Introduction

Agroforestry systems (AFS) consist of landuse practices that intentionally integrate trees and shrubs into crop and/or animal farming systems. Five broad categories of AFS sequester more carbon than traditional agriculture (Wilson and Lovell 2016). Silvopasture integrates trees in pastureland with livestock and can store 45 percent more aboveground biomass (AGB) than perennial pasture (Udawatta and Jose 2012). While rows of trees and/or shrubs are integrated with agronomic or horticultural crops in alley cropping, windbreaks and shelterbelts are planted adjacent to crops or grazing areas in a linear fashion for wind protection. Riparian buffers allow perennial vegetation adjacent to water bodies. The most complex form of AFS is forest farming, which includes a diverse mix of perennial species modeled to mimic natural woodlands (Gene Garrett and Buck 1997). Multi-layer understory crops, woody biomass, and soil carbon in AFS can sequester around 0.29 to 15.21 Mg/ha/yr of AGB (Nair et al. 2010), making it one of the promising natural climate solutions (Griscom et al. 2017).

The promotion of AFS as a natural climate solution requires cost-effective monitoring of carbon biomass. There are two approaches to measuring AGB. While the in-situ measurement through field observations is more accurate than the ex-situ method that relies on remote instruments, it is resource intensive and have limited spatial coverage. Remote sensing is an ex-situ technique that assesses the physical characteristics of landcover by measuring the reflected and emitted radiation from a distance (Wang et al. 2019). It has been widely used in AGB estimation of forestry (Timothy et al. 2016; Pádua et al. 2017; Ahmad et al. 2021). AFS are more complicated than forestry because of understory crops, diverse vegetation, and heterogeneous structure (Udawatta and Jose 2012). Remote sensing has been applied in AFS since 1990s (Unruh et al. 1993; Houghton et al. 1993). However, with the greater access to high-resolution data and advances in AGB estimation methodologies, their use has increased significantly in the last decade (Fig. 1) (Czerepowicz et al. 2012; Chen et al. 2015; Schneider et al. 2018). Optical imageries, Light Detection and Ranging (LiDAR), Synthetic Aperture Radar (SAR), and other remotely derived imageries along with field observations provide information on canopy cover, tree height, and other characteristics need to accurately estimate AGB.

Fig. 1
figure 1

Cumulative and yearly frequency of the reviewed papers on AGB estimation by different methods

Full size image

Machine learning (ML) can model complex spatial patterns using a variety of large input data with higher accuracy (Wu et al. 2016). It has been increasingly used in AGB estimation of forests (Maxwell et al. 2018; Zhang et al. 2020) and gradually applied in AFS as well (Filippi et al. 2014; Suchenwirth et al. 2014; Güneralp et al. 2014). With the increasing use of remote sensing and ML in AFS, there is a need to synthesize the emerging literature. The main goal of the study is to review remote sensing and machine learning techniques for AGB estimation of AFS. The paper is outlined as follows. The methods used for the literature review are briefly discussed. The main findings related to various factors affecting AGB performances, including data sources, methodological approaches, and covariates use are discussed next. The paper concludes with a discussion on the opportunities and challenges of using remote sensing in AFS. The synthesis can aid researchers, practitioners, and growers in better understanding the possibilities and limitations of remote sensing and machine learning applications for carbon monitoring.

Materials and methods

Keyword searches of AFS-related terms were conducted using major databases, including Google Scholar, Web of Science, Scopus, and related journals. The following query was used: [Types of AFS] AND [aboveground biomass] AND [Types of remote sensing]. A total of 125 manuscripts written between 1991 and 2022 were shortlisted for review of which 33 articles were selected for detailed review based on the inclusion/exclusion criteria (see supplementary materials for more details).

Results and discussion

The performance of AGB estimation depends upon multiple factors, such as the spectral and spatial resolution of data sources, types of covariates, methods of AFS delineation and AGB estimation, and types and sizes of AFS. Some of the key factors potentially affecting the accuracy of AGB, measured by the coefficient of determination (R-squared) are discussed as follows. Statistical tests like the t-test and Kruskal–Wallis test performed to evaluate the statistical significance of the findings are also discussed.

Effect of data sources

There are three types of remote sensing data used in AGB assessment of AFS: optical imageries, Light Detection and Ranging (LiDAR) and Synthetic Aparture Radar (SAR), and multi-source. Optical imageries are the most frequently used (31 out of 33 studies) data source. The visible, near-infrared, and short-wave infrared reflectance from objects generate vegetation indices, texture information, and other parameters needed for AGB assessment. While satellites like Landsat provide global coverage, their coarse spatial resolution reduces the accuracy (Timothy et al. 2016). Therefore, high (0.5–2.0 m) and very-high resolution (lower than 0.5 m) imageries are commonly used in AFS studies (Macedo et al. 2018). Optical imageries produced an average R-squared value of 0.696.

LiDAR and SAR are also used for AGB assessments. Nine out of thirty-three studies used LiDAR whereas only two studies used SAR. LiDAR transmits and receives pulses of laser energy to measure the height, canopy closure, volume, and biomass of forest stands (Gatziolis and Andersen 2008). When combined with other data sources, some LiDAR studies have measured AGB with high accuracy. For example, Kanmegne Tamga et al. (2022) estimated the AGB of food forest by including tree height and tree density information from multispectral and LiDAR data (R-squared = 0.91, RMSE = 3.78 Mg/ha). Accuracy is also affected by the level of aggregation. LiDAR information can be aggregated at the plot- or plant-level (Chen et al. 2015). While plot-level assessment aggregates the AGB of individual trees for a given area, tree-level assessment produces higher AGB as it provides disaggregated information of individual trees (Graves et al. 2018). Unlike LiDAR which is airborne, SAR is a space-borne sensor that emits and receives long wavelength energy produced by its sensors. Since the low frequency (L- and P-band) bands can penetrate through trees, it performs better than optical imageries. For example, Naidoo et al. (2016) improved the AGB accuracy (R-squared between 0.83 and 0.88) by integrating L-band HH and HV backscatter of SAR with Landsat image reflectance.

While optical and LiDAR data sources are useful, combining these data products can help integrate information on different aspects of AFS. A total of 18 out of 33 papers used more than one data sources (Table 1). Multispectral and hyperspectral imageries along with LiDAR and field observations provide nuanced information about the biophysical characteristics of vegetation and thereby improve the accuracy. Wang et al. (2016) estimated the AGB of individual trees in shelterbelts with accuracy of 97 percent and R-squared of 0.61 by combining airborne LiDAR and airborne spectrographic imagery. However, fusion of data products doesn’t always improve performance. Luo et al. (2017) estimated the AGB of forest biomass by integrating LiDAR and hyperspectral imagery and found that spectral characteristics were good estimators of AGB and LiDAR only marginally improved the performance (R-squared increased by 5.8 percent). This is partly attributed to complementary information contained in these data products.

Table 1 Summary of key variables of the reviewed literature
Full size table

Methods for AFS delineation and AGB estimation

The AGB estimation using remote sensing generally involves the following steps. AFS are delineated automatically or semi-automatically using landcover classifications, or manually using GIS, remote sensing and field observation. AGB values are estimated for sample field plots through destructive or non-destructive sampling. The final step involves AGB estimation of the overall study area by establishing a relationship between field-based AGB and remote sensing and other variables (Wang et al. 2019).

For AFS delineation, five out of thirty-three studies used automated or semi-automated landcover classification. Along with spectral characteristics, scholars also used textural information to classify trees (Lourenço et al. 2021). Seven out of thirty-three studies used Object-based Image Analysis (OBIA) for AGB delineation. Unlike pixel-based classification that separates individual pixels directly, OBIA aggregates image pixels into spectrally homogenous image objects using an image segmentation algorithm (Fig. 2) (Liu and Xia 2010). OBIA had a classification accuracy between 79 and 89 percent.

Fig. 2
figure 2

Object-based classification of scattered trees (Source: Gonçalves et al. 2019)

Full size image

The AGB estimation consisted of regression and machine learning models (Fig. 3). Regression models were used in 15 out of 33 studies whereas 16 studies used machine learning. Two studies utilized statistical methods that upscaled field-level AGB to a larger area. Overall, machine learning had an average R-squared of 0.815 whereas that for simple linear regression was 0.686 (Fig. 2). However, when a t-test was performed between two groups (machine learning and non-machine learning), there were no statistically significant differences between them.

Fig. 3
figure 3

R-squared by AGB estimation methods

Full size image

Choice of covariates

The covariates used in AGB estimation models included spectral indices and bands, textures, biophysical variables, and geomorphometric variables (Table 2). Spectral indices are widely used in AGB assessments because of their high correlation with the vegetation index (Fig. 4) (Laosuwan and Uttaruk 2016). Of all the indices, the Normalized Difference Vegetation Index (NDVI) is the most frequently used index to measure vegetation greenness (Fig. 5). However, it can suffer from over-saturation and become insensitive to woody parts of AFS where most of the carbon is stored (Lu 2003). Indices like Atmospherically Resistant Vegetation Index (ARVI) reduces over-saturation by minimizing atmospheric aerosol brightness (Bordoloi et al. 2022).

Table 2 Highly cited covariates from the reviewed literature
Full size table
Fig. 4
figure 4

Frequency of highly used spectral indices. Acronyms: ARVI Atmospherically Resistant Vegetation Index, EVI Enhanced Vegetation Index, GNDVI Green Normalized Difference Vegetation Index, NDVI Normalized Difference Vegetation Index, NDWI Normalized Difference Water Index, SAVI Soil Adjusted Vegetation Index, SR Simple Ratio, VARI Visible Atmospherically Resistance Index

Full size image
Fig. 5
figure 5

NDVI image of a farm with orchards and vineyards in Central Indiana (Source: Authors, Imagery: USA National Agricultural Imagery Program, Spatial resolution: 1 m, Image Date: June 6, 2020)

Full size image

Besides spectral indices, spectral bands, band ratios, and textures, along with biophysical and geomorphometric variables were also frequently used covariate types (Table 2). A total of 29 out of 33 papers utilized these variables. (Table 1). Among them, single band and band ratios are also good predictors of AGB. Prasondita et al. (2019) estimated the AGB of food forest using spectral bands and indices and found that single band 7 of Landsat 7 ETM + is the best predictor (R-squared = 0.44, RMSE = 52.85 tonnes/ha).

While spectral indices or spectral bands can provide some information on AGB, scholars have included more than one type of covariates (i.e., spectral, texture, biophysical, and geomorphometric). The average R-squared of studies that used more than three types of covariates was 0.87 whereas that for one type of covariate was 0.67 (Fig. 6). However, the result was not statistically significant when Kruskal–Wallis test was performed.

Fig. 6
figure 6

R-squared by the count of covariates

Full size image

Machine learning

ML has been widely applied in AGB assessments of forestry (Chen et al. 2018), and they are slowly emerging in AFS. The accuracy of ML depends on the input data, the number of explanatory variables used, data resolutions, and other parameters (Morais et al. 2021). Nineteen different algorithms were used in AGB estimation. Random Forest (RF), Support Vector Regression (SVR), and Stochastic Gradient Boosting (SGB) had slightly higher R-squared than other models (Fig. 7). RF is the most frequently used machine learning algorithm because it has been implemented in remote sensing literature since the early 1990s (Strahler and Jupp 1990) and it can capture the complexity and non-linearity of relationships with less sensitivity to noise in the training data (Mascaro et al. 2014; Safari et al. 2017). Other popular algorithms include Multivariate Adaptive Regression Splines (MARS), Stochastic Gradient Boosting (SGB), and Support Vector Regression (SVR). Filippi et al. (2014) estimated the AGB of the riparian forest using MARS, SGB, and Cubist and found SGB to be the most accurate predictor. Though there were no significant differences between MARS and SGB estimates, SGB was less sensitive to input variables than MARS and Cubist. Likewise, Safari et al. (2017) found that RF and MARS outperformed Support Vector Machine (SVM) and Boosted Regression Tree (BRT) for low biomass forests. RF also performed better than Artificial Neural Networks (ANN) that emulates human learning through inter-connected processing units called nodes in a forest setting (Zhang et al. 2020).

Fig. 7
figure 7

The R-squared of frequently used machine learning algorithms for AGB estimation. Acronyms: ANN Artificial neural network, GBR Gradient boosting regression, LR Linear regression, RF Random forest, SGB Stochastic gradient boosting, SVR Support vector regression, XGBR Extreme gradient boosting regression

Full size image

However, some studies found other models to be better performing than RF. Pham et al. (2020) noted Extreme boosting regression (XGBR) to be performing slightly better than CatBoosting regression (CBR), Gradient boosted regression tree (GBRT), RF, and SVR in a riparian system (R-squared = 0.622, RMSE = 27.36 Mg/ha).

Besides these standard models, scholars have also developed hybrid algorithms for AGB estimation in forestry. Su et al. (2020) combined RF and spatial kriging to create a random forest regression kriging model that addresses the spatial autocorrelation effect not covered by RF alone. Pham et al. (2020) used the combination of XGBR and genetic algorithm to build an XGBR-GA model that outperformed CBR, GBRT, SVR, and RF regression for mangrove forests in Vietnam.

While machine learning models had higher R-squared in many studies, they did not always improve the AGB estimation. Almeida et al. (2019) used LiDAR and hyperspectral data to estimate the AGB of the Brazilian Amazon forest. The authors compared the performance of a linear model with Ridge Regularization, SVR, RF, SGB, and Cubist and found that LiDAR and hyperspectral images had a more substantial impact on AGB estimation than ML algorithms.

Opportunities and challenges for AGB estimation of agroforestry systems

Some of the unique opportunities and challenges for estimating the AGB of AFS are as follows:

Opportunities

  1. 1.

    Advances in AGB estimation of forest and agriculture systems There have been significant advances in AGB estimation of forests and agriculture (Lu et al. 2016). As a combination of agriculture and forestry, AFS can benefit from methodological advances in these sectors (Bégué et al. 2018; Ahmad et al. 2021). For example, riparian buffers and food forests are very similar to forestry systems, and therefore forest-related methodologies could be applied to these systems with some modifications. Similarly, the AGB estimation of alley cropping could benefit from multitemporal remote sensing applications in agriculture settings (Karlson et al. 2020).

  2. 2.

    Advances in cloud computing One of the main barriers to the wider adoption of remote sensing has been data and computation needs. Over the last decade, there have been significant advances in cloud computing platforms such as Google Earth Engine (GEE), Microsoft Azure, and Amazon Web Service. In addition, some of these services are tailored toward remote sensing analysis. For example, GEE, a platform released by Google, allows users to have access to multiple geospatial datasets and achieve parallel programming using GEE’s in-built library (Kumar and Mutunga 2018; Amani et al. 2020). These platforms are increasingly used for biomass estimation (Yang et al. 2019; Xie et al. 2022), and they are likely to expand in the future.

  3. 3.

    High-resolution data sources There has been rapid advancement of spatial data acquisition methods that produce high spatial and temporal resolution datasets. High-resolution sensors like GeoEye, WorldView, IKONOS, EROS-B, Pleiades, and PlanetLab are already being used in AFS research. In addition, Unmanned Areal Systems (UAS) has also expanded data collection capacity at a very-high resolution (Pádua et al. 2017). The availability of these data products and widely accessible cloud computing platforms is expected to expand biomass research of AFS.

Challenges

  1. 1.

    Identification of agroforestry features Accurate identification of agroforestry over large areas would require very high-resolution imagery and specialized methods due to the small size and/or narrow width of AFS (Czerepowicz et al. 2012). These systems have been mapped with high accuracy for Trees Outside Forests, windbreaks, and riprain buffer (Liknes et al. 2017), however, narrow features and presence of shadows hinder accurate AFS delineation.

  2. 2.

    Measurement of understory vegetation The areal remote sensing is generally less accurate in estimating the understory canopy biomass. The insufficient LiDAR returns from heterogeneous vegetation cover create uncertainty in assessing the understory canopy AGB (Li et al. 2015). Advanced models like the radiative transfer model can help detect the contribution of understory crops towards measuring the vegetation cover, but they are applicable to a small areal extent (Hornero et al. 2021).

  3. 3.

    Allometric equations The trees grown in the open space of AFS accumulate more branch biomass than those grown in forest areas. Many allometric equations used for agroforestry AGB estimations are derived from the forest environment, potentially underestimating the biomass calculation (Zhou et al. 2011). In addition, since AGB differs by geography, agroforestry type, tree composition, tree age, and site quality, existing equations inadequately capture these variations (Chave et al. 2005).

  4. 4.

    Biomass estimation error The diverse tree density of AFS creates additional errors for the AGB estimation. Algorithms underestimate AGB in a forest with dense canopy cover due to saturation of pixels but over-estimate it in the area with thin canopy cover due to sub-pixel heterogeneity. Background soil reflectance can be problematic to estimate AGB in forested areas with low tree density. To some extent, the issue could be minimized by using high-resolution data and multiple parameters for feature extraction. In addition, the accuracy of remote sensing relies on the in-situ AGB measurements of sample plots for model calibration and validation (Wang et al. 2019).

  5. 5.

    Access to geospatial data and technology The cost and access to satellite imagery and technology are still major barriers for its wider application (Smith and Doldirina 2008). While many medium- and coarse-resolution imageries are freely available, high-resolution imageries are costly. Similarly, accessibility of satellite imageries is another hurdle (Turner et al. 2015). While data providers have developed specific websites and tools to facilitate easy access (e.g., https://sentinel.esa.int/), the limited information and technical barriers restrict their use, especially in developing countries where AGB is widely practiced.

Conclusion

This study reviewed methods of AGB estimation of AFS using remote sensing. A total of 33 papers were reviewed in detail covering all types of AFS across the world. Since remote sensing can assess a large area of land in a quick and efficient manner, it has been widely implemented in AGB research. High-resolution imageries are increasingly used for detecting heterogenous AFS through pixel-and object-based classification methods. For AGB estimation, scholars used regression, machine learning and statistical models. However, the study finds no statistically significant differences between the models in terms of performances (R-squared). Scholars have incorporated different types of covariates, including spectral bands, spectral indices, texture, biophysical and geomorphometric variables. Similarly, there’s no statistically significant differences in model performance with the addition of covariate types. While the performance of ML varies by the input parameters, spectral and spatial resolutions, and type of sensors, non-linear algorithms such as Random Forest and Support Vector Machine are some of the most widely used algorithms. The advancements in cloud computing like Google Earth Engine and the availability of high-resolution dataset presents opportunities for wider use of remote sensing in AFS biomass estimation.

Remote sensing and machine learning technologies will become increasingly valuable in the coming years, as working lands are tapped as spaces to store carbon. Compared with other types of agricultural landscapes, AFS is uniquely suited as habitats for applying for carbon credits, because much of the sequestered carbon is detectable in the aboveground biomass. Unlike carbon stored the in soil, the aboveground pool can be verified, estimated, and monitored without extensive on-the-ground sampling. The ability to apply carbon credits to site-scale plantings, with direct financial benefits for individual landowners, would be an important breakthrough for distributing carbon storage broadly across the landscape. The technologies explored in this paper offer a foundation for moving that dialog forward.