Capturing the Scale Dependency of Erosion-Induced Variation in CO2 Emissions on Terraced Slopes

Net soil CO2 emissions are not independent of topography but tend to decline with increasing slope gradients. Such decline has been attributed to increased runoff and greater soil loss on steep slopes, leaving the soil less habitable for microorganisms. However, the specific variations of slope gradients and thus the associated soil properties relevant for CO2 emissions, especially from terraced slopes, are often disguised by the coarse resolution of digital terrain models (DTMs) based on commonly available open-source data. Such misrepresentation of the relationship between topography and soil CO2 emissions carries the risk of a wrong assessment of soil-atmosphere interaction. By applying a slope dependent soil CO2 emission model developed from erosion plots to nearby sloping and partially terraced cropland using two DTMs of different spatial resolutions, this study tested the significance of these resolution-induced errors on CO2 emission estimates. The results show that the coarser-resolution Shuttle Radar Topography Mission (SRTM) underestimated CO2-C emission by 27% compared to the higher-resolution DTM derived from Unmanned Aerial Vehicles (UAV) imagery. Such difference can be mostly attributed to a better representation of the proportion of flat slopes in the high-resolution DTM. Although the observations from erosion plots cannot be directly extrapolated to a larger scale, the 27% underestimation using the coarser-resolution SRTM DTM emphasizes that it is essential to represent microreliefs and their impact on runoff and erosion-induced soil heterogeneity at an appropriate scale. The widespread impact of topography on erosion and deposition on cropland, and the associated slope-dependent heterogeneity of soil properties, may lead to even greater differences than those observed in this study. The greatly improved estimation on CO2 emissions by the UAV-derived DTM also demonstrates that UAVs have a great potential to fill the gap between conventional field investigations and commonly applied coarse-resolution remote sensing when assessing the impact of soil erosion on global soil-atmosphere interaction.


INTRODUCTION
A substantial number of studies have utilized a combination of field data and remote sensing to assess the potential impacts of soil erosion on atmospheric CO 2 at various spatiotemporal scales (Nadeu et al., 2012;Yue et al., 2016;Borrelli et al., 2017;Wilken et al., 2017;Lugato et al., 2018). One of the common complications in this research is the compatibility between the resolution of digital terrain models (DTMs) and the field topographic features and processes to be represented (Lu et al., 2020). Presumably, large scale topographic maps (e.g., 1:10,000) can produce DTMs of more desirable accuracy, but highly accurate topographic data is not always available, processable by available computers, or is often too expensive to be acquired. Yet, freely accessible data sources with lower resolutions often bear great uncertainties when assessing slope-scale soil erosion risk and impacts. The rates of erosion by water and gravity, including tillage and intentional soil redistribution, are all slopedependent and much more sensitive to minor topographic variations than a coarse resolution grid cell DTM would reflect (Kirkby, 2010;Balaguer-Puig et al., 2018;He et al., 2020). This in turn introduces potentially significant uncertainties when accounting for erosion-induced spatial redistribution of soil water and nutrients (Li et al., 2018), and hence the vertical CO 2 exchanges with the atmosphere in a landscape characterized by small-scale soil redistribution (Du et al., 2020b).
Previous studies have demonstrated that slope-induced erosion can potentially affect the spatial distribution of soil water, nutrients and microbial communities (Du et al., 2020a, Du et al., 2020bHu et al., 2016;Wang et al., 2014, Wang et al., 2017b, which all possibly introduce variations to CO 2 emissions from eroded soil (Nitzsche et al., 2017;Hu et al., 2020b). In particular, an exponentially declining pattern of soil CO 2 emissions was observed when slope gradients increased from 0 to 20° , highlighting the potential uncertainty in simulated greenhouse gas contributions from sloped and terraced fields to the atmosphere. Assessing emissions accurately therefore requires an effective DTM to represent slope dependent CO 2 emissions in soil-atmosphere models. However, the impact of features such as narrow terraces or short slopes on soil respiration is not captured by raster cells representing hectares to square kilometers in size commonly used in regional or global scale CO 2 emission models (Yue et al., 2016;Borrelli et al., 2017;Lugato et al., 2018). Even if those models had a greater spatial resolution, easily available open source topography data, such as the 30 m × 30 m Shuttle Radar Topography Mission (SRTM) data of the US Geological Survey or similar elevation data, do not capture the impact of microreliefs on soil respiration. Laser scanning techniques, such as airborne laser scanning (ALS) or terrestrial laser scanning (TLS), can create high resolution and accurate DTMs, but the heterogeneous point densities, undesired return signals from vegetation, prolonged working hours with multiple scans, and sophisticatedly designed tie points to align orientations of individual scans largely hinder their application to capture microreliefs in rugged landscapes with complex terrains (Baltensweiler et al., 2017;Ou et al., 2021). With fast development of photogrammetric techniques and increasing accessibility of Unmanned Aerial Vehicles (UAVs) and highresolution cameras, improved DTMs for a large area, even with complex terrains, can be easily obtained within a short time by stitching highly overlapping images (Schwanghart et al., 2013;Krenz and Kuhn, 2019;Hu et al., 2020a). Hence, this promises great opportunities to accurately assess the scale-dependent variation of CO 2 emissions on terraced slopes.
This case study was conducted on a terraced field on the Chinese Loess Plateau (Figure 1), where soilscapes have been fragmented intensively by establishing check-dams and terraces (Ostwald and Chen, 2006;Sun et al., 2014;Wang et al., 2017a). The Chinese Loess Plateau thus recommends itself as a suitable area to identify the potential role of a misrepresentation of slope dependent soil redistribution on CO 2 emissions. To assess the potential errors for such topographic scale-dependent CO 2 emissions on eroding cropland, a simple slope dependent CO 2 emission model developed on the same loess soil, was adopted and applied to two DTMs, derived from SRTM and UAV imagery respectively. The aim of this study was to identify the potential errors of CO 2 emissions on terraced slopes induced by ignoring microreliefs.

Study Area
The study area is in the Wangdong Catchment (35°13′N-35°16′N, 107°40′E-107°42′E, average altitude: 1,220 m a.s.l.), southwestern part of the Chinese Loess Plateau. Highly erodible loess soil combined with concentrated rainfall in summer, and further accelerated by intensive cultivation, once made this region one of the most eroded region in the world. In consequence, the plateau had been incised into tableland (fragmented plateau), slopes (or gullies) and deposited sediment in valley bottoms (Wang et al., 2017a). To expand agriculture land, terraces had been established on almost every aspect of all the slopes to increase infiltration ( Figure 1), which were then converted back to forest, grassland or natural rehabilitation in 1990s to curb soil erosion.

Acquisition of Terrain Models From Shuttle Radar Topography Mission and Unmanned Aerial Vehicle
Two DTMs with different spatial resolutions were applied to the Wangdong catchment that had been remained natural rehabilitation since 1990s, specifically on an east-facing terraced slope (about 0.1 km 2 ) with uneven heights and widths ( Figure 2). One DTM was based on the open source SRTM from the US Geological Survey (30 m × 30 m), accessed on June 15 th , 2018 with raster projected to WGS 1984 UTM Zone 48N. The other DTM was created based on the Structure from Motion (SfM) photogrammetric technique using the UAV imagery obtained on April 19 th , 2018. The slope gradients of the study area were then calculated from the two DTMs with ESRI ArcGIS. The UAV used in this study was a Phantom 3 Professional quadrocopter with a gimbal manufactured by DJI (SZ DJI Technology Co., Ltd., Shenzhen, China). The camera was DJI FC300X (GoPro Inc, San Mateo, California, United States ), with a 1/2.3" 12.4-megapixel CMOS sensor and an exposure time automatically adjusted. Images were taken at an altitude of 100 m above the plateau with an angle of 90°and an average overlap of 80%. Given that the width of terraces on the selected slope range from 3 to 15 m, the ground sampling distance (GSD) of 5.25 cm was considered adequate to reflect the essential features of terraced slopes. Navigation was GPS-based (under coordinate system WGS 1984 UTM Zone 48N) using an image acquisition program of Pix4D Capture App on iPad min A1550. We acknowledged that the resolution of the DTM could be compromised by the changes of altitude along the descending slope. Hence, five ground control points (GCPs) were placed before the arial flight at different altitudes with a "zig-zag" pattern to the furthest point one could manage to descend. The ground coordinates of the GCPs were recorded by a real-time kinematic positioning apparatus (RTK, CHC-i60, Huace, Shanghai, China) coupled with a HCE300 field surveying controller (Huace, Shanghai, China). The coordinates were recorded under China CGCS 2000 by Gaussian Projection. The horizontal accuracy of the RTK system was ± (10 + 1 × 10 −6 × distance) mm and the vertical accuracy was ± (20 + 1 × 10 −6 × distance) mm.
All the UAV images were processed using the commercial photogrammetry software Pix4Dmapper Pro (Pix4D SA, Lausanne, Switzerland) and georeferenced according to the GPS positions following the steps listed in Krenz and Kuhn  (2019). In brief, after imported into the software, all the images were automatically matched, and bundle-block adjusted. The accurate geolocations of the five GCPs were then added and marked manually on different images to produce a densified point cloud and a 3D textured mesh, which were further used to obtain a DTM (with resolution of 5 × GSD) and a georeferenced orthomosaic. Slope gradient was determined for each pixel of the UAV imagery using QGIS based on the approach developed by Zevenbergen and Thorne (1987). Since the potential influences of vegetation on CO 2 emissions were out of the reach of this study, areas covered by vegetation were identified on the DTM and then excluded from calculation. Areas with a gradient >30°were not covered by soil, and thus were digitized manually from the UAVbased DTM and excluded from the calculations to avoid any overestimation of emissions. The SRTM-DTM did not show any areas with slopes >30°, and consequently, no areas were excluded in agreement with the purpose of this study.

Slope-Dependent Soil CO 2 Emissions
The slope dependent CO 2 emission model (Eq. 1) was developed from the year-round observations on four erosion plots with different slope gradients: 0.5°, 5°, 10°, 20° (Figure 3). The development of this slope dependent CO 2 emission model has been reported in Hu et al. (2020a). The slope dependent CO 2 emission model reads as: where y represents the yearly CO 2 -C (g m −2 yr −1 ) at slope gradient of x (°) with an R 2 of 0.90. The four east-facing erosion plots were established in the nearby Changwu State Key Agro-Ecological Experimental Station by refilling with the same loam soil (Cumulic Haplustoll as in the USDA Soil Taxonomy System) developed from loess deposits as in the terraced fields (Du et al., 2020a). After establishment, the erosion plots were remained uncultivated to simulate the natural rehabilitation as in the terraced fields in the Wangdong Catchment. Specific descriptions on erosion plots and their setting up have been reported in Du et al. (2020a) (Figure 3). In brief, each erosion plot was 5 m × 1 m with three replicates, to represent the common slope gradients on the Loess Plateau (Shen et al., 2016). At the lower end of each erosion plot, there was an outlet channeled into a sediment trap to collect runoff and sediment samples (Figure 3). During the experimental year from October 2014 to September 2015, there were in total 534.5 mm precipitation, with five rainfall events that generated detectable runoff and sediment discharge. The runoff and soil net erosion rate reached 4.56 mm and 125.4 g m −2 , respectively . Although the erosion plots employed in this study were slightly smaller than typical terraces in the study area (3-15 m), previous observations have proved their adequate length and inclination to distinguish the effects slope gradients on soil CO 2 emissions (Wang et al., 2017b;Du et al., 2020a). The soil CO 2 emission rates on each eroding slope were measured every 10 days during the experimental year, following the protocol described in Du et al. (2020a). With the semi-arid climatic conditions and well-drained loess soil, we presumed that the likelihood of carbon emission as methane was low (Brzezińska et al., 2012). Therefore, carbon mineralization through CO 2 emissions was described as CO 2 -C in this study to illustrate that our measurements captured only CO 2 emissions without considering other gaseous manners.

RESULTS AND DISCUSSION
In the coarse 30 × 30 m SRTM-derived DTM, only the major geomorphological features such as plateau and slopes were clearly distinguished (Figure 4). It did not capture the narrow width of the terraces (3-15 m) and attributes mostly intermediate slope angles to the terraces. The UAV-derived DTM, on the other hand, reflected the terracing on slope length and gradient well, i.e., flat steps separated by retaining walls or naturally steep ravine sideslopes. Applying the CO 2 emission model to the selected terraced FIGURE 3 | Slope-dependent CO 2 emission model (A) developed from erosion plots (B) nearby the Wangdong Catchment, Chinese Loess Plateau. More descriptions on the slope-dependent CO 2 emission model, please refer to Hu et al. (2020b); to learn more about the setting up of erosion plots, please refer to Du et al. (2020a).
Frontiers in Environmental Science | www.frontiersin.org June 2021 | Volume 9 | Article 688286 field generated a CO 2 -C emission of 392 g m −2 yr −1 for the SRTM-derived DTM and 451 g m −2 yr −1 for the highresolution UAV-derived DTM ( Figure 5). When weighted by slope gradient distribution in the study area, the difference in CO 2 -C emission was even greater, generating 537 g m −2 yr −1 for the UAV-derived DTM and 423 g m −2 yr −1 for in the SRTMderived DTM ( Figure 5). The differences between the slope dependent CO 2 -C emission calculated from the two DTMs of the study site clearly demonstrate an inherent risk associated with ignoring the actual topography. Although the slope-dependent CO 2 emissions can vary with pixel size when the UAV-derived DTM was of different resolutions, the difference between the CO 2 emission values between SRTM-and UAV-derived DTMs can be considered as indicative of the sensitivity of emission calculations to the spatial patterns of soil along slopes affected by erosion and deposition. We must acknowledge that the comparison of CO 2 emissions derived from the two DTMs in FIGURE 5 | Comparison of yearly CO 2 -C emission from different slopes on a terraced field in the Wangdong Catchment calculated from two digital terrain models. Areas with a gradient of more than 30°were not covered by soil anymore, and therefore were manually excluded from the calculations made with the UAV-based DTM.
Frontiers in Environmental Science | www.frontiersin.org June 2021 | Volume 9 | Article 688286 5 this study is obviously biased by the terraced topography, thus leading to a larger area with intermediate slope gradients and in turn lower emission totals. However, a bi-modal distribution of slope gradient frequency is not uncommon in terraced land (Wei et al., 2016;Chen et al., 2017). Hence, reliable estimation on CO 2 emissions must be microrelief-specific, especially when considering that terraced fields constitute at least 10% of the entire cropland area globally (1.1 billion ha, Roser and Ritchie, 2019) and distribute widely in almost all the major grain producing regions. For instance, 22 million ha in Middle east and North Africa (Dixon et al., 2001), 13 million ha in China (Chapagain and Raizada, 2017), 133 million ha rice field in Asian-Pacific region that is often terraced (Papademetriou, 2000), as well as smaller areas such as vineyards in Europe and South America (Chapagain and Raizada, 2017;Varotto et al., 2019).
In more general terms, the greater emissions from the UAVderived DTM illustrate that when the raster cells in DTMs were greater than the actual topographic features, it would potentially lead to significant errors of agricultural land CO 2 emissions. Such errors are dependent on how close the pixel size is to the microreliefs and would thus be particularly pronounced in landscapes with short slopes and contrasting erosion and deposition-induced soil development along them (e.g., Van Oost et al., 2009;Greenwood et al., 2015;Hu et al., 2016;Lu et al., 2020). However, such discrepancy induced by using improper pixel size to model slope dependent impacts on surface processes or biogeochemistry is not uncommon in recent global and regional-scale studies addressing erosion and climate change (e.g., 250 m × 250 m in Borrelli et al., 2017; 1 km × 1 km in Lugato et al., 2018). They would be too coarse to capture the potential impacts of finer scale topographic features on runoff, erosion and deposition, and thus soil respiration. Soil disturbance introduced by human activities, such as establishing terraces, conducting tillage and land leveling practices, or planting different vegetations, can also introduce spatial and temporal variations in slope-dependent CO 2 emissions Li et al., 2020), further highlighting the essence to create DTMs with proper pixel size and resolution in assessing regional carbon balance.
Overall, the slope dependency of CO 2 emissions in this study is a showbox exhibiting the decisive role of the DEM resolution and pixel size in calculating regional carbon balance. Certainly, the flight protocol and resolution of DTM employed in this study cannot be directly applied to estimate CO 2 emissions from other regions with local reliefs, different soil properties and land use types. A sensitivity study with different resolutions of UAVderived DTM is essential to understand at which pixel size occurs the most significant variations in the slope dependent CO 2 emissions. The greatly improved estimation of CO 2 emissions by the UAV-derived DTM further demonstrates that UAVs with high spatial resolution have a great potential to fill the gap between conventional field investigation and commonly applied coarse-resolution remote sensing for assessing the impact of soil erosion on soil respiration. While the resolution of global DTMs might still not capture microreliefs relevant for erosion, combining UAV-derived DTMs with soil biogeochemical data could be used to enable the development of parameters that capture the influence of topography on soilatmosphere interaction at a landscape scale. Moreover, sequential UAV investigations at decadal intervals also promise a good opportunity to enable stakeholders to evaluate the flattening of deserted terraces evolved over repeated erosion processes, and the thus induced variations in regional carbon balances.

CONCLUSION
This study applied a slope-dependent CO 2 emission model to calculate yearly carbon emissions from a partially terraced study site on the Loess Plateau, with the aim to identify possible errors introduced by DTMs of different spatial resolutions. After weighted according to slope gradient, the estimated CO 2 -C emission was 27% greater for a highresolution DTM than for the coarse-resolution SRTM-based DTM. The difference was attributed to the overestimation of slope steepness in the SRTM-based DTM and associated lower respiration rates. While this difference is biased by the terraced slope we selected for this study, our results still highlight that soil CO 2 emissions were sensitive to small topographic features. This indicates that topography and associated surface processes such as erosion and deposition, and soil properties on sloping land, must be represented on a scale that is adequate to capture soil-atmosphere carbon fluxes in globalscale biogeochemical models. The use of UAVs offers the possibility to create topographic data with high spatial resolution, filling the gap between conventional field investigation and commonly applied coarse-resolution remote sensing for assessing the impact of erosion on soil respiration.

DATA AVAILABILITY STATEMENT
The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author.

AUTHOR CONTRIBUTIONS
YH assisted the field investigation with the UAV, analyzed the slope-dependent CO 2 model and wrote the manuscript; VS processed the UAV data and SRTM data; BK conducted the UVA flight in the field and supervised VS for data processing; SG offered local geo-and climate-information and field facilities; NK conceived the concept and structured the manuscript.