Mapping potential surface ponding in agriculture using UAV‐SfM

Among the environmental problems that could affect agriculture, one of the most critical is ponding. This may be defined as water storage on the surface in concavities and depressions due to soil saturation. Stagnant water can seriously affect crops and the management of agricultural landscapes. It is mainly caused by prolonged rainfall events, soil type, or wrong mechanization practices, which cause soil compaction. To better understand this problem and thus provide adequate solutions to reduce the related risk, high‐resolution topographic information could be strategically important because it offers an accurate representation of the surface morphology. In the last decades, new remote sensing techniques provide interesting opportunities to understand the processes on the Earth's surface based on geomorphic signatures. Among these, Uncrewed Aerial Vehicles (UAVs), combined with the structure‐from‐motion (SfM) photogrammetry technique, represent a solid, low‐cost, rapid, and flexible solution for geomorphological analysis.


| INTRODUCTION
Water storage on terrain surfaces may be a serious problem in agriculture, especially in lowland landscapes. It can be described as a surplus of water that remains on the surface due to prolonged rainfall, snowmelt, or irrigation (Chen et al., 2000;P asztor et al., 2015;Saadat et al., 2020). The negative effects are manifested in various forms.
First, it causes serious damage to crops at the biological level, mainly related to the lack of oxygen supply (Gupta et al., 2004;Singh, 2015).
Also, soils exposed to long ponding periods may suffer a reduction in fertility (Dunin, 2002). In this framework, the crop yield may decrease significantly (McFarlane & Williamson, 2002). Another aspect refers to the in-field accessibility at the proper time for optimal agricultural management (i.e. trafficability; Chipanshi et al., 2018).
The first cause of ponding could be found in the poor capacity of the soil to drain water, a circumstance that occurs when the water input exceeds the potential infiltration (Haghnazari et al., 2015). In particular, the soil's hydraulic conductivity is significantly influenced by texture, skeleton, aggregates, and the pore system (Archer et al., 2002;Ben-Hur et al., 2009;Saxton et al., 1986;Servadio et al., 2001). For example, soils with high clay content are interesting, where permeability changes according to moisture conditions (Bronswijk, 1991;Yaalon & Kalmar, 1984). Also, non-optimal agricultural practices, such as heavy machinery, can worsen this condition, especially regarding soil compaction (Berli et al., 2004;Bottinelli et al., 2014;Pijl et al., 2019a). This process reduces soil porosity and permeability, leading to changes in its characteristics and behaviour (Batey, 2009;Keller et al., 2019;Soane & Van Ouwerkerk, 1994). The rising of the shallow water table due to subsurface water is also a cause of ponding (Mueller et al., 2005;Salvucci & Entekhabi, 1995).
Another important factor causing water stagnation is the absence of drainage systems which can drain the water into the hydrographic network. The terrain topography may provide an important contribution, especially considering a factor such as a slope that can lead to the accumulation of runoff in specific surface concavities (Tarolli et al., 2019). Specifically, microtopography plays a key role. Indeed, terrain microstructures affect both the formation of runoff and the amount of water stored on the surface (Frei & Fleckenstein, 2014). In concave areas, the soil is rapidly subject to saturation, therefore surface stagnation can occur more frequently. This condition may easily happen in lowland fields if not properly managed.
Since topography is one of the determining factors in waterlogging, the analysis of the terrain's morphological characteristics could provide useful information for identifying potentially vulnerable areas and preventing waterlogging. Indeed, thanks to the rapid technological developments of the last decades, remote sensing offers new interesting prospects for knowledge of the Earth's surface. It becomes a strategic tool for identifying critical issues during agricultural production and provides support for designing sustainable solutions. High-resolution topography (HRT) techniques permit different opportunities to investigate agricultural surfaces, allowing 3D reconstruction of even the smallest surface signatures. HRT may be performed using diverse acquisition platforms (Cucchiaro et al., 2020;Gupta, 2017;Xue et al., 2008). For example, LiDAR allows the topographic reconstruction of an agricultural landscape at large spatial scale through an airborne laser scanner (ALS; Ladefoged et al., 2011;McCoy et al., 2011) or at a detailed scale, through a terrestrial laser scanner (TLS; Barneveld et al., 2013;Eitel et al., 2010). Nowadays, a rapid and low-cost solution is the structure-from-motion (SfM) photogrammetry technique paired with multi-stereo view (MSV) algorithms (hereafter SfM). This has been used in different studies in agriculture: to analyse microtopography on surfaces managed with conservation agricultural management (Tarolli et al., 2019); to monitor the bank erosion in drainage network (Prosdocimi et al., 2015); to measure the surface roughness of cultivated terrain surfaces (Martinez-Agirre et al., 2020;Snapir et al., 2014); and to estimate soil loss by erosion (Vinci et al., 2017). SfM may be used through different acquisition platforms that permit analysis at different spatial levels, from plot to field scale (Dong et al., 2017;Jay et al., 2015;Nguyen et al., 2016).
For larger-scale surveys, the use of UAVs (unmanned aerial vehicles or drones) allows the realization of high-resolution DEMs, indispensable support for the knowledge of the processes taking place on agricultural surfaces. For example, they are used in agriculture to map erosion and deposition (Meinen & Robinson, 2020); identify potentially unstable areas due to surface runoff and provide suitable solutions (Pijl et al., 2019b); measure the height of crops (Chang et al., 2017); and estimate the biomass (Bendig et al., 2014;Gil-Docampo et al., 2020). A DEM rich in detail and capable of describing terrain signatures opens up considerable opportunities for understanding the processes occurring on it. It may also be a useful tool in understanding the surface water storage phenomena. Tarolli et al. (2019) proposed an innovative and effective method to evaluate stagnant water's potential depth on an agrarian surface using small experimental plots (2 Â 2 m) and a DEM of 0.02 m pixel resolution derived using a handled mirrorless camera. However, they did not test the method at a larger scale or validate it with real ponding conditions. Other literature examples where DEMs have been used for similar purposes concern the investigation of snow patterns in agricultural landscapes (Lapena & Martz, 1996), laboratory analysis for water storage quantification (Chaplot & Le Bissonnais, 2003;Kamphorst & Duval, 2001;Ullah & Dickinson, 1979), or catchment-scale waterlogging analysis in agriculture (Merot et al., 1995).
This work aims to provide a robust, rapid, and low-cost method for mapping ponding on an entire farm's soil surface. To do this, a methodology based on a high-resolution DEM (0.10 m cell) derived from UAV-SfM is proposed. To our knowledge, UAV use for detecting ponding in agriculture is a novelty. Specifically, this approach investigates the terrain's signatures through a geomorphological indicator able to map the depressed areas of the terrain. For those areas, morphologically prone to water ponding, the spatial extent and the potential waterlogging depth were detected. The quality of the mapping was then compared with a solid statistical assessment based on a detailed field survey campaign, considering real ponding conditions.

| STUDY AREA
The study area is located within the municipality of Rovigo (RO), in northern Italy (45 06 0 07 00 N; 11 52 0 10 00 E; Figure 1a). It is a lowland area (slope < 1%), devoted to agriculture and historically subject to land reclamation, as well as numerous territories of the Po and Adige Valley (Curtis & Campopiano, 2014). Over the last years, the average annual precipitation was 750.4 mm, with a peak in 2013 of 1006.4 mm. The rainiest month was May, with an average cumulated rainfall of 84.8 mm (ARPAV, 2020). The phenomenon of standing water in the field was reported all across the year in the more depressed area. Typically, it remains on the terrain surface from a few days to a week, and more frequently during spring and autumn. These periods were also the rainiest, with more than 200 mm per season.
This fact was particularly critical for management, as sowing is usually scheduled in April. Indeed, the presence of pond water on the surface may cause delays in working time, widespread absence of seed germination, and seedling mortality.
The study area is cultivated with maize, soya, and wheat, embracing the conservation agriculture (CA) farming system (Hobbs et al., 2008;Carretta et al., 2021). No-tillage operations were carried out in the field, and the crops are harvested using agricultural machinery. The area is divided by a network of ditches (Figure 1b), and at the time of the survey, it was free of crop coverage. According to the regional soil map (ARPAV, 2019), the zone is characterized by soils originating from the depressed region of the alluvial plain and mainly composed of silt and clay. Moreover, it is characterized by a high available water capacity (AWC, 225-300 mm) and moderately low permeability (0.36-3.6 mm/h). The condition of topographic depression, in addition to soil properties, makes the area easily susceptible to surface water stagnation. This fact is also observable from the past years'

| METHODS
The presented method for potential surface ponding mapping involved a first phase of in-field campaign for data collection. Subsequently, various tools were used for data processing, mapping, and result assessments. The workflow that describes the main conceptual steps is proposed in Figure 2.

| Data acquisition and analysis
The photogrammetric survey (Figure 2, A1) was performed after the harvest period, to ensure optimal conditions for applying the SfM technique, showing the agricultural surface as free as possible from vegetation. The survey was carried out using a DJI Zenmuse X4S camera (20M pixels, focal length 8.8 mm, 1-inch CMOS sensor) mounted on a professional quadcopter (DJI Matrice210v2). The drone was parametrized to provide 70% side-85% front picture overlap, F I G U R E 1 (a) Territorial framework of the study area (Rovigo, northern Italy). (b) The area surveyed by UAV-SfM and plots for water ponding analysis (P1, P2, P3, P4). The ditches/roads network is drawn on an orthophoto provided by Esri ® (2016). (c, d) Orthophotos provided by Google ® from 2015 and 2018, respectively. Arrows indicate problems affecting crops related to water ponding. (e) Water ponding from the field. Photo by Sofia Michieli. (f) UAV picture that shows wet zones in the study area. Arrows indicate some ponded areas. Photo by Sara Cucchiaro [Color figure can be viewed at wileyonlinelibrary.com] essential for the image-matching algorithms used in SfM (Eisenbeiss & Sauerbier, 2011;Eltner et al., 2016). Considering the flat morphology of the study area, free of complex features requiring specific analysis, the flight was performed in automatic mode. The UAV followed a route consisting of 16 strips and took nadiral photos (Figure 2, Ai).
The survey covered a wider area than the water ponding investigation zones only, including the region between northern and southern plots. coordinate system). This step was essential to georeference the SfM outputs and evaluate the photogrammetric errors (Grayson et al., 2018). Table 1 describes the main information about the UAV-SfM survey.

| Data processing
Before image processing, pre-calibration of the drone camera was carried out using Agisoft Lens ® software. This preliminary step allowed an automatic lens calibration routine using an LCD screen as a target.
In this way, it was possible to estimate the camera model parameters and lens distortion coefficients. Then, this information was refined in the SfM step, where the pictures dataset was processed using the photogrammetric software Agisoft PhotoScan Pro ® 1.4.5. This extracts 3D point clouds from the images and orthomosaics based on SfM and MVS algorithms (James et al., 2017;Javernick et al., 2014). In the first step, the software was able to generate and display a sparse point cloud ( (Han et al., 2017). This operation was performed through the Statistical Outliers Removal (SOR) filter. Afterwards, the point cloud was filtered manually, removing all points not belonging to the terrain (Jensen & Mathews, 2016). Where the UAV-SfM technique failed in detecting the ground in critical spots (in particular dense crop residues), some points belonging to low ground cover were left to avoid unrealistic information gaps in the final output.
F I G U R E 2 Workflow for potential water ponding mapping in agriculture using UAV-SfM. In detail, the diagram shows the process from in-field data acquisition to SfM-MSV steps, post-processing and output generation/validation. * Refers to the use of Agisoft PhotoScan Pro ® 1.4.5 software; ** refers to CloudCompare software; *** refers to Esri ArcGIS ® Desktop (version 10.6.1.9270) software [Color figure can be viewed at wileyonlinelibrary.com] These problematic areas were limited, however, and did not alter the final water ponding mapping. The points of the filtered dense cloud were finally interpolated and gridded within Esri ArcGIS ® Desktop (version 10.6.1.9270) software using the Natural Neighborhood Interpolation algorithm to create the DEM (Garnero & Godone, 2013;Sibson, 1981; Figure 2, Ci, C1). This method was selected because it can leave a rougher terrain morphology, avoiding the smoothing effects given by other techniques (Pirotti & Tarolli, 2010). Indeed, a rougher and more realistic surface could detect even the smallest concavities, thus helping in ponding detection.
Error assessment was performed for both sparse point cloud and DEM, generated after dense cloud filtering (Figure 2, C4). Indeed, the transformation of a cloud into a gridding elevation surface introduced several uncertainties or artefacts that influence the accuracy/quality of the produced surface. Therefore, in particular in the vertical component, these errors were evaluated through a statistical comparison between the Z values of CPs (measured in the field by GNSS) and the equivalent Z measures of DEMs. First, following Höhle and Höhle's (2009) approach, the outliers were removed by applying a threshold selected from an initial calculation of the error measures. Then, RMSE (Estornell et al., 2011), standard deviation (SDE), mean error (ME), and the normalized median absolute deviation (NMAD), a robust estimator for the SDE more resilient to outliers in the dataset (Cucchiaro et al., 2020;Gonçalves et al., 2018;Höhle & Höhle, 2009) were calculated.

| Statistical validations
The i and j elements are the number of rows (based on the comparison raster) and columns (based on the reference raster) that correspond to the number of classes used for the analysis. In contrast, elements n are the result of pixel comparison between input data. The matrix was subsequently normalized as described in Equation 2, so that each of its elements n will be expressed as a function of the total number of pixels p (Pontius & Millones, 2011): where p represents the normalized value of the n ij element in the original matrix, J is the number of classes, and N i is the total number of pixels for each row. Based on the normalized version of the confusion matrix X, Cohen's kappa standard accuracy k(X) is calculated as: Here, C is the proportion agreement, calculated as the normalized sum of true positive pixels, and indicator of concordance: According to Pontius and Millones (2011), E is instead the overall expected agreement. It is calculated assuming a random spatial allocation of the classes in the comparison raster, given the proportions of those classes in the reference and comparison raster. The expected agreement (e g ) for an arbitrary class g is proposed in Equation 5, while E in Equation 6 refers to their sum: 3.4.2 | REA and in-field measurement: water depth assessment During the survey, the depth of the stagnant water on the agricultural surface was measured at 47 sample points using a stick meter and their positions are shown in Figure 4.
T A B L E 2 Confusion matrix (X) presented by Pontius and Millones (2011) and used to evaluate the accuracy with which the proposed method can map water ponding Comparison refers to the binary raster based on the REA. b Reference refers to the rasterization of the polygons identifying the presence of pond water in the field derived by the orthomosaic.
These effective depth measurements were statistically compared with what was estimated using the morphological approach investigated. Indeed, the degree of correlation between the two quantitative variables was evaluated, which was necessary to validate the methodology's reproducibility (Lin, 1989). The Pearson coefficient r xy , also known in statistics as bivariate correlation, was used. This indicates a linear correlation between two variables, its direction, and how statistically robust the agreement is (Adler & Parmryd, 2010).
Equation 7 was used to verify the correlation between real and estimated water depth measurements. In general, positive values describe a linear relationship, with a maximum value of 1 that defines a perfect match: where cov(x,y) is the sample covariance between two variables x (water depth measured in the field) and y (water depth estimated by  As can be observed from the evaluation of the errors, the SfM technique has proved to be valid and robust in the digital reconstruction of the elevation. Furthermore, during the interpolation and gridding process of the point cloud, different interpolation techniques could be tested to further mitigate errors, but this analysis is too specific for this paper's scope. Indeed, the choice of the appropriate interpolation algorithm is a debated issue in the literature, as well as the optimal neighbourhood size setting (Milenkovi c et al., 2015). However, the DEM generated following the proposed procedure was suitable for the aim of this research, offering high-resolution and quality data. infiltration capacity of the soil (Horn et al., 1995). Stagnations of a few metres wide can be observed (in some cases less than 1 m), which extend along the direction of the maximum plot length. Finally, although this is not the main purpose of this work, it is also important to mention the role of soil type and its propensity to be waterlogged.

| REA application
As described earlier, the soil in the study area mainly consists of clay and silt, with moderate permeability and formed in depressed alluvial zones.
In this context, appropriate solutions may concern the use of drainage systems. Several studies in the literature illustrate how they may be used to combat waterlogging in agriculture. For example, Singh (2017) provides an overview of different drainage technologies; Ritzema et al. (2008) test several subsurface drains to counteract stagnation in irrigated land in India, according to soil type and annual rainfall; other applications are described for case studies in Australia (Cox & McFarlane, 1995), Pakistan (Qureshi et al., 2008), and Turkey (Bahçeci & Nacar, 2009). Also, another solution consists of modifying the profile of the agricultural surface through appropriate soil movements. In this way, by working on the topography, it is possible to avoid depressed areas and create a slope gradient that drains the water runoff towards the ditches. In this regard, other research has investigated the potential of a correct combination between field topography and ditches (Appels, 2013;Dunn & Mackay, 1996;Shore et al., 2013), leading to efficient water management and the solution of waterlogging.
F I G U R E 7 Potential pond water depth estimated on a 0.10 m DEM through the REA indicator. The raster is classified into water-depth classes to which a colour is assigned (transparent for 'No ponding').
(a) Whole study area; (b) focus on plot P1; (c) focus on plots P2, P3, P4. Particularly interesting are the blue stripes that are evident in the middle of these plots, created due to heavy mechanization [Color figure can be viewed at wileyonlinelibrary. com] T A B L E 5 Summary table of the statistical assessment described in the text. According to Pontius and Millones (2011): C (dimensionless), used in the calculation of k(x), is the proportion agreement; E (dimensionless), used in the calculation of k(x), is the overall expected agreement. According to Cohen (1960): k(x) (dimensionless) is the kappa standard and measures the accuracy with which the proposed approach can predict the extension of the area potentially occupied by pond water. r xy , Pearson's coefficient (dimensionless), measures the correlation between the water depth measurements made in the field at sample points and the values predicted by the proposed approach

| REA assessment
Regarding the assessment of the results, the two statistical tests described above are performed. Looking at the first analysis, it is possible to observe in Table 5 and Figure 8 how the proposed method can map the extent of the area potentially subject to water ponding with a reasonable estimation. It is interesting to note the high value of k(x), an indicator of concordance, which shows how the REA can classify a pixel as ponding in an acceptable way. The second statistical assessment shows an important correlation between the two datasets (Table 5), also confirmed by scatterplots in Figure 9. Looking at these, a positive correlation can be observed between the values measured in the field and those estimated by the method, with a variance tending to remain within the confidence interval displayed. This is mainly the case for P1 and P2, where higher values of r xy were found than for P3. In P4, a more noticeable discrepancy may be observed.
Looking at Figure 9d, it is possible to note an underestimation of the stagnant water depth for lower values and an overestimation for higher values. Specifically, there is one point where the estimated depth of standing water is about 6 cm greater than that measured.
The proposed method is affected by limitations that may influence the water depth estimation, especially if some standing not-clear water remains on the surface. Being a purely geomorphological approach, soil and surface water characteristics and weather conditions are not considered input. For this reason, it is assumed that the same starting conditions exist throughout the study area. Indeed, this work aims to offer stakeholders a quick and cost-effective standing water mapping tool; implicit approximations are therefore inevitable but acceptable, considering that a complex problem is described with an expeditious methodology. Also, limitations can be found in the instruments and methodologies used to perform the survey and to process the data. The first limit regards the point cloud's errors,  Other research in the literature deals with mapping areas prone to water ponding using remote sensing techniques. For example, Merot et al. (1995) tested a topographic approach. The authors mapped waterlogging in agriculture using the saturation overland flow concept proposed by Beven and Kirkby (1979), and implemented in the TOPSOIL model. Different from this paper's method, it increases the complexity of the methodology and data processing time considerably. Then, they used a 40 m resolution DEM derived from topographic 1:25 000 maps, suitable for a catchment scale, but a limit for field-scale analysis. Also, Mandal andSharma (2001) andPandey et al. (2012) proposed a different remote sensing approach, namely the use of multispectral satellite imagery, tested for a large-scale lowland irrigated landscape. Their analysis permitted the mapping of potentially susceptible areas to water ponding at regional and subregional scales. A similar methodology was also proposed by Liu et al. (2018), which analysed areas prone to waterlogging using commercial satellites' multispectral products. Although this methodology is widely used in the literature, some limitations may be found for water ponding mapping. Indeed, they mainly described the spatial extent of the standing water at large scale. Instead, by adopting a geomorphological method such as the one proposed in this study, it is also possible to map the depth of surface water at farm scale, thus offering the stakeholders important information for better management. Furthermore, the spatial resolution of these data is another interesting issue. Although multispectral imageries with pixels smaller than 1 m are available on the market, the detail is still insufficient to map ponding phenomena at microtopographic level. To overcome this problem, it would be useful to use a UAV equipped with a F I G U R E 9 Scatterplots generated comparing the water depth measured in the field with that estimated by the REA: (a) P1; (b) P2; (c) P3; (d) P4. Graphs include the linear correlation line (95% confidence interval). Points are themed using the chromatic scale in Figures 4, 6 and 7 and based on the measured value. Elaboration was performed using R-4.0.1 for Windows (R Core Team, 2000) through the ggplot library (Wickham, 2016) [Color figure can be viewed at wileyonlinelibrary.com] multispectral and thermal camera. The method proposed in this work could be improved in the future by adopting a complete UAV SfM + Multispectral + Thermal solution. In this way, the REA from SfM may be used to map water ponding and the multispectral and thermal camera to assess the results, especially regarding the standing water's spatial extent. For example, the relationship between surface temperature and soil moisture (Turner et al., 2011) could be exploited to identify areas potentially subject to ponding.
In general, the assessment of this study's results suggests that the proposed method could map the potential extension and depth of stagnant water in agriculture with reasonable estimation. The role of geomorphology as a key element in water ponding generation and understanding in agricultural landscapes is particularly interesting. For this reason, the research tries to exploit the developments of modern topographic techniques for the identification of critical areas on a large scale. Thanks to the UAV-SfM surveys, the mapping of stagnant water becomes rapid in data acquisition and processing, offering crucial information to stakeholders at low cost. The method, which gains robustness thanks to statistical validation, becomes an attractive tool for preparing agricultural land, identifying areas in subsidence that need reprofiling or ad hoc tillage. Therefore, our research has great applicability in lowland farming and becomes a decision-making utility in designing targeted solutions.

| CONCLUSION
Stagnant water in lowland agriculture landscapes represents a threat to crops and cultivation management. In this work, a novel approach is proposed to map the ponding problem using a UAV-SfM technique.
This primarily allows us to perform a low-cost analysis compared to other topographic surveying techniques, such as LiDAR. Also, thanks to the rapid data acquisition and processing, the method is also robust at farm scale, proposing an upscale from the experimental plots. Specifically, the morphometric indicator relative elevation attribute was used to map the potential water ponding phenomena in such landscapes. The high resolution obtained becomes a key starting point.
Indeed, UAV-SfM proved to be very useful for investigating the geomorphic signatures that characterize the agricultural surface. The proposed technique was able to reconstruct in detail even the smallest convexity and concavity of the terrain, offering an interesting insight into the potential occurrence of stagnant water. Furthermore, in this work, an accurate statistical comparison between the results obtained and an in-field campaign of water depth measurement is proposed.
The research outcomes illustrate how a purely morphological approach can be used to identify the potential surface water stagnation phenomena, through expeditious and low-cost topographic surveys using UAV systems.
The results provide useful information for better managing the analysed site, highlighting criticalities related to some non-optimal farming practices. Besides, it could be useful in water stagnation risk prevention and may be used as support for sustainable agricultural management.