Tree-dominated ecosystems are crucial for climate change mitigation and adaptation as they support biodiversity, store carbon (C), provide habitats, and offer nutritional and economic benefits. Tree density, size, and distribution within an ecosystem are key attributes and central to decision-making regarding landscape restoration efforts aiming to increase tree-based ecosystem services and promote optimal tree growth. Among many existing landscape restoration initiatives, agroforestry, which involves the integration of trees within croplands, is increasingly being adopted in the Global South. Agroforestry has the potential for concurrent intensification and diversification of production and enhanced carbon storage. Furthermore, trees within croplands fertilize and stabilize the soil, and provide numerous benefits to farmers including enhanced income as well as improved food and nutrition security. Contrastingly, natural forests provide different ecosystem services, such as high C stocks, biodiversity, and ecological habitats. However, natural forests are threatened by both natural and anthropogenic processes, often leading to deforestation and forest degradation. The restoration potential of degraded natural forests (hereafter, degraded forests) is high, and once restored, regrowing forests can store carbon up to 20 times faster as compared to old-growth forests, which is in particular the case in the tropics. Identifying degraded forests within protected areas is however challenging. As a result, many large-scale restoration programs prioritize agricultural land, where tree cover is typically low and presumed to be easier to increase rapidly. Furthermore, it is challenging to monitor the success of large-scale restoration efforts within degraded forests, as quantifying newly established individual trees is difficult using traditional satellite systems and often requires labor-intensive field inspections that may not be representative for large areas. The same difficulties apply to farmlands, where trees are typically sparsely planted, and single trees are too small to be identified by conventional satellite systems, especially when trees are young. Consequently, studies often rely on proxies to evaluate the success of ecosystem restoration, such as the greenness of an area. However, greenness is a measure of the presence and abundance of green vegetation, and it is not always related to tree cover. Tree-level traits such as density, crown area, and carbon stock dynamics related with tree plantations and forest regeneration have rarely been reported for large areas over extended time periods. Furthermore, conventional methods for monitoring restoration activities usually focus on larger land units, excluding smallholder farms with less than 10 ha. Therefore, the contribution of smallholders to landscape restoration goes often unnoticed and remains largely unrecognized. Here, we conduct a nation-wide wall-to-wall mapping of trees (defined here as woody plants with a crown area larger than 3 m²) in 2008 and 2019 in Rwanda, and study changes in their number, crown area and carbon stock, with a particular focus on newly established trees in farmlands and degraded forests. Rwanda has invested a remarkable effort in restoration, and is thus an interesting case to illustrate potential gains in forest and non-forest biomass from restoration in tropical countries. We demonstrate how our data can be utilized to estimate future potential carbon sinks needed to reach net zero emissions by 2050, and to what extent carbon gains resulting from smallholder farms and forest restoration contribute towards achieving this target.

The introduction section cites several relevant studies on the importance of tree-dominated ecosystems for climate change mitigation and adaptation, the increasing adoption of agroforestry in the Global South, the challenges of monitoring large-scale restoration efforts, and the limitations of using proxies like greenness to assess restoration success. These references provide context for the study's focus on tree-level analysis of carbon stocks in Rwanda's farmlands and degraded forests. The cited literature highlights the gap in existing research, namely the lack of detailed, tree-level monitoring of restoration efforts, particularly on smallholder farms, which motivates the current study's methodology and contribution.

This study uses a combination of high-resolution aerial and satellite imagery (0.25 x 0.25 m² in 2008 and 0.5 x 0.5 m² in 2019) from Rwanda to map individual trees at a national scale. The images were pre-processed to correct for differences in resolution, spectral bands, and cloud cover using deep learning models. A UNet-based deep learning model, trained on a large dataset of manually delineated tree crowns (325,540 total), was used to segment individual tree crowns in both datasets. The minimum crown size threshold was set at 3 m², based on visual inspection. Clumped trees in dense canopy areas were separated in a post-processing step to improve accuracy. The study uses existing national forest cover maps and agricultural spatial extent maps to delineate farmlands and degraded forests. For carbon stock estimation, the researchers employed an allometric approach, using previously established equations to relate crown diameter (derived from crown area), diameter at breast height (DBH), and above-ground biomass (AGB). AGB was then converted to above-ground carbon stock (AGC) using a biomass-to-carbon conversion factor. The uncertainty of this approach was evaluated using data from 296 permanent field plots with tree measurements. The methodology also incorporates data on farm plot delineations from the National Land Authority of Rwanda to analyze tree gains on smallholder farms. The accuracy of the tree crown segmentation model was evaluated using an independent test set (52 labeled patches, 17,306 trees), resulting in a high accuracy rate. The accuracy of change estimates between 2008 and 2019 was further validated against measurements from 296 permanent field plots. Finally, a cost-benefit analysis of tree restoration potential was conducted using existing cost estimations from the Rwandan Ministry of Environment. Specific methods were used to account for uncertainties introduced during the estimation of tree biomass and carbon stocks from crown area measurements. The study specifically addresses the limitations of previous approaches, which frequently relied on proxies such as greenness rather than direct tree-level measurements.

The study found that smallholder farmers in Rwanda planted an average of 3.2 trees per farm plot between 2008 and 2019, resulting in a national gain of approximately 50.4 million new trees. The carbon sink from these new trees in farmlands was 0.13 Mg C ha⁻¹ yr⁻¹, significantly lower than the carbon sink in restored degraded forests (0.76 Mg C ha⁻¹ yr⁻¹). The total number of trees in degraded forests increased from 1.5 million to 2.1 million during the same period. The study estimated total carbon stock in agroforestry areas increased from 5.2 ± 0.3 Tg C in 2008 to 6.3 ± 0.36 Tg C in 2019, with 29.7% of this increase attributed to newly established trees. In degraded forests, carbon stocks increased from 0.38 ± 0.12 Tg C in 2008 to 0.79 ± 0.26 Tg C in 2019, with 23.7% from new trees. The carbon density of newly established trees was significantly higher in degraded forests (8.4 Mg C ha⁻¹) compared to agroforestry fields (1.4 Mg C ha⁻¹). Considering the national greenhouse gas (GHG) emissions in 2019 (1.4 Tg C), the combined carbon sequestration from agroforestry and forest restoration offset roughly 10% of these emissions. However, a combined restoration scenario reaching optimal tree density (300 trees ha⁻¹ in agroforestry and 700 trees ha⁻¹ in degraded forests) has the potential to offset about 80% of national GHG emissions. This scenario requires a significant investment (~1.163 billion USD over 20 years), but provides a substantial return on investment from increased crop and timber yields, erosion prevention, and carbon credit revenues. The average number of trees per smallholder farm increased from 10.9 to 14.1, suggesting a significant contribution from these small farms to national carbon sequestration efforts despite their comparatively lower carbon sequestration rates per unit area compared to degraded forests. The study also highlights the accuracy of the tree level mapping with an independent test set indicating an accuracy of 93-95%.

The findings demonstrate the substantial contribution of both agroforestry and forest restoration in Rwanda to carbon sequestration and the potential to achieve significant emission reductions. The six-fold difference in carbon sequestration rates between restored degraded forests and agroforestry highlights the importance of prioritizing forest restoration efforts while simultaneously engaging smallholder farmers in tree planting initiatives. While farmlands contribute substantially to the total carbon sink due to their large area, the lower carbon sequestration per hectare emphasizes the need for optimized agroforestry practices to maximize both agricultural yields and carbon storage. The study's approach of mapping individual trees using high-resolution imagery and deep learning provides a powerful tool for monitoring the effectiveness of landscape restoration projects. This enables more accurate assessments of carbon stock changes at the tree and farm level, leading to improved understanding of the contribution of smallholder farmers to climate change mitigation efforts. The cost-benefit analysis suggests a considerable return on investment, underscoring the economic viability of large-scale restoration initiatives and their potential to engage smallholders in carbon markets. The integration of these findings with ongoing initiatives like large-scale tree planting campaigns provides a critical step toward Rwanda's net-zero emission goals. However, the success of these initiatives relies on careful planning, optimization of agroforestry practices, and the development of effective mechanisms for engaging smallholder farmers in carbon markets.

This study demonstrates the significant contribution of trees on smallholder farms and forest restoration to Rwanda's carbon sequestration efforts. The high-resolution mapping of individual trees revealed a large increase in tree density in both agroforestry areas and degraded forests between 2008 and 2019. While degraded forest restoration showed significantly higher carbon sequestration rates, the vast area of farmland implies a substantial overall contribution from smallholder farmers. The study highlights the potential for achieving substantial emission reductions through optimized restoration efforts, but also points to the need for further research on effective ways to engage smallholder farmers in carbon markets and on the economic and logistical challenges of large-scale restoration projects. Future research should focus on enhancing the methodological approach by integrating additional data sources like LiDAR for improved accuracy and extend the methodology to other regions using freely available satellite imagery.

The study acknowledges limitations regarding data quality, particularly related to image quality, as the results are sensitive to the availability and quality of aerial and satellite images. The cost-benefit analysis presented a simplified estimation that disregards social and landscape variations; a more comprehensive econometric forecast is recommended for practical implementation. The study focuses solely on aboveground carbon stocks, neglecting belowground carbon; incorporating belowground carbon would provide a more comprehensive view of carbon sequestration. Furthermore, the analysis does not account for all potential ecosystem services related to the planted trees, focusing primarily on carbon sequestration. The analysis relies on the assumption that GHG emissions will remain constant at 2019 levels, which may not hold true in the future. Finally, the study acknowledges that inclusion of smallholders in carbon markets has political, economic, and regulatory implications that were outside the scope of the current study.