The global atmospheric carbon concentration has increased from 277 ppm (ppm) in 1750 (beginning of the industrial era) to 412 ppm in 2020 (Friedlingstein et al. 2022). Increasing carbon levels have led to a further rise in global temperature and changes in other climate quantities. Carbon capture, utilization, and storage options are pivotal in mitigating climatic effects and achieving Sustainable Development Goals (SDGs), particularly clean energy (SDG 7) and climate action (SDG 13). Forests are an important part of the global carbon budget, capturing and storing several tons of carbon in their living biomass, and are therefore, a useful option for carbon sequestration (Fahey et al. 2010). Thus, a rapid, accurate, and non-destructive measurement of forest carbon stock and change is crucial for planning forest management and carbon emission reductions.

Simply, forests account for almost 86% of the global vegetation carbon pool. However, carbon capture and storage could vary with stand, age, composition, precipitation, temperature, radiation, management, disturbance, and harvesting frequency, among other factors (Dixon et al. 1994). Inventory-based methods and allometric approaches are commonly used to estimate carbon stored in the aboveground biomass. Inventory data are a part of complete, systematic, and annual sampling and surveys of plots within each state under the Forest Service, Forest Inventory and Analysis (FIA) program of the U.S. Department of Agriculture Forest Services (Bechtold & Patterson, 2005; Smith et al. 2019). Allometric approaches, on the other hand, are based on regression models using morphological traits (Avalos et al. 2022; Jones et al. 2020). These methods appear to have comparatively low coverages and suffer from sampling and estimation errors (Weiskittel et al. 2015; Picard et al. 2015). Drones capture forest images to derive morphological/structural parameters, which could be used in allometric equations. Images captured with RGB or Light Detection and Ranging (LiDAR) sensors on drones generate dense point clouds for monitoring forest cover and therefore provide a low-cost alternative to inventory methods and satellite-based sensing. Integrating drone imagery and Machine learning (ML) models could provide a useful tool in carbon estimation over an extensive area with improved precision; thereby facilitating forest management and long-term carbon capture and storage.

Drones are increasingly being used for on-demand, rapid, cost-effective, and near real-time characterization, monitoring, and mapping of forest cover at much finer scales. Miller et al. (2017) used a multirotor drone with a Canon ELPH 520 (Canon Inc., Tokyo, Japan) HS point and shoot digital camera (RGB sensor) and highlighted the potential to precisely measure tree height and biomass in monoculture plantations when they were used with accurate terrain models; thereby, facilitating carbon budgeting. Jones et al. (2020) used a 20-megapixel Sony RX100iii camera mounted onto a 3DR Iris+ drone to capture RGB images in generating a three-dimensional model and orthomosaics of Mangrove stand in Southern Australia, which were then used in the estimations of canopy area and canopy height to further model diameter at breast height (DBH) and diameter at 30 cm (basal diameter; D30). Subsequently, two approaches, drone image-based measurement of tree structure (height, canopy area, DBH, D30) and drone-image based measurement of structural characteristics in allometric equations, were used in predicting above-ground biomass for carbon estimation. Indeed, potential exists to integrate drone-based forestry features with data-driven approaches like ML, Deep Learning, and time-series forecasting models in estimating and modeling forest carbon storage.

Machine Learning is “a branch of artificial intelligence and computer science which focuses on the use of data and algorithms to imitate the way that humans learn, gradually improving its accuracy” (IBM, 2020). Deep Learning is a subset of ML that utilizes artificial neural networks. We see the potential for using different Auto-Regressive Integrated Moving Average (ARIMA) models along with different variants of Deep Learning like Long Short-Term Memory (LSTM) units and Gated Recurrent Units (GRUs) to forecast forest indices for carbon estimation. A few studies have utilized ML models and drones in classifying forest areas and estimating coverage showing 93% accuracy (Haq et al. 2021); detecting forest tree species, forest fires, and disease and insects with a median accuracy of 80% (Abid, 2021; Syifa et al. 2020; Diez et al. 2021); estimating timber quality and carbon storage (Baccini et al. 2012; Wang et al. 2021; Mascaro et al. 2014); investigating plantation and restoration, and conducting policy analysis (Miller et al. 2017; Firebanks-Quevedo et al. 2022). For example, Convolutional Neural Network (CNN)- based Deep Learning models for real-time target detection and recognition like YoLov3 and YoLov5 have been used in predicting sawn timber moisture content (Wang et al. 2021). Other CNN-based algorithms like a combination of ResNet50 and SegNet architecture were utilized in forest image segmentation (Bhatnagar et al. 2020). Images collected by hyperspectral and LiDAR cameras or drone acquired-RGB data were used as inputs in the Deep Learning models to detect forest tree species (Mosin et al. 2019; Mascaro et al. 2014) and forest anomalies like forest fires and insect manifestation (Diez et al. 2021). Random forest ML and airborne LiDAR technology have also shown potentials for better estimating and modeling forest carbon and developing carbon density maps (Baccini et al. 2012; Mascaro et al. 2014). While research on forecasting growth attributes or structure of trees is limited, interest is mounting to facilitate carbon estimation and mapping for larger areas with improved accuracy.

Overall, we see the potential to utilize drone images to estimate important forestry features and indices like basal area, canopy cover, height, volume, and Normalized Difference Vegetation Index (NDVI) (Zhang et al. 2016). Such structural characteristics could be used in developing regression models or used in allometric equations in determining the carbon content in aboveground biomass at a point in time. The inclusion of climatic variables like rainfall, temperature and humidity in forecasting models further opens opportunities to refine algorithms for site-specific predictions. However, results from drones and ML models must be validated. Select destructive sampling of trees provide a way to cross-validate carbon estimation (Jones et al. 2020). Additional research on drones and ML would provide additional knowledge to our current understanding of forest carbon.