementary strengths and structural limitations that must inform task-specific selection.

The ViT/Transformer attains the highest pooled accuracy of any architecture (~ 9.5) and leads in multi-modal fusion capacity (~ 8.5). This reflects its attention mechanism’s capacity to integrate heterogeneous input streams across large spatial extents. However, these strengths incur a significant operational cost. ViT records the lowest scores among all architectures for edge deployability (~ 3.0), data efficiency (~ 3.5), and transferability (~ 2.0). These are the three dimensions most critical for real-world deployment in resource-constrained forestry environments. Its transferability score is especially low, representing the weakest value across the entire chart. This aligns directly with the severe cross-biome degradation findings discussed in Sect. "Model transferability across ecological zones".

MobileNet (TinyML) presents the exact inverse profile. It leads all architectures in edge deployability (~ 9.0), data efficiency (~ 7.5), and transferability (~ 6.0), while recording the lowest accuracy score (~ 7.0). This accuracy–efficiency trade-off defines the TinyML design paradigm. It makes MobileNet the most operationally viable architecture for on-device inference on UAV platforms or remote field deployments where computational resources are highly limited. EfficientNet-B3 occupies a complementary niche. It ranks second in edge deployability (~ 7.5) with a more balanced mid-range accuracy (~ 8.0). This provides a highly practical intermediate option between raw performance and deployment feasibility.

The standard CNN presents the most balanced overall profile, with no dimension falling below approximately 3.5. It achieves moderate-to-strong scores across all six axes, a finding consistent with its status as the most widely adopted architecture in the corpus. The 3D-CNN achieves high accuracy (~ 9.0), ranking second only to ViT. However, it scores poorly in edge deployability (~ 5.0), reflecting the substantial computational overhead required for volumetric convolutions over LiDAR point clouds.

The most diagnostically distinctive profile belongs to the GNN (PointNet + +), which records the highest interpretability score of any architecture in the chart (~ 7.5). This provides a unique competitive advantage for applications where spatial reasoning about inter-tree relationships must be auditable or ecologically interpretable. Transferability remains the weakest axis across the entire chart for most architectures. Only MobileNet (~ 6.0) and GNN (~ 5.5) achieve scores above 5 in this category. This reinforces the core conclusion drawn throughout this review that cross-biome and cross-sensor generalisation remains the principal unresolved challenge for operationally deployable deep learning systems in forest remote sensing.

Authors

1. Gabriel Osei Forkuo
2. Stelian Aleandru Borz

Cite this article

Forkuo, G.O., Borz, S.A. Deep learning-based computer vision in forest monitoring and management: a systematic review. Biodivers Conserv 35, 208 (2026). https://doi.org/10.1007/s10531-026-03412-x

Download citation

• Received: 16 March 2026

• Revised: 23 June 2026

• Accepted: 25 June 2026

• Published: 06 July 2026

• Version of record: 06 July 2026

• DOI: https://doi.org/10.1007/s10531-026-03412-x

Keywords