Tree Biomass Estimation in Agroforestry for Carbon Farming: A Comparative Analysis of Timing, Costs, and Methods

Abstract

Agroforestry systems (AFSs) enhance long-term carbon sequestration through tree biomass accumulation. As the European Union’s Carbon Farming Certification (CRCF) Regulation now recognizes AFSs in carbon farming (CF) schemes, accurate tree biomass estimation becomes essential for certification. This review examines field-destructive and remote sensing methods for estimating tree aboveground biomass (AGB) in AFSs, with a specific focus on their advantages, limitations, timing, and associated costs. Destructive methods, although accurate and necessary for developing and validating allometric equations, are time-consuming, costly, and labour-intensive. Conversely, satellite- and drone-based remote sensing offer scalable and non-invasive alternatives, increasingly supported by advances in machine learning and high-resolution imagery. Using data from the INNO4CFIs project, which conducted parallel destructive and remote measurements in an AFS in Tuscany (Italy), this study provides a novel quantitative comparison of the resources each method requires. The findings highlight that while destructive measurements remain indispensable for model calibration and new species assessment, their feasibility is limited by practical constraints. Meanwhile, remote sensing approaches, despite some accuracy challenges in heterogeneous AFSs, offer a promising path forward for cost-effective, repeatable biomass monitoring but in turn require reliable field data. The integration of both approaches might represent a valid strategy to optimize precision and resource efficiency in carbon farming applications.

1. Introduction

Agroforestry is a land-use system that integrates trees, shrubs, or other woody perennials in traditional agricultural systems, including crops or animals [1]. Thanks to the tree carbon sequestration potential in aboveground and belowground biomass (respectively, AGB and BGB), agroforestry systems (AFSs) have gained increasing research interest to enhance long-term carbon sequestration in agriculture [2] and support the Voluntary Carbon Market.

The Carbon Farming Certification Framework (CRCF) Regulation is designed to support carbon removal initiatives by introducing EU-wide standards for quality, verification, and certification. It also outlines procedures for recognizing and managing certification schemes, as stated in Article 1 (1) of the CRCF. In the context of carbon farming, the proposal builds on the foundations set by the LULUCF Regulation (EU) 2018/841 [3], which provides a framework for tracking and documenting carbon removal efforts. Importantly, data collected from carbon farming must align with national greenhouse gas inventory requirements. To promote efficiency, the regulation advocates for cost-effective biomass measurement techniques using tools such as digital databases, geographic information systems (GISs), and satellite-based remote sensing technologies, including those provided by the Copernicus Sentinel programme [4].

The EU envisions carbon farming as a key element in its broader strategy to neutralize residual emissions from high-emission sectors. However, challenges persist. Carbon farming is susceptible to reversal risks, and the carbon sequestration potential of various methods may be overstated due to limited understanding of complex ecosystem dynamics and their susceptibility to long-term climate change impacts [5,6]. While the adoption of carbon farming practices appears inevitable, their actual contribution to mitigating climate change continues to be debated. Biomass estimation is a key requirement for carbon trading and farming schemes, also known as Emissions Trading Systems (ETSs), which aim to remove or avoid greenhouse gas emissions. However, recent studies have highlighted significant weaknesses in these systems. A systematic assessment showed that fewer than 16% of 1 Gt CO₂e of assigned carbon credits corresponded to real emission reductions, leading to a shortfall of over 840 Mt CO₂e [7]. These findings have triggered calls for fundamental reform of carbon crediting mechanisms, as discussed at COP29 [7]. The regulation of such markets is already improving with the deployment of new technologies such as satellite imagery and LiDAR (Light Detection and Ranging)–laser scans that build three-dimensional images to estimate carbon stocks and verifiably upload data to ensure transparency. Another innovation is the extension of the carbon credit concept to cover so-called biodiversity credits as recently proposed by the Global Biodiversity Framework [8]. In the European context, AFSs are now officially included in the carbon farming practices of the CRCF Regulation [3]. The CRCF Regulations requires a complete, highly accurate, and transparent quantification of CO₂ removal in carbon farming activities. In particular, CO₂ removal monitoring should rely on a proper strategy combining on-site measurements with remote sensing or other advanced methodologies [9].

In forestry, tree biomass and carbon stocks can be measured directly through tree destructive measurements or estimated indirectly using biomass expansion factors (BEFs), which derive AGB from stem volume data [10]. Indirect estimations are usually preferred as stem volume data is easily provided by regional or national forest inventories, which can well predict tree AGB in homogeneous forest growth conditions [11]. When heterogeneous tree growth conditions make biomass accumulation variable or new tree species are considered, allometric models based on field-destructive measurements are necessary for assessing biomass stocks [1]. Allometry quantitatively relates complex tree parameters to measure with other parameters (or predictors) that are easier to obtain. Stem diameter at breast height (DBH) and tree height (H) are typical allometric predictors [12,13]. A large number of forestry allometric equations have already been developed and included in GlobAllomTree, an online platform that contains over 5000 tree allometric equations distributed across 73 fields [14].

In AFSs, trees may follow different growth patterns from forest trees in terms of tree density and competition [1,15]. In particular, tree competition in AFSs is reduced due to a lower tree density [16]. Moreover, different light conditions [17], frequent pruning for fruit [18] or timber production [19], and fertilization practices may alter growth rates, tree slenderness coefficient, and crown development [20]. This significantly reduces the accuracy of forestry allometric models by changing the relationship between predictors and biomass. Therefore, tree biomass estimation in AFSs requires site- and species-specific allometric models based on time-consuming and expensive in situ destructive (‘ground-truth’) measurements [21]. Such difficulties have led to a limited literature on allometric models in AFSs, with various sampling methods applied.

In addition to field-destructive measurements, remote sensing has emerged over the last few decades as a faster and less laborious ex situ technique for deriving AGB data from single-tree volume estimations [22]. Remote sensing includes a wide range of technologies for tree biomass estimation, which range from high-resolution satellite imagery [23] up to unmanned aerial vehicle (UAV) scanning [24,25]. While remote sensing offers highly accurate AGB estimations in forestry [23], agroforestry understory cropping, heterogeneous vegetation, and structure affect the accuracy of AGB estimations [22]. Remote sensing biomass models based on forestry allometric equations may potentially underestimate the higher branch biomass of agroforestry [26]. Moreover, higher background soil reflectance due to lower tree density in AFSs complicates remote sensing AGB estimation [27]. At the same time, the integration of remote sensing data with AI algorithms and multi-sensor platforms is significantly enhancing the ability to estimate AGB with greater spatial detail and transparency, especially in tropical regions where unreported land-use changes—such as the expansion of oil palm plantations—can now be detected and quantified [28,29]. Therefore, remote sensing biomass estimations may still need calibration and validation from in situ destructive AGB measurements [22].

INNO4CFIs (Nature-Based Business Model and Emerging INNOvations to enhance Carbon Farming Initiatives (CFIs) while preserving Biodiversity, Water Security and Soil Health “https://inno4cfis.eu/” (accessed on 23 July 2025) is a European research project that aims to innovate carbon farming through afforestation practices in AFSs. Rooted in European already established and ex novo agroforestry plantations, INNO4CFIs focuses on improving remote sensing accuracy of AGB estimations in AFSs through a cross-validation with ground-truth data obtained by in situ destructive AGB measurements [30].

In this paper, we reviewed the general literature on methods and approaches for AGB estimation in AFS for carbon farming. The review of these data allowed us to analyze the different impact that destructive and remote sensing methods might have in determining the carbon accumulation of agroforestry systems. We addressed the following questions: (1) What are the advantages/disadvantages of using destructive methods for AGB measurements? (2) What are the limitations of using remote sensing technologies in AGB detection? (3) What is the impact of these different approaches in the carbon credit market?

Based on these considerations, we hypothesize that a mixed approach combining ground-truth data with remote sensing could increase accuracy of AGB estimation while maintaining costs low. To the best of our knowledge, no studies have directly evaluated the timing and costs per surface or plant unit in destructive and remote sensing AGB estimations, including post-processing data in a real case study. Here, we analyzed time and costs from a pilot AFS realized within the INNO4CFIs project. In particular, this pilot AFS is composed of ten rows of olive (Olea europaea L.) and cypress (Cupressus sempervirens L.) trees, intercropped with fava bean (Vicia minor L.).

2. Traditional and Innovative Methods for Estimating AGB in AFSs

2.1. Traditional Field-Destructive AGB Measurements

2.1.1. Field Activities

Both in forestry and agroforestry, field-destructive tree biomass measurements consist of three main parts: tree selection for sampling, AGB and/or BGB harvesting and finally determination of tree fresh and dry biomass weight [31,32]. However, the field methods used in AFSs widely vary due to several constraints. In particular, methodologies for sampling tree AGB may range from random [21,33] to systematic sampling [34,35] due to variable availability of tree sampling. Systematic or stratified samplings are commonly used in experimental AFSs with no tree-felling restrictions. In an intercropping system with five different tree species, for example, Barzgar et al. [21] adopted a stratified random sampling that divided the tree population into strata representing the diameter classes, in which they randomly selected the trees to sample. However, several studies focusing on high-value or protected tree species substantially reduced the sampling number to a minimum sample size [33], sampled naturally felled or diseased trees [34], only branch biomass [35]. Moreover, many tree carbon sequestration studies located in cocoa (Theobroma cacao L.) [36,37,38], mango (Mangifera indica L.) [39] and avocado (Persea americana Mill.) [38]. AFSs used only non-destructive dendrometric (DBH and H) measurements and adopted previous species-specific allometric equations from literature. Therefore, tree availability is a major constraint and a cause of discrepancy in agroforestry tree biomass estimation studies. When destructive measurements are taken, most of the studies sampled AGB according to the different tree components (trunk, branches, small branches and leaves), from which the total fresh weight is determined [40]. The whole AGB fresh weight is then converted to the oven-dry weight by determining the moisture content (MC) of the subsamples [33,40].

BGB is a relevant but challenging carbon pool to measure due to the high workload required by root system excavation [12,41]. Therefore, such difficulty has led literature to mainly focus on AGB estimation in AFSs, with a very limited number of studies also including BGB estimation. BGB can be alternatively estimated using an indirect root-to-shoot ratio, where BGB usually accounts for 20%–40% of total tree biomass [33,34,42]. However, the root-to-shoot ratio may vary substantially according to species- and site-specific conditions. For instance, a previous study reported a 30% BGB allocation in Indian AFSs using Celtis australis L. [33], whereas another study found a 15.21% BGB allocation in AFSs of the same country using Tecomella undulata (Sm.) Seem. [34]. In an extensive study on olive (Olea europaea L. cv ‘Leccino’) biomass allocation, Brunori et al. [43] showed that root/shoot ratio may also vary depending on tree basal diameter. Such variability makes the root-to-shoot ratio not suitable for accurate carbon sequestration analyses. BGB can be directly and accurately measured through destructive measurements, where coarse roots (larger than 2 mm) are excavated using minirhizotrons, soil cores and profile walls techniques [12,41]. Commonly, a square plot corresponding to the tree area is excavated to a predefined depth, which can vary according to species-specific root growth patterns. For example, Borden et al. [44] reached 1 m depth for BGB measurement in Thuja occidentalis L. and Juglans nigra L., while Zu et al. [45] decreased soil depth to 50 cm for Larix kaempferi L. Following root system excavation, all coarse roots for each sample tree are weighed in the field and subsequently oven-dried at 105 °C for root gravimetric water content analysis.

2.1.2. Data Analysis for Allometric Equation Development

Tree carbon sequestration in forestry is directly related to the quantification/estimation of biomass stock, divided into various vegetation types [13,46,47]. As biomass stock estimation requires destructive measurements, indirect methods and sampling strategies are usually preferred [13,46]. The most common approach involves developing biomass functions through regression models relating biomass or volume to easily measurable variables, such as DBH and H [46]. Numerous studies have confirmed the strong correlation between morphological variables, especially DBH, and total dry biomass [48,49,50]. Simplified models often rely solely on DBH, whereas more complex models may incorporate additional variables, such as H and wood density, thereby improving prediction accuracy [13,47]. The choice of variables and model parameters is crucial to optimize both accuracy and bias control in biomass estimation [13].

According to IPCC guidelines, a default carbon fraction of 0.47 is typically used to convert biomass estimates into carbon stock values [51].

However, the accuracy of allometric models largely depends on the quality of input data. In AFSs, where species composition and management intensity vary greatly, localized destructive sampling remains essential for developing site-specific equations [13,46]. Field measurements required to calibrate these models must take into account local agroforestry conditions to ensure reliability.

While DBH and wood density are generally easy to measure and provide consistent estimates, incorporating H can significantly improve model performance. However, height measurements often carry greater observational errors, particularly in dense or closed-canopy systems [13,52]. In agroforestry contexts, species- and site-specific allometric models—linking DBH, H, and wood density to AGB—remain the most effective approach [13,53,54]. However, biomass estimation models must account for structural variability influenced by site conditions and natural disturbances [47]. A general equation for estimating tree biomass using allometric models is:

$$Y={\beta }_0{X}_1^{{\beta }_1}{X}_2^{{\beta }_2}...X_J{\beta }_J+\epsilon$$

(1)

where Y represents tree biomass, $X_J$ are biometric predictor variables (e.g., DBH and H), ${\beta }_J$ are model parameters, and ε is the error term (Equation (1)). Specific biomass estimation models incorporating both diameter and height include:

$$㏑w=a+b㏑D+c㏑H$$

(2)

$$㏑w=a+b㏑\left(D^{2}H\right)$$

(3)

where W represents AGB, D represents DBH (cm), H represents height (m), and a, b, a′, and c are regression coefficients (Equations (2) and (3)).

Numerous allometric equations have been developed for forest trees [55,56,57], whereas fewer have been developed for trees in AFSs. Among commonly cultivated species, poplar (Populus sp.) and willow (Salix sp.) clones have received significant attention due to their high biomass production [58,59,60]. A selection of published equations is summarized in

Table 1. Allometric equations suitable for AFSs (Y is AGB; C is circumference; and AGBC is aboveground biomass carbon). The number of the analyzed individuals (n), the standard model fitting metrics used (R² and RMSE), and the reference of each study are reported.

Type	Equation	n	R2	RMSE	Reference
Linear simultaneous AFSs (Uganda)—Various species	$㏑\left(Y\right)=-2.16+1.05㏑\left(H\right)+\epsilon$	12–16	0.30	0.761	[60]
	$㏑\left(Y\right)=-3.04+2.79㏑\left(H\right)+\epsilon$		0.739	0.208
	$㏑\left(Y\right)=2.19-0.79㏑\left(DBH\right)+3.27㏑\left(H\right)+\epsilon$		0.759	0.189
Teobroma cacao in AFSs (Cameroon)	$㏑\left(Y\right)=-1.613+1.83㏑\left(DBH\right)+\epsilon$	35	0.942	0.198	[42]
	$㏑\left(Y\right)=-1.598+1.628㏑\left(DBH\right)+0.36㏑\left(C\right)+\epsilon$		0.945	0.190
Wild cherry trees (Prunus avium L.)—AFSs (Germany)	$㏑\left(Y\right)=-1.674+2.286㏑\left(DBH\right)+\epsilon$	70	0.982	-	[61]
	$㏑\left(Y\right)=-0.340+0.92㏑\left(DBH\right)-0.167DBH$ $+0.0321㏑\left(DBH\right)+DBH\epsilon$		0.782	-
Quercus rubra (Canada)	$AGBC=0.008{\left({DBH}^{2}H\right)}^{1.0086}$	12	0.984	30.3	[21]
Juglans nigra (Canada)	$AGBC=0.0093{\left({DBH}^{2}H\right)}^{1.052}$	16	0.982	18.2
Picea abies (Canada)	$AGBC=0.0037{DBH}^{2.536}$	13	0.988	23.15
Robinia pseudoacacia (Canada)	$AGBC=0.0013{DBH}^{2.536}$	10	0.967	19.09

.

Based on the data shown in Table 1, the prediction efficiency of allometric curves based on AGB data from destructive measurements varies greatly depending on the factors in the study areas. The variability explained by allometric curve models (average of 0.851) is high, representing a high level of reliability of these data as inputs to these models. Due to the practical difficulties and high costs associated with excavating root systems, BGB is rarely measured directly. Instead, it is commonly estimated using empirical root-to-shoot ratios or functions derived from AGB, an approach broadly adopted in biomass studies despite its inherent uncertainties [13,61]. Variable agroforestry tree growth conditions require differentiation between biomass equations for inner and outer rows in a field/plantation. Border trees receive more sunlight and wind exposure than inner-row trees, affecting growth patterns [13,45]. While DBH, H, and biomass may differ across spatial scales, previous studies have found no significant differences in allometric relationships [13,46].

The decision between using existing allometric equations or developing new ones represents a key challenge in tree biomass estimation. Verifying the suitability of available equations typically requires direct measurements or destructive sampling, followed by rigorous statistical testing. According to Picard et al. [12], equation accuracy can be assessed by comparing predicted and observed biomass values, using residual analysis and calculating prediction errors. Common validation criteria include relative bias, relative root mean square error (RMSE), and the proportion of observations falling outside a defined confidence interval (e.g., 95%). A reliable model is characterized by a uniform distribution of residuals and minimal deviation between predicted and measured biomass. Importantly, uncertainties associated with the selected allometric model should be included in the total biomass estimation error. Developing new equations becomes necessary when existing equations prove inadequate due to species-specific variability or distinct site conditions. As per IPCC guidelines [51], this requires harvesting biomass from at least 30 trees representing the full range of diameter classes. If the resulting regression lacks statistical strength (e.g., low R²), additional sampling is needed, which implies increased time and cost. In recent years, remote sensing has emerged as a promising alternative, potentially reducing reliance on destructive methods thanks to improved availability of high-resolution data. Nevertheless, AFSs, remote sensing accuracy is often compromised by complex vegetation structures, crop interference, and species heterogeneity. While remote sensing performs well for trees with large branches (e.g., >5 cm), destructive sampling remains essential for calibrating and validating remote estimates, especially for young plantations, dense understory conditions, or species with challenging morphologies. For instance, Olea europaea L., a common Mediterranean species, presents difficulties in accurately estimating DBH, height, and crown development via remote sensing including LiDAR. Moreover, destructive methods are crucial for generating baseline data on underrepresented species such as Cupressus sempervirens L., widely used in the Mediterranean region for windbreaks, timber, and reforestation [62]. Within the INNO4CFIs project, destructive sampling is being conducted to support the development of precise non-destructive biomass estimation models. Despite its high accuracy, this approach entails significant logistical, financial, and temporal constraints, which must be carefully weighed when designing biomass monitoring protocols.

2.1.3. Advantages and Limitations of Destructive AGB Measurements

Destructive measurements for tree AGB estimation in AFSs represent a resource-intensive approach, which could have some important advantages in supporting the calibration and improving the accuracy of remote sensing-based AGB estimations. Preliminary in situ measurements may assist remote sensing in accurately delineating the small size and/or narrow width of AFSs [22]. Destructive measurements may also provide allometric equations that can estimate the large branch biomass of trees in AFSs, thus limiting remote sensing potential underestimation due to the use of unsuitable forestry allometric models [63]. Moreover, background soil reflectance can affect remote sensing AGB estimation in low tree density areas. AGB destructive measurements of sample plots can validate remote sensing accuracy [27]. Destructive measurements may still play a prominent role in the AGB estimation of new AFSs, where understory crops affect AGB estimation when using remote sensing [22]. Finally, destructive measurements offer valid support whenever new species are used in AFSs.

Nevertheless, destructive measurements have important limitations that always need to be considered. First of all, time-consuming and expensive field campaigns are required for tree selection and sampling. The several operations needed (tree felling, dissection and weighing) usually require proper machinery (chainsaw and excavator, if also sampling BGB) and are personnel-intensive [31]. A field team composition is commonly composed of felling and weighing units. Overall, these units involve at least two loggers and other three operators (a digger driver and two additional operators handling the logs and the branch bundles) [12]. In

Table 2. Cost estimation associated with field-destructive tree AGB measurement in EU countries where INNO4CFIs AFSs are located.

Country	Cost per Tree (€/Tree−1)	Source
Italy	37.09	Local public prices (Tuscany region)
Spain	79.26	Local public prices (Zamora Province)
Belgium	300–500	Local private prices (Hainaut Province)

, we reported national cost information about tree destructive measurements from three European countries (Italy, Spain, and Belgium). Large cost variation may be caused by different felling conditions, which are usually less and more expensive in agricultural and periurban conditions, respectively. Not less importantly, destructive measurements impact on carbon farming potential by reducing tree total number in AFSs.

Similarly to AGB, the time-consuming and laborious workload associated with BGB destructive measurements restricts sample size and measurement repeatability [64,65]. Furthermore, root excavation methods create disturbance to the rhizosphere equilibrium and may damage other tree root systems, if proper precautions are not taken [64,65]. Overall, all these disadvantages have made BGB destructive measurements unsustainable as a primary long-term method of root carbon sequestration analysis. However, small-scale BGB measurements can still offer a prominent source of calibration and validation for root biomass estimates of recently developed remote sensing techniques such as Ground Penetrating Radar (GPR).

2.2. Innovative Methods: Remote Sensing-Based AGB Estimations

2.2.1. Remote Sensing and Machine Learning

The development and launch of multiple multispectral satellites for earth observation since the 1980s brought a new paradigm on analyzing imagery. Not only using manual delineation to determine an object, but also using the spectral pattern [66]. This accompanied with the increase in performance of computers, birth of the modern methodology to analyze the earth’s environment using remote sensing imagery and machine learning as we understood [22].

Multispectral remote sensing as stated refers to remote sensing that utilizes not only using common optical wavelengths of red, green, blue in 400 to 700 nm electromagnetic spectrum that our eyes can see but expanding that. This commonly means adding at least near-infrared spectral to capture more information. Infrared, which most humans cannot see, could be represented as a substitution to one of three RGB (red, green, blue) channels to be visualized. The result became known as a standard false colour composite (Figure 1) which is usually used near-infrared in the red channel, red in green channel, and green in blue channel. The reason for utilizing near-infrared is due to its ability to penetrate and reflect more of vegetation, which can provide information of said object as shown in (Figure 2).

The new information given by more spectra would give more features or predictors in machine learning. This in turn makes remote sensing better utilized for supervised and unsupervised machine learning. Many machine learning models such as decision trees and support vector machines have been used by many researchers and projects to model many earth phenomena such as deforestation, land cover, carbon stock, flood, heatwave, and more [69]. However, multispectral reign has its limit, especially when satellite remote sensing spatial resolution continues to be more detailed. Where one-pixel value does not determine the object but needs the whole neighbourhood or even more context. In this sense, a new method would need to be integrated such as object segmentation or convolution. This allows a more holistic approach to classify an object in satellite remote sensing imagery [70]. This methodology showed its advantage when classifying a house, park, street, trees, or any other complex non homogenous object.

Unmanned aerial vehicles (UAVs) can carry a variety of sensors, including RGB and multispectral cameras. Estimating AGB from UAV-based imagery can greatly benefit from machine learning techniques, which can complement or replace conventional workflows. Machine learning approaches are particularly effective for tasks such as image segmentation and feature extraction from digital surface models or vegetation indices. For instance, convolutional neural networks (CNNs) have been successfully applied to automate the classification of crop types and status [71]. Furthermore, machine learning algorithms can be used to predict AGB from those features derived from RGB and multispectral data. In this context, multiple linear regression, support vector machines, artificial neural networks, and random forest regression were compared in [72], with the latter identified as the most robust model for estimating maize biomass by integrating both structural and spectral variables.

2.2.2. Satellite-Based AGB Estimation

In recent decades, satellite-based remote sensing has emerged as a prominent tool for AGB estimation in forest temperate [73], Mediterranean [74] and tropical [75] ecosystems. Despite some technical challenges, remote sensing has received increasing interest for AGB estimation in AFSs. Current state-of-the-art techniques integrate optical, radar, and LiDAR satellite data with machine learning or statistical models to estimate AGB [22].

Optical imageries use visible, near-infrared, and short-wave infrared reflectance from terrestrial objects to create vegetation indices such as NDVI (Normalized Difference Vegetation Index and EVI (Enhanced Vegetation Index) [22,74]. These indices help estimate vegetation cover, productivity, and stress conditions, which are key factors for biomass monitoring [76]. Nowadays, optical remote sensing imagery varies within different spectral, spatial and temporal resolutions for AGB estimation at different scales. For instance, coarse- to moderate-resolution optical data is used for AGB estimations at global, continental or national level [77]. On the contrary, high (0.5–2.0 m) to very high (lower than 0.5 m) resolution optical data is used for AGB estimations at local scale [78]. In AFSs, several papers [79,80,81] estimated AGB using the free multispectral very high resolution datasets provided by Sentinel-2 (European Space Agency). Conversely, free available coarse to medium optical resolution imageries, such as USGS EarthExplorer (Landsat Series), MODIS and ASTER, are restricted to AGB estimations at the larger scales [82]. Commercial providers such as Pleiades (Airbus) and WorldView-3 (Maxar Technologies) (ESA, 2024) also offer high-resolution imagery for AGB estimation. However, limitations such as cloud cover, narrow features and presence of shadows may substantially affect observations in AFSs [22,83].

Synthetic Aperture Radar (SAR) technology uses active microwave signals to penetrate clouds and provide data independent of lighting conditions. SAR is sensitive to vegetation structure and moisture content, providing valuable insights into biomass. In AFSs, free Sentinel-1 (ESA) and commercial ALOS-2 PALSAR (Japan Aerospace Exploration Agency, JAXA) datasets are used for AGB estimations [81,84,85]. Differently from optical imagery, SAR is unaffected by cloud cover and can capture canopy structural features over time [86]. Conversely, SAR data processing requires specialized expertise and SAR backscatter can saturate at high biomass levels [86]. In particular, extraction of AGB information from SAR images may be complex in highly spatial fragmented and heterogeneous areas, such as in the Mediterranean AFSs, where discontinuous tree canopies due to irregular foliage density are difficult to characterize in SAR images [84,87].

LiDAR technology actively measures distances by emitting laser pulses and timing their return. This approach enables advantages in three-dimensional modelling of vegetation structures, such as tree height, canopy closure and biomass of AFSs [22,77,88]. Current operational LiDAR missions include GEDI (Global Ecosystem Dynamics Investigation, onboard the ISS) and ICESat-2 (Ice, Cloud, and Land Elevation Satellite-2), both led by NASA and freely accessible. GEDI offers sample-based coverage limited to latitudes between ~51.6° N and ~51.6° S, and ICESat-2 has a sparser vegetation sampling focus. Integration with other remote sensing datasets is often necessary for large-scale mapping. Complex data processing also demands significant computational resources [88].

In general, given the resolution of satellite acquisitions, an additional limitation is the detection of small trees with still undeveloped crowns (e.g., newly planted AFSs up to 2–3 years old depending on the species). In contrast, the ability to detect in inaccessible or dangerous areas (i.e., areas with high social instability) where AFSs are widespread, or where tree felling is not allowed or feasible -for legal or economic reasons- is undoubtedly an advantage of satellite remote sensing [22].

All the different types of satellite data, both free and commercial, are summarized in

Table 3. Satellite-derived data type description.

Satellite	Data Source	Access Link	Type
Sentinel-2	Copernicus Open Access Hub	Sentinel Hub	Free
Airbus Pleiades	Airbus	Airbus Portal	Commercial (from 17 to 50 EUR/sq KM)
Landsat-8/9	USGS Earth Explorer	Earth Explorer	Free
Sentinel-1 (SAR)	Copernicus Open Access Hub	Sentinel Hub	Free
ALOS-2 PALSAR	JAXA	JAXA Data Portal	Commercial
GEDI (LiDAR)	NASA Earthdata	NASA Earthdata	Free
ICESat-2	NSIDC	NSIDC	Free

.

Considering an area of interest of ~50 sq/km as in the case of the INNO4CFIs Living Hub of Follonica, the costs for the relative biomass measurements were estimated and listed in

Table 4. Cost estimation associated with satellite-based AGB estimation.

Element	Brief Description	Price Range (€)
Satellite-based script development	Python-based (software version 3.10) script for pre- and post-processing satellite imagery, as well as for developing, testing, and deploying the ML-based model to extrapolate AGB layers and related information.	15,000–20,000
Biomass model run (per execution)	-	500–1000
Commercial satellite data	Time to be considered for new satellite imagery tasking: 1 to 2 weeks (depending on tasking requirements and cloud presence).	750–1500
Open-source satellite data	-	Free

.

2.2.3. UAV-Based AGB Estimations

Within remote sensing systems, UAVs or drones can play a relevant role in increasing sustainability and efficiency of land management activities. As extensively reviewed by Pádua et al. [25], a broad range of UAVs and sensors is available for remote sensing in AFSs. Different from expensive large and medium-size UAVs, small, mini and micro UAVs generally weigh less than 20 kg and offer a few hours autonomy in limited distance range [89]. Small-size UAVs differentiate in fixed-wing and multi-rotor systems [25]. While fixed-wing UAVs are well suited for wide-area agricultural surveying due to their high cruising altitude and centimetre-level spatial resolution, the lower cruising altitude and sub-centimetre resolution of multi-rotor UAVs make them ideal for detailed monitoring of AFSs [25]. UAVs imaging or sensing capability is based on the sensors used. RGB, infrared, multispectral, hyperspectral and LiDAR sensors can be easily mounted on UAVs [25]. Advanced Deep Learning models of object-based recognition, which are expensive and require high computational capacity, or Machine Learning models, which are simpler and less expensive are then used to translate acquired data into biomass. According to a recent review by Chehreh et al. [90], RGB is the most used sensor in AFSs, followed by LiDAR and multispectral cameras. In particular, UAVs equipped with RGB and multispectral sensors have been used for the non-destructive estimation of various plant traits (e.g., canopy cover, tree height, growth rate) as well as for the calculation of vegetation indices of abiotic and biotic stress [91]. However, these sensors differ fundamentally in the data structure provided as output. RGB sensors capture high-resolution imagery, which is usually processed into an orthophotography mosaic by image stitching [92] or into digital surface models (DSMs), usually obtained through 3D reconstruction algorithms [93]. The introduction of structure-from-motion (SfM) algorithms has further increased 3D accuracy [94,95]. By transmitting a repeated optical laser pulse to the surface, LiDAR develops a 3D surface point cloud [93]. This point cloud is available for several downstream applications in AFSs, from forest growth monitoring [95] to plant height measurements [96]. RGB and multispectral cameras are usually preferred due to the high LiDAR costs. However, accuracy of UAV-based AGB estimations in AFSs may currently be affected by uncertainty about tree crown detection, which is probably caused by sparser canopy and low leaf density. For instance, Surový et al. [97] found differences in tree height estimation accuracy of agrosilvopastoral systems. Height root mean square error (RMSE) was larger in coak oak (Quercus suber L.) than in holm oak (Quercus ilex L.) and umbrella pine (Pinus pinea L.) due to the tiny cork oak apical shoots, especially in small trees. In the current state of the art, this would require calibration with field crown diameter and DBH measurements [97]. Tree height determination is another source of variability for UAVs-based AGB estimations. To this regard, Diaz-Varela et al. [98] tested a new workflow for estimating tree height in olive orchards with discontinuous (open vase cultivation system) and continuous (hedgerow) tree canopy. Through image acquisition by a RGB sensor on a fixed-wing UAV, they generated both an orthomosaic image and a digital surface model using structure-from-motion image reconstruction. Tree height accuracy was reported for continuous canopy, while accuracy was lower for discontinuous canopy. In a Tectona grandis L.f. silvopastoral system (SPS), Hernández-Cole et al. [99] found that canopy closure hindered UAV-based individual crown diameter measurement. UAV-based AGB estimation was significantly lower than field AGB destructive measurements, which could be caused by inaccurate tree orthomosaic identification due to canopy closure or the choice of the DBH prediction model. However, increasing flight tests at low altitude may facilitate UAV tree identification in case of high-crown coverage [99].

Despite this issue to be solved, UAVs or drones have great potential for efficient remote AGB estimation in AFSs. Indeed, drone-based imaging should complement satellite imagery retrieval in remote sensing applications. On the one hand, obtaining results with satellites requires less human intervention compared to drones. Raw and processed satellite imagery can be accessed through various portals offering free or paid services. Many of these portals also provide application programming interfaces (API) that facilitate the automation of the retrieval process.

Conversely, drone-based data acquisition is a labour-intensive process that includes preparing necessary documentation to comply with national and international regulations, designing the flight plan, setting up the UAS, executing the mission, and managing and processing the collected data. During the mission, the crew must consist of at least two individuals, each possessing the necessary competencies for the specific operational scenario, to ensure the highest levels of safety and efficiency. One pilot is responsible for controlling the drone, adhering to the planned flight path, and avoiding obstacles. The second pilot serves as a visual observer, maintaining situational awareness and monitoring the surroundings for potential hazards [100]. On the other hand, tasking is more flexible with drones than with satellites for two main reasons. Firstly, drone missions can be scheduled to meet date, time and weather constraints. Secondly, the payload can be tailored to fit the study requirements. In conjunction with flight plan parameters, this results in higher, adjustable, spatial, spectral, and temporal resolutions [101].

Consequently, drone-based models can play a significant role in downsizing satellite-based data. In this sense, monitoring within-field crop variability at finer spatial, spectral and temporal resolution can help farmers take informed decisions as a part of agricultural management. By way of illustration, Li et al. [102] compared Sentinel-2 and UAV data for crop monitoring in the context of precision agriculture. Particularly, they focused on the characterization of biophysical plant parameters of wheat and barley crops located in Brandenburg, Germany. On average, higher correlations of UAV-based multispectral data with agronomic parameters were found for leaf nitrogen (N), leaf area index (LAI), mean and maximum plant height, and fresh biomass, compared to Sentinel-2 imagery. Large-scale patterns could be derived from both Sentinel-2 and UAV imagery; however, the former was influenced by management-driven features such as tramlines, which could not be accurately georeferenced. Consequently, agronomic parameters were better correlated with UAV data than with Sentinel-2 data. Nevertheless, combining data from both sources could help improve crop monitoring for farmers and reduce costs. Another limitation of drone-based LiDAR used in new AFSs plantations is that small trees are poorly captured and the support pole can be incorrectly included in the biomass.

2.2.4. Root Biomass Quantification Through Ground Penetrating Radar (GPR)

As traditional BGB destructive measurements are laborious and soil-disruptive, substantial research effort has been invested in recent decades to develop indirect methods for root biomass quantification [103]. Ground penetrating radar (GPR) is a geophysical technique that detects alterations in ground physical structures [104]. Through a system composed of a control unit and two antennas, one for radar pulse transmission and the other for signal reception, GPR emits high-frequency EM energy waves that propagate into the ground [103]. While radar waves propagate through ground interfaces, signal reflection is induced by differences in subsurface dielectric permittivity and recorded by the GPR unit [104], which creates a final radargram or reflection trace. In a radargram, radar two-way travel time is related to the polarity and amplitude of the return signal [105,106]. In particular, water is characterized by a high dielectric constant, which generates a marked signal reflection in coarse roots with a larger water content than the near soil matrix [103]. However, depth penetration capacity of GPR signal is influenced by the GPR antenna frequency and the subsurface electrical conductivity. For instance, soil water and clay content may cause radar signal attenuations. Therefore, GPR measurements well adapt to dry and low-content clay soils [45]. When applied to root coarse biomass, minimal diameter resolution increases at higher antenna centre frequencies. For example, 400–500 MHz allows 1–4 cm root diameter size resolution, which increases up to 0.25–0.5 cm diameter resolution with 1500–2000 MHz [103]. GPR-based coarse root biomass estimation has been tested in many worldwide forestry species [45,107,108], where estimation accuracy was found to be site-specific. For instance, in a study on coarse root biomass in Larix kaempferi L., Zhu et al. showed that coarse root estimation was overestimated (+38%) or underestimated (−13%) depth-dependently, as compared with destructive measurements [45]. In Pinus palustris Mill., Butnor et al. found that below-stump biomass was underestimated, with the underestimation increase being correlated with tree DBH. When tree diameter was below 15 cm, root biomass was underestimated by 80% [109].

In AFS, GPR-based root biomass estimation has been so far studied in a few studies. In a 25-year-old intercropping system including five different species (Populus deltoides x Nigra L., Juglans regia L., Quercus rubra L., Picea abies L. Karst, Thuja occidentalis L.), Borden et al. assessed the accuracy of tree-scale estimates by relying on complete excavation of the coarse root system [110]. They found that GPR slightly overestimated root biomass of 4% and 24% in Q. rubra and P. abies, respectively. On the contrary, GPR underestimated root biomass of 32% and 16% in Populus sp. and T. occidentalis, respectively. When data were pooled, Borden et al. showed a linear relationship (r² = 0.75) between GPR-estimated and destructively measured biomass [110]. Such species-specific accuracy discrepancies, especially in terms of underestimation, may be due to low GRP resolution of small roots (less than 1 cm), undetected of coarse roots deeper than or outside of the GPR signal radar [110]. In African cocoa (Theobroma cacao L.) agroforestry systems both under shade and in monoculture [44], GPR-based structural root estimation correlated with DBH size and showed good accuracy as the DBH increased.

Overall, GPR technology shows some limitations to the detection of tree root biomass. Previous studies [109,110] found that smaller coarse roots induce less radar signal reflection, thus decreasing the capacity of biomass data collection. GPR accuracy also reduces for deeper roots as depth causes signal attenuation [45]. Moreover, shallow root systems are likely less explored by GPR signals [110]. Therefore, subsurface conditions and root growth patterns play a key role in GPR accuracy of BGB, with soil conditions being more or less site-dependently conducive for radar analyses. However, GPR offers important advancements of root biomass estimation. Despite GPR needing calibration from destructive measurements, these are significantly reduced in terms of workload and sampling minimum size as compared with traditional destructive measurements [44]. GPR-based measurements can be easily repeated over time, thus offering the possibility of temporal-scale studies. Although GPR best performs in limited environmental conditions such as well-drained soils with low-clay content, radar signal modelling has reached further developments in the last decade and might have potential for accuracy increase in field conditions [45].

2.3. A Case Study in Follonica (Tuscany, Italy)

Despite several papers testing individually or contemporaneously field-destructive and remote sensing-based methods for AGB measurements in AFSs, no studies have so far included an analysis of the time and costs associated with such methods. The INNO4CFIs research project focuses on improving remote sensing accuracy of AGB estimations in AFSs through cross-validation with in situ destructive measurements.

In the INNO4CFIs reference site of Follonica (Tuscany, Italy), AGB estimation in two cypress varieties (C. sempervirens L.) and nine olive varieties (O. europaea L.) grown in AFSs was obtained through destructive measurements and remote sensing satellite- and drone-based approaches. The experimental agroforestry trials are located within the experimental farm of ‘Santa Paolina’ (Follonica, Tuscany, Italy). Founded in 1967 and led by the Italian National Council (CNR), the station is a pilot farm in the area for research in agroforestry systems and fruit tree biodiversity conservation. The climate is Mediterranean subarid, with January representing the coldest month (9 °C as monthly mean) and July representing the warmest month (24 °C as monthly mean). The mean of the annual precipitation is 650 mm, concentrated in autumn and spring, while summer is dry and subject to drought periods. The soil of the Santa Paolina farm is mainly sandy-loam with poor chemical fertility due to a low content of organic matter. As found in specific physico-chemical soil lab analyses (

Table 5. Physical and chemical soil properties of the agroforestry fields at Santa Paolina farm (expressed as the mean of three samples, collected on 5 May 2024).

Parameter	Unit	Quantitative Value	Qualitative Value
pH	-	6.56	moderate acidity
EC	mS/cm−1 at 25 °C	0.109	low
TN	N g/kg−1	0.67	low
P	P2O5 mg/kg−1	8.66	low
K	K2O mg/kg−1	230	high
Ca	Ca mg/kg−1	1197	high
Mg	Mg mg/kg−1	153	high
Mg/K	ratio	0.67	low
SO	%	1.3	low
C/N	ratio	11	medium-low
CEC	meq/100 g−1	11.23	low
texture	-		sandy clay loam

), low values in nitrogen (N), phosphorus (P), organic matter (SO) affects this soil, jointly with an unbalanced Mg/K ratio due to the high content of Potassium (K) that could inhibit the uptake of Magnesium (Mg).

The destructive measurements, composed of preliminary non-destructive measurements for tree selection and followed by destructive tree samplings, were performed between December 2024 and January 2025. Contemporary remote sensing measurements were structured as follows: (i) drone-based multispectral measurements, (ii) optical satellites remote-sensing, (iii) LIDAR satellites remote-sensing, and (iv) SAR satellites remote-sensing. This range of technologies guarantees a robust representation of the existing methods, spanning from the “gold standard” of the destructive samplings to non-destructive methodologies with higher technology and innovation content.

2.3.1. Field-Destructive AGB Measurements

For the destructive measurements, a propaedeutic non-destructive campaign selected the tree individuals better representing the average DBH and H of the cypress and olive AFSs, distributed over a plot area of 0.8 and 0.6 ha, respectively. Overall, 18 cypress (5–22 cm diameter and 4.7–9.6 m height range) and 27 olive (10–17 cm range and 2.6–4.6 m height range) tree individuals were selected for felling between December 2024 and January 2025. After tree felling, the fresh weight of each tree component (stem, branches larger than 5 cm, twigs containing foliage and branches smaller than 5 cm) was measured using a field scale. Three subsamples were collected from each tree component and oven-dried at 105 °C to measure the dry weight. The field research team, composed of two loggers and four other people involved in measuring each tree component, completed the destructive measurements of 45 trees in 3 work days (corresponding to 18.9 h). The work of weighing the samples in the laboratory involved an additional commitment of one staff unit for one week.

2.3.2. Satellite-Based AGB Estimations

To support AGB estimation in the Follonica region, a series of five very high-resolution Airbus Pléiades imagery acquisitions were planned and collected through Airbus’ tasking service. The acquisition schedule was designed to cover seasonal variation and optimize observation conditions over the experimental plots of the Living Hub (see

Table 6. Airbus imagery acquisition planned in Follonica.

Acquisition Number	Acquisition Date Range
1	1 April–10 May 2024
2	1 July–10 August 2024
3	1 October–15 October 2024
4	1 January–10 February 2025
5	1 March–9 April 2025

).

Each acquisition covers an area of approximately 50 km², ensuring full coverage of the study region (see Figure 3).

Each imagery was subjected to a standardized pre- and post-processing pipeline (involving radiometric and geometric corrections, co-registration and cloud masking) as reported in Figure 4. After preprocessing, the data were ingested into the developed AGB estimation model.

2.3.3. UAV-Based AGB Estimations

The operational scenario in Follonica is subject to strict flying restrictions due to the proximity to a heliport and an urban environment. This demanded operating under the specific category, following the standard scenario STS-01, as defined by the Italian d-Flight system. Therefore, the in situ operation required submitting a coordination request to ENAC, as well as potential airspace reservation. Both the operator and pilots must comply with registration certificate, operational declaration, insurance, operations manual, emergency plan for the operator, STS theoretical and practical training, certified flight hours for the pilots and active Direct Remote Identification (DRI) on the drone. The process of preparing, submitting and approving this documentation may take several months and must be completed before the actual mission.

Data capture was carried out using a compact and efficient system based on the Mavic 3M Flysafe (Flying Eye, Biot, France), which complies with C5 requirement. This UAS features an RGB camera and a multispectral camera, which comprises four multispectral sensors: green (G), red (R), red edge (RE), and near-infrared (NIR). Additionally, the sensor platform integrates an RTK module for centimetre-level georeferencing. This equipment enables the generation of point clouds and 3D models via photogrammetry, as well as vegetation indices to assess vegetation physiological status.

The drone operations started with a pre-flight stage to prepare the required documentation. Once in the Follonica Living Hub, a visual inspection of the trees and the surroundings was conducted to adjust the equipment setup and the flight configuration to the AFS characteristics. Both nadir and oblique flights were planned, with spatial resolutions ranging from 1.25 to 2.5 cm/pixel, depending on flight altitude (between 15 and 55 m). This approach, combined with high overlap (90% frontal, 80% lateral), allowed for accurate modelling of canopy volume and calculation of parameters such as canopy cover fraction. These conditions also determined the number of batteries required, which ranged from one to four per hectare, depending on crop density and phenological stage at every plot. The system was complemented by an Emlid Reach RS2+ GNSS antenna (Emlid Tech Kft., Budapest, Hungary) mounted on either a surveying pole or a levelled tripod, depending on the operation to be carried out. Custom-made methacrylate plates, intended as ground control points (GCPs), were evenly distributed across the plots. Their positions in the plots were recorded progressively as they were laid out, using the GNSS antenna in rover mode. This step can indeed be performed either before or after the flight, with the sole condition that the plates are not removed until both tasks have been completed. Differential corrections are transmitted via NTRIP in FIX mode to ensure high geometric accuracy.

After completing the GCP location process, the Emlid antenna was setup in an elevated, interference-free location and a common geodetic reference system was configured to avoid the need for reprojecting between the images and the GCPs. Connectivity between devices (drone, GNSS station, remote control) was ensured via SIM cards for continuous data transmission. This dual acquisition approach was adopted because GNSS signal losses (base-drone) may occur during acquisition, affecting accuracy. If this happens during flight, some images may lack exact geolocation invalidating the data or making the construction of accurate models very difficult. This issue is mitigated with the use of GCPs, which allow reprocessing of affected areas using known coordinates, although this increases initial processing time. The overall pre-flight procedure can take up to one and a half hours, depending on crop size. The flight stage involves mainly the operation of the drone, adhering to pre-flight plans and regulations. The pilot monitors telemetry and maintains visual line of sight with the drone, while the co-pilot oversees the environment (birds, aircraft, unauthorized persons). The duration of this phase depends largely on the crop size and condition, and hence the need to exchange batteries. A MicaSense reflectance panel (AgEagle Aerial Systems Inc., Wichita, Kansas, USA) was used before and after each flight to enable the comparison of vegetation indices among different dates. In the Follonica Living Hub, the average duration of the flight stage was one hour per hectare.

In the post-flight stage, the recorded data were downloaded and projects configured in the Pix4D suite (Pix4D SA, Prilly, Switzerland). Pix4Dmapper performs image alignment, radiometric correction, and reflectance map generation, enabling the calculation of vegetation indices. Pix4Dmatic is optimized for large datasets, generating dense point clouds and 3D reconstructions with high geometric accuracy through automated photogrammetric processing. Intermediate results are usually obtained in 5–6 h. Reflectance images and orthomosaics were imported into QGIS (QGIS.org, Switzerland) to calculate and segment vegetation indices. This process may take about 15 min once the workflow has been automated. The point cloud is then classified in Pix4Dsurvey, separating ground from vegetation to generate precise 3D models of terrain, canopy, canopy cover and volumes. This process may take 15–20 h depending on the area. Overall, the workflow process used in the Follonica case study is summarized in Figure 5, while the overall carried out costs are listed in

Table 7. Description and amount of drone-based measurements costs in the Follonica’s case study.

Element	Brief Description	Price Range (€)
Drone and accessories	DJI Mavic 3 Multispectral with RTK module, extra batteries, landing pad	5000–5500
GNSS and surveying	Emlid Reach RS2+, tripod, survey pole, levelling base	2300–2800
GCPs	Custom methacrylate plates	40–170
Communications system	SIM cards for device connectivity	30–60 (per unit/year)
Analysis and modelling software	Pix4Dmapper, Pix4Dmatic, Pix4Dsurvey (licences)	8500–10,000
GIS and geospatial analysis	QGIS (free and open-source software version 3.40.9)	Free

.

2.3.4. Comparison of Timing and Costs from the Activities in Follonica

To provide a practical overview of timing and costs relative to AGB estimation activities in Follonica, we compared the different methods used. While the timing and the equipment required by the destructive measurements can be easily assessed, remote sensing cost analysis needs a deeper investigation. In our study, UAVs and satellites implied two main types of costs: (i) the costs directly related to the use of the UAV/satellite system on the area, together with the time needed for the measurements (amount of hours employed to deliver the task); (ii) the costs relative to the licences, the costs of goods with partial wear and its accessories, the satellite-based script development (post-processing data) (Table 4). Importantly, post-processing analysis requires specific skills and training representing an additional cost to be considered.

Whereas possible, we have accounted three ways to determine the costs, (i) the analytical way, quantifying the costs directly sustained in the agroforestry case study in Follonica, (ii) the synthetic way, based on the price list of the public works of Tuscany region or the comparison of the experimental data with market costs, (iii) the off-farm procurement, based on the costs of the service from private enterprises at regional scale heard for comparing the costs with the private market.

Moreover, another useful parameter to address the farm management in the Carbon Farming business is the cost sustained for tree felling (€ tree⁻¹) (

Table 8. Costs per hectare (€ ha⁻¹) to quantify AGB in the experimental AFSs in Santa Paolina farm (Grosseto province, Tuscany, Italy). Data were acquired between February and April 2025. The values marked with “*” do not include V.A.T. (Value Added Taxes) as they are estimated on the cost per hour of the research personnel employed on the task.

AGB Estimation Method (€ ha−1)	Analytical Estimate (€ ha−1)	Synthetic Estimate (€ ha−1)	Off-Farm Procurement (€ ha−1)
destructive sampling	965.19 *	929.89	1157.14
non-destructive sampling (allometric equations)	266.44 *	-	-
UAV-based measurement	400–670	430.00	650.67
Satellites-based (optical, LiDAR and SAR) measurements	1250–2500	-	-

). DBH and H measurements required three minutes (three personnel units). Tree felling, data analysis and journey costs need also to be included in the final evaluation. As mentioned before, this study entails analytical, synthetic, and off-farm cost estimates (

Table 9. Costs per tree (€) of AGB destructive measurements performed in Follonica experimental AFSs (Grosseto province, Tuscany, Italy). The AGB data were acquired between February and April 2025.

Destructive Sampling	Analytical Estimate (INNO4CFIs Real Costs)	Synthetic Estimate (Regional Price List)	Off-Farm Procurement (Private Market)
cost per tree (€)	30.03	28.93	36.00

). Off-farm procurement data were provided by regional enterprises, the synthetic value for the destructive samplings was obtained from the Public Work List of the Tuscany region. The synthetic value for the UAV-based measurement was obtained by comparing the experimental data with market costs. Finally, analytical values for the e-based measurement and the satellites data were provided by INNO4CFIs’ partners.

3. Discussion and Conclusions

Accurate estimation of AGB in AFSs is critical for carbon farming initiatives and their integration into Voluntary Carbon Markets (VCMs) and the EU’s CRCF [3,4]. Reliable AGB assessments underpin the quantification of carbon sequestration potential and directly influence the credibility and scalability of carbon credits derived from AFSs. However, methodological challenges—particularly the trade-off between precision and scalability—remain a significant barrier. Traditional destructive sampling methods, which involve harvesting and weighing tree components, are regarded as the most accurate approach for AGB estimation. Their precision, especially when used as input data for site- and species-specific allometric equation development, is unmatched [12,13]. Yet, these methods are inherently labour-intensive, time-consuming, and ecologically invasive, making them impractical for large-scale or repeated assessments. Moreover, due to the challenge posed by developing specific allometric equations, standard allometric equations are often used [13,49]. This occurs even in cases where allometric equations developed for a given species, site or cultivation form are not available, as in the case of AFSs [14]. In fact, they often fail to account for the unique growth patterns observed in AFSs, which are influenced by management practices such as pruning, fertilization, and reduced tree density [15,32]. This difficulty underscores the need for developing custom AGB estimation models that reflect the structural complexity of agroforestry [9,12,38,39,40,41]. In this context, remote sensing technologies offer promising alternatives [22,77,80]. Since 2015, the Copernicus project, launched by ESA, has opened the market to free satellite data. Satellite-based platforms—utilizing optical and SAR data—enable non-destructive, broad-scale biomass estimation [22]. However, the heterogeneity of AFSs, characterized by variable tree densities, mixed vegetation species and layers, and diverse soil backgrounds, poses significant challenges to the accuracy of satellite-derived biomass models [22,83]. Additional confounding factors such as cloud cover, narrow field geometry, and canopy shadows further constrain the utility of optical satellite imagery in agroforestry contexts. Consequently, satellite data would hopefully require calibration with high-precision ground-level data, derived from destructive sampling. Commonly, remote sensing methods still require field-measurements of tree parameters such tree DBH and height. When new tree species are investigated for the first time, direct biomass data from field destructive measurements are needed for remote sensing model calibration [45,111,112]. In this context, UAVs, or drones, represent a significant advancement in remote AGB estimation. UAVs equipped with RGB, multispectral, or LiDAR sensors can generate high-resolution 3D canopy models, height maps, and vegetation indices with extremely high spatial resolution, capable of capturing fine-scale structural details of vegetation [22,77,97]. UAVs are particularly well-suited for small-scale or heterogeneous AFSs, where their operational flexibility and fine spatial resolution enable detailed monitoring of biomass dynamics over time. By looking at current literature [96,98,99], UAV-derived AGB prediction models can achieve high accuracy, especially when using machine learning methods like Random Forest, combining spectral and structural inputs, and selecting appropriate vegetation indices [27,113]. The highest accuracies were obtained when models integrated multisensor data and robust regression methods [114,115]. However, model performance is highly dependent on input data quality, sensor type, and environmental conditions, sometimes more so than the model type itself.

However, UAV-based approaches face limitations related to flight regulations, operational costs, and limited area coverage per flight [22,113]. Despite their individual limitations, the integration of satellite and UAV-based remote sensing can yield a synergistic framework for AGB estimation. UAVs can provide detailed, site-specific data to calibrate broader-scale satellite models, thereby improving the accuracy of regional biomass assessments. This hybrid approach offers a scalable, non-destructive alternative to traditional methods, with potential to significantly reduce the cost and labour associated with biomass monitoring. This integrated approach may represent the best scenario that can only be implemented in funded research projects (such as INNO4CFIs), for obvious reasons of technical and economic sustainability.

The Follonica case study represents a valuable case study that compares traditional and emerging approaches to AGB estimation in AFSs. This experimental AFS, as part of the INNO4CFIs project, provides a comprehensive overview of the technical, logistical, and economic dimensions involved in the application of both destructive and non-destructive biomass quantification methods. Integrating destructive samplings, UAV- and satellite-based remote sensing allowed for cross-validation of AGB estimates and highlighted the strengths and limitations of each methodology. Destructive samplings served as the reference for biomass quantification and calibrated remote sensing biomass outputs. However, the high labour intensity and operational cost make this method unsuitable for large-scale or recurring measurements. The UAV operations, while being logistically demanding due to airspace restrictions and regulatory compliance, offer valid spatial accuracy and cost-efficiency. Satellite-based observations (optical, LIDAR, SAR) provide a broader spatial coverage and seasonal frequency, despite higher initial costs. Both UAV and satellite imagery processing require considerable computational power and algorithm development, as well as high technical expertise. Nevertheless, their scalability and non-invasiveness make them promising for regional biomass monitoring once adequately calibrated with field data. The Follonica case also underscores the economic trade-offs between accuracy, coverage, and repeatability. While destructive sampling is precise, its high cost limits its feasibility in agricultural systems. UAV-based measurements represent a practical balance, offering relatively low cost per hectare and operational flexibility, particularly in small or heterogeneous plots [87]. Satellite data, though more expensive, are indispensable for upscaling biomass assessments across landscapes and integrating with regional carbon inventories. Another important consideration is the temporal and logistical demands associated with each method. Field operations for destructive samplings were completed in just three days, but required intensive labour. UAV flights took approximately one hour per hectare, plus extensive pre-flight preparations, and ground control setup time. Meanwhile, satellite imagery acquisition depended on external providers and atmospheric conditions but required minimal or field measurements. High initial remote sensing costs could be economically sustainable for agricultural consortia or trade associations. The greatest cost source of satellite-based AGB estimations is model development. Therefore, the use of these tools in carbon farming needs appropriate technical support and knowledge transfer. The Follonica case demonstrates that a mixed-technology approach (destructive measurements combined with UAV and satellite observations) can provide better and accurate AGB estimates. Meanwhile, satellite methods, once calibrated, have potential as tools for regional-scale monitoring and carbon certification protocols. Looking forward, hybrid workflows combining periodic destructive validation, frequent drone flights, and satellite monitoring represent a promising pathway to operationalize carbon estimation in Mediterranean agroforestry systems. Prior knowledge of the tree species present in AFSs remains essential. Projects such as INNO4CFIs are essential for bridging the gap between traditional and emerging technologies and for developing reliable frameworks to support the expansion of carbon farming in AFSs. In addition to the technical and economic feasibility, CF initiatives should account also for the emission of CO₂ of the processes. A careful Life Cycle Assessment (LCA) of the three methods is currently not available in the literature. This aspect could be very decisive in adopting one methodology or another. Especially for data processing, the computing powers required for satellite and drone-based approaches, the infrastructure needed, and their operation, result in a non-negligible carbon footprint that will need to be taken into account in future LCA studies.