Remote Sensing-Assisted Physical Modelling of Complex Spatio-Temporal Nitrate Leaching Patterns from Silvopastoral Systems ^†

Highlights

What are the main findings?

• Simple computer vision retrieval of LAI supported the point-scale model to map complex spatio-temporal variation in soil nitrate leaching.
• Nitrate leaching in the studied silvopastoral systems was high and varied spatio-temporally from 7 to 860 kg N ha⁻¹ year⁻¹.

What is the implication of the main finding?

• The potential of remote sensing retrieval of crop and soil data is large and underutilized to support models for estimating nitrate leaching.
• Future studies should investigate “deep learning meets process-based model” as a hybrid approach to capture complex nonlinear patterns of soil nitrate leaching.

Abstract

Affordable optical data from Unmanned Aerial Vehicles (UAVs) coupled with process-based models could constitute an integrative platform to map complex spatio-temporal patterns of nitrate leaching and reduce uncertainties in tightening the nitrogen (N) cycle of silvopastoral systems. This study uses field data from a commercial farm in Denmark with lactating sows housed in paddocks with pastures flanking a central zone of poplars, either pruned (P) or unpruned (tall, T), each with resources (feed and hut) on the same (S) or opposite side (O) of the tree zone. The poplar leaf area index derived from canopy cover using a computer vision approach on true-colour UAV imagery was fed to a process-based model alongside soil data and geostatistical analyses to derive the soil water balance across the paddocks and explicitly map the variation in soil nitrate leaching. The results showed clear patterns not seen before of nitrate leaching hotspots shifting from high values in the pre-study year without animals to diluted lower values in the main study year involving the pigs. The results also showed a seasonal and spatial variation of 7 to 860 kg N ha⁻¹ year⁻¹, a wide leaching range otherwise difficult to capture, by employing only a process-based model using mean effective parameters. Nitrate leaching was in the order PO > PS > TO > TS. The N cycle was tightened with T regardless of S/O. The approach could be improved with more machine learning-aided process-based modelling to operationally monitor complex silvopastoral systems to alleviate nitrate leaching in outdoor pig systems.

1. Introduction

Integration of remote sensing data and process-based models in agricultural and environmental research and innovation aims to improve the agronomic, ecological and socio-economic sustainability of agroecosystems. However, the integration is not a trivial task due to both data assimilation and scale issues. Process-based models can simulate many processes in the system based on physical principles, typically in one dimension, and provide a balance between system complexity, model parameterization, and project budget. Remote sensing support is either in forward mode, to dynamically update the model’s initial conditions and constrain its simulations with observed reality, or in inverse mode to improve model parameterization. Nonetheless, in the context of agro-environmental sciences, the main advantage of the integration is the more accurate, near-real-time description of the actual soil–plant condition along various stages of the growing season [1,2], which affects other processes such as soil water and nitrogen (N) uptake, evapotranspiration, and nitrate leaching out of the root zone. Therefore, to ease the coupling with remote sensing, it is desirable to suggest an integrative framework with simple yet reliable remote sensing data types and image analysis methods that offer both accuracy and spatialization.

Remote sensing data to support process-based models at no cost and at a large scale such as the satellite multispectral missions of MODIS, Landsat and Sentinel [3,4,5,6] are unable to contribute directly and accurately to following the water and N dynamics of (agro)ecosystems mostly due to their coarse spatial resolution. For analyses at high spatio-temporal scales, UAV data are more useful. Zhang et al. [7] compared Red–Green–Blue (RGB; DJI Mavic2 Pro from DJI, SZX, Shenzhen, China) and multispectral (DJI P4, DJI, SZX, Shenzhen, China) UAV data from rice fields in China to estimate leaf area index (LAI, the perpendicular projected area of a one-sided leaf in m² to that of ground surface in m²) and showed the high potential of both data for practical research. Ferraz et al. [8] also showed the high predictive capacity of RGB UAV images for retrieving bioparameters of sesame in Brazil. Both studies used machine learning on image data for algorithm training with field data. However, near-machine learning approaches based on computer vision principles such as image thresholding, segmentation and edge detection/feature matching might also be suitable despite their simplicity, especially when field data and computational capacities are limited and frameworks already involve complex process-based models. Image thresholding involves near-fuzzy classification of each pixel based on different intensities, starting with luminance and brightness (hue, H; saturation, S; and value, V), followed by a weighted average of the RGB bands and eventually supervised thresholding [9,10], which is particularly suitable for homogeneous canopies typically seen in managed agricultural systems. Image thresholding has been reported as promising for calculating LAI (or also plant area index, i.e., total leaf and cordon area in m² to that of ground surface in m², as well as canopy cover, CC, i.e., fraction in % of vertical projection area covered by canopy) from not only UAV but also smartphone app RGB data to detect changes in canopy structure after the application of canopy management practices [10]. However, strict thresholds can also filter part of the leaf or canopy pixels and underestimate the projected LAI or CC; hence, the approach should ideally involve threshold analysis.

Chen et al. [11] estimated LAI from UAV (multispectral) images of wheat fields in Australia to suggest an unsupervised and plot-scale plant phenotyping framework involving a process-based (radiative transfer) model. This forward support to process-based models has been widely used and has shown improved prediction of aboveground traits such as crop yield by integrating RGB- and multispectral UAV-based estimates of not only LAI but also leaf N concentration or plant height, with process-based models, e.g., SWAP [12], AquaCrop [13], DSSAT [14], WOFOST [15], SAFY [16], and RiceGrow [17]. Scientific assessments of belowground processes such as nitrate leaching from soils with remote sensing-assisted process-based models are considerably sparser, as the processes are belowground and involve lag phases and also the models require parameterization beyond crops to realistically simulate the evapotranspiration and soil water transport needed to estimate nitrate leaching. An earlier study suggested the integration of LAI estimated by remote sensing data and soil physical and hydraulic variables estimated by geostatistics and Daisy, a process-based model [18], as a plausible approach to not only reveal complex and dynamic patterns of water transport phenomena and nitrate leaching, but also handle uncertainties in space and time for further research and decision making [19]. Manevski et al. [20] used Daisy to simulate the spatial distribution of nitrate leaching in silvopastoral systems with poultry and willows using largely approximated LAI distributions and not considering soil spatial heterogeneity. Nitrate leaching is known to have a leptokurtic distribution (thicker tails) with extreme data values (outliers or high variability) compared to a normal distribution due to the high spatio-temporal variability of both soil nitrate and water contents [20,21]. Schuster et al. [22] integrated digital methods involving tractor-mounted multispectral sensors, and satellite data from Sentinel-2, with the PROMET model and georeferenced measurements of subsoil nitrate stocks to reveal explicit spatial variability for large croplands in Germany. These patterns might easily be complicated by trees that utilize soil nitrate and have shown potential to reduce nitrate leaching in arable cropping [23,24], and outdoor pig and poultry farms [20,25]. Especially before the leaching season peaks, rapidly growing and pruned trees might absorb nitrate from the deeper soil layers, which is not the case for grass [26]. From an agro-engineering perspective, positioning resources—hut, feed trough—influences the animal behaviour where they distribute manure, with a tendency for a more even distribution in tree areas if available [26]. To the best of the authors’ knowledge, no studies have been conducted to investigate the role of affordable and simple remote sensing in quantifying the explicit spatio-temporal distribution of nitrate leaching in complex silvopastoral systems.

The main aim of this study was to firstly integrate a relatively simple remote sensing approach and a complex process-based model to simulate the water balance of pasture paddocks on sandy soils with outdoor pigs distributed in different resource positions and tree treatments and secondly, quantify the spatio-temporal distribution of nitrate leaching in the paddocks. Silvopastoral systems integrate trees, forage and livestock on the same land for mutual benefit [27]. Their investigation and optimization are challenged by the numerous interlinked, concurrent and cause-and-effect processes and feedbacks [3,24,27]. In outdoor pig production on pastures, a rather low nitrogen (N) digestion utilization efficiency (NUE) of 10–44% [28] against high N input in feed creates a large surplus of soil nitrate from manure deposited unevenly and non-randomly [29,30]. This creates “hotspots” since the pigs’ rooting and trampling behaviour destroys the pasture [30] which would have otherwise utilized the soil nitrate. As a result, high soil nitrate leaching has been reported, easily exceeding 200 kg N ha⁻¹ in a year [31], and these values are likely underestimated given the inability of the methods in previous studies to integrate spatio-temporal variation. The specific objectives were to (1) delineate LAI dynamics based on UAV true-colour imagery, (2) integrate LAI dynamics to parameterize the Daisy model and simulate the paddocks’ water balance, (3) estimate nitrate leaching by the percolation-weighted concentration method with the measured nitrate values, (4) statistically analyze the spatial and temporal variability of nitrate leaching data within paddocks and between treatments, and (5) calculate paddock N mass balances consisting of inputs and outputs.

2. Materials and Methods

2.1. Experiment Setup and Farm Data

A field experiment was conducted on a commercial farm with sandy soil in Denmark on four paddocks, each measuring 63 × 15 m and each hosting one lactating sow and its piglets at a time. The middle zone of each paddock—a third of the total area—contained the poplar tree hybrid O.P.42, Populus maximowiczii × P. trichocarpa, planted in 2011 at 2–3 m spacing (1900 trees ha⁻¹; Figure 1). In April 2022, the trees were either cut to 2 m height, with the resulting biomass chipped and the wood chips placed on the soil surface beneath the trees (hereafter denoted as pruned, P) or left uncut (i.e., tall, T), with each treatment occupying two paddocks. The rest of the paddock area was planted with spring barley with undersown grass–clover (year 1) followed by three batches of sows in 2023–2024 (year 2) (see Figure A1 in Appendix A). The historical management on this farm followed a crop rotation with sows every second year, alternating with spring barley undersown with grass–clover. In the year before measurements, four batches were managed: in the first two, huts and feed were placed on opposite sides of the trees; in the last two, both were on the same side.

Figure 1. Layout of paddocks with different vegetation zones and containing feed (red circles) and water (red squares), and huts (grey boxes with indicated batch number) positioned either on the same (S) or opposite sides (O) with trees in between. The trees were either not pruned (tall, T) or pruned (P). Each paddock was equipped with 51 suction cells (black dots) pooled by three (encirclements) for soil water sampling. Collection boxes are shown as black rectangles. The four examined paddocks were part of a larger experiment on 16 paddocks (right plot) as reported in Ullfors et al. [32]. Layout not to scale.

For the experiment, grass–clover grew during the winter 2022–2023 before sows of the crossbreed Nordic Landrace × Yorkshire (TN70 by Topigs Norsvin) were introduced to the paddocks in three batches, each starting mid-April, late July and early November 2023, and stayed in the paddocks for eleven weeks, one week prior to farrowing and ten weeks during lactation (Figure 2). Within this period, resource position treatment was introduced: feed and hut positioned either on same (S) or opposite side (O). Farm data needed for N balance calculation included eliminated and weaned piglets (number, weight), feed amount (by feeding curve for sows), feed ingredients and straw usage. Piglets’ feed consumption was not determined at paddock level since they moved between paddocks. Yields of barley grain and straw were estimated from previous farm records.

Figure 2. Illustration of the time overlap of different data collection and animal husbandry batches.

2.2. Remote Sensing Data Collection and Analysis

Remote sensing field campaigns with UAVs were conducted on four occasions in 2023 and six occasions in 2024. The UAV was a DJI Mini 3 Pro (DJI, SZX, China) with a 48-megapixel camera flown at ca. 43 m height over the entire field containing the paddocks to collect true-colour (RGB) images. The system used a 1/1.3 in CSMOS (48 MP) image sensor with a field of view of 82.1° (24 mm format equivalent), aperture of f/1.7, shooting range of 1 m and above, and electronic shutter speed of 1/8000–2 s.

The RGB images after each flight were downloaded on a local PC. For each date and pruning regime, one image was chosen and cropped to show the tree area only. To retain the area covered by green leaves only, a semi-machine learning analysis based on computer vision was conducted by converting the RGB values of the photos to HSV and applying thresholds. HSV is a widely applied tool that splits the colour information into H (type of colour), S (colour intensity), and V (brightness), yielding binary image classification less sensitive to illumination changes, shadows, and highlights, which are common problems in RGB. Moreover, RGB mixes colour and brightness and different lighting can drastically change pixel values, but HSV isolates colour (hue) and the classification remains accurate even under variable brightness. Therefore, the method is well suited for discriminating objects with similar lighting intensity but different colours (vegetation and soil, as in this study, but also water), and it is widely used in image thresholding and segmentation, reducing classification complexity [33].

At this point, it is important to emphasize that the DJI Mini 3 Pro RGB camera has been used for various scientific tasks such as assessing forest ecology [34,35,36,37] and crop protection [38]. However, it does not provide in-depth geometric and radiometric calibration as many multispectral camera systems do; therefore, its use should be carefully tailored for the study objective. There are factory-calibrated and auto-calibrated parameters ensuring high image quality, in addition to several user specifications to ensure reliable image acquisition [39], which were deemed ample for the image thresholding method based on HSV. In fact, H and S channels are relatively insensitive to image radiometry; however, the V channel involves brightness, which affects intensities under different lighting conditions and can potentially reduce the extent of the object of interest or capture unnecessary background. Here, such risk was deemed minimal over the highly homogeneous canopies of poplars in the two treatments.

The thresholds were found by comparing the colour values of the green leaf area against branches and surrounding soil, which were chosen manually with numerous attempts using GIMP-GNU Image Manipulation Software (v2.10.38) [40]. The initially chosen thresholds were applied to three photos representing different seasons and were varied to visually assess the fit. Suarez et al. [33] provides a Python-based tool for high-throughput phenotyping of CC also based on HSV for citrus tree canopies as an example. The discriminated green leaf pixels were used to obtain CC (Figure 3). Since images were available from April 2023 to April 2024 only, CC was assumed to be the same in 2022 with a 20% reduction for pruned trees to mimic pruning in April 2022. Also, the spatial variation in CC was remarkably low (see, e.g., Figure 3) and notable variation was detected only between the two main treatments P and T. Therefore, in order to avoid redundant computation, only zones per treatment were depicted for further analysis. The CC was used to determine LAI using Beer’s law, $LAI=-\left(log\right(1-CC)/k)$, with a canopy extinction coefficient (k) of 0.5 [41,42]. Also, once a month from April 2023, the herbaceous vegetation cover for the subzones (Figure 2) was determined qualitatively by experienced technical staff using a standardized scale [25,43].

Figure 3. Example of threshold determination using GIMP-GNU Image Manipulation Software using the colour values of green leaf area and other image features such as branches and surrounding soil to estimate leaf area index of the poplars in the experimental paddocks. The increase in green area in the centre on 10 October 2023 is a patch of not-well-harvested canopy that remained greener for a prolonged time before the late autumn defoliation. Full vertical lines depict mean and median of each histogram plot, with V+S exaggerating skewness as values can exceed 1.0 and approach 2.0 which is normal for canopy images with shadows and sunlit surfaces.

2.3. Soil Water Sampling and Water Balance Modelling

In each paddock, 51 ceramic suction cups were installed in April 2022 at 1 m depth (Figure 1). The aim was to cover the spatial variation in soil nitrate destined for leaching as much as feasible. The sampling interval was two weeks (Figure 2). Soil water samples were collected from November 2022 before pig introduction in spring 2023 by applying a negative pressure of about 70 kPa usually 1–3 days before sampling. For economic reasons, the three samples of each row had to be pooled before the nitrate concentrations were determined colorimetrically on an auto-analyzer. If a cup was malfunctioning, water collected from the remaining cups was used. This left 17 data points per paddock (i.e., 68 data points for the entire study field with four paddocks) on soil nitrate.

The water balance at each location with measured soil nitrate in each paddock was modelled with the Daisy model, version 5.92 [18]. It is a one-dimensional, deterministic and detailed process-based model describing water/energy flows in the soil–plant–atmosphere continuum. In brief, the water fluxes considered are precipitation (gain), irrigation, if any (gain), evapotranspiration (loss), surface run-off, if any (loss), and surface fluxes, as well as deep percolation (i.e., drainage; loss) and capillary rise (gain). Water input used for wallowing, a behaviour critical for thermoregulation, sunburn, skin care, and overall comfort of pigs, was not included as it was limited in space and not quantified. The dynamics of soil water, and thus drainage, are modelled by a numerical solution of the Richards’ equation. Reference evapotranspiration is calculated according to the FAO Penman–Monteith equation [44]. Further details can be found in Hansen et al. [18].

The model was run from 1 April 2022 to 1 April 2024 with the required inputs of daily weather data and information on soil, plant and field management. Weather data included daily precipitation (mm), air temperature (°C), relative humidity (%), wind speed (m s⁻¹), and global radiation (MJ m⁻²) obtained from the national grid database of the Danish Meteorological Institute. Precipitation was corrected for uncertainties caused by evaporation, wind and wetted surfaces [32]. Vapour pressure was calculated from relative humidity and air temperature according to FAO standards [45].

For the soil parameterization, soil texture and organic matter (OM) content were derived from topsoil (0–20 cm) samples, and for the subsoil (20–100 cm; Figure A2 in Appendix A), values for 30–100 cm soil were obtained from a 3D map of the soil texture of Denmark at 10 m resolution [46]. Soil texture and OM content over the whole experimental area were markedly similar, depicting a high degree of homogeneity. Clay, silt, sand and OM contents were, respectively, 4.7, 6.2, 85.2 and 3.9%, the C/N ratio was 13.7, the bulk density was 1.41 g cm⁻³ for the topsoil, and corresponding values for the subsoil were 6.4, 3.7, 89.4, 0.5, 13 and 1.43. The soil hydraulics were described by Brookes and Corey’s inverse model with topsoil values for saturated hydraulic conductivity and soil water of 10.1 cm h⁻¹ and 35%, and soil moisture at field capacity and wilting point of 18 and 9%, respectively, whereas values for the subsoil were 10.6 cm h⁻¹, 37.1%, 21% and 3.7%. The soil surface evaporation factor in Daisy (parameter EpFactor) was reduced by 35% in the paddocks with pruned trees compared to the tall trees to account for mulching with wood chips, conserving soil water content and reducing soil evaporation [47,48].

For plant information, the “permanent” vegetation module described by Boegh et al. [49] was used to simulate the poplar zone with plant height set to 5.3 and 18.3 m for pruned and tall trees, respectively. The LAI of the poplars determined from the UAV imagery for pruned and tall trees was given as an input to Daisy for the prediction of interception and evapotranspiration (crop canopy transpiration factor in Daisy, parameter EpFac of 1.2). For the grass zone, the management of the crop was specified with historical data on sowing, ploughing, seed bed preparation and harvest of the previously grown spring barley; LAI of the grass–clover grazed by the sows was estimated to have a start value of 2 in April based on the literature [50,51] multiplied by the observed fraction of vegetation cover. Simulations were set for each pruning treatment in the tree zone and each vegetation cover observation subzone in the grass zone. This degree of detail was chosen to account for the high variation in vegetation cover.

2.4. Nitrate Leaching Calculation

The measured soil nitrate concentrations at each pooled location and the respective modelled daily percolation were combined to determine daily leaching values according to an improved trapezoidal rule method by Lord and Shepherd [52]; i.e., instead of assuming measured concentrations to represent the average flux concentrations, the concentrations were interpolated and weighted by the simulated daily percolation [20,53], as illustrated in Figure 4. The obtained daily nitrate leaching was accumulated to yearly sums from 1 April to March 31 for 2022–2023 (year 1) and 2023–2024 (year 2).

Figure 4. Illustration for estimating nitrate leaching according to the trapezoidal rule method: instead of assuming measured concentrations to represent average flux concentrations (A) [52], concentrations were interpolated and weighted by Daisy-simulated daily percolation (B) [53]. The figure illustrates underestimation of nitrate leaching by the original method and improvement by incorporating the daily variation in percolation.

2.5. Statistical Analysis and Nitrogen Balances

The annual nitrate leaching data based on 68 observations across all four paddocks were spatially interpolated for each year by ordinary kriging (prediction and variogram) using the krige function from the gstat R package (RStudio version 4.4.2) on a 1 × 1 m grid. The results were used to show the spatial variation in nitrate leaching and calculate average leaching for each zone in each paddock and year, and for each paddock. Effects of treatment and zone on annual nitrate leaching were analyzed with a generalized linear mixed model. Effects of zones and treatments on weekly leaching were analyzed with a generalized additive model (GAM) with smooth functions to link predictor variables to the response variable. Here, different group levels were fitted with varying functions.

For each paddock, surface mass N balance for years 1 and 2 was calculated as the difference between inputs and outputs [20,25,54]. Inputs were feed for sows and piglets, sow weight loss, straw for bedding, seeds for barley, and grass–clover fixation by clover and atmospheric deposition [25]. Outputs were piglets, harvested barley grains and straw. An overview of the data and the factors is shown in Table A1 in Appendix A. Corresponding soil N balance was calculated in greater detail: by subtracting from the surface balance the sum of N outflows from the 1 m soil, i.e., leaching, ammonia volatilization and denitrification, the flat soil surface ensured no N losses via surface runoff.

3. Results

3.1. Weather and Canopy Characteristics

The 2 study years were similar to the 10-year average values, with slightly lower precipitation in year 1 (Table A2 in Appendix A). The water balance of the study period was deemed to represent the climatic normal. Poplar LAI values estimated from the UAV imagery—and given as input to the Daisy model—varied seasonally and this was depicted well by the data (Figure 5), with the highest values of 9.2 and 7.7 in tall and pruned trees, respectively. In the grass zone, LAI decreased in autumn 2023 whereafter it remained low. Grass–clover CC visual assessments for year 2 (April 2023 to April 2024) supported the grass zone LAI by showing high CC before and during paddocks’ animal occupation for the summer months before decreasing sharply in late autumn (Figure A3 in Appendix A).

Figure 5. Development of leaf area index (LAI, m² m⁻²) derived from UAV imagery (tree zone, B; April 2023 to November 2024) and visual assessment (grass zones, A and C) in the paddocks. Treatments include resources (feed and hut) positioned either on same (S) or opposite sides (O) with trees in between, and trees either not pruned (tall, T) or pruned (P). Shaded areas show the standard error of different subzones used for visual assessment. Tree zone B values in 2021–2022 were adopted from 2022–2023 for T, also for P but with 20% reduction to mimic effect of pruning in April 2022.

3.2. Simulated Water Balance for the Experimental Paddocks

The hydrological years 2022/23 (year 1) and 2023/24 (year 2) were comparable in air temperature and precipitation, slightly wetter in year 2 compared to year 1 (Figure 6). As a result, Daisy showed more delayed percolation in the winter of year 1, before the pigs were introduced in the paddocks, compared to year 2 when percolation was more gradual and overall higher. Percolation was lower in the tree zone, 643 and 751 mm in years 1 and 2, respectively, compared to 701 and 859 mm in the grass zone. Simulated actual evapotranspiration was on average 22% higher for poplar than for grass. The change in water storage was 234 mm on average across years and vegetation zones. These water balance components provide information on the transport factor for nitrate leaching.

Figure 6. Simulated water balance components in grass–clover and poplar vegetation. Values are mean of treatments, with notable variation (shaded area, standard deviation) only for poplar in year before experiment as trees were pruned and the difference faded in year during experiment (see Figure 1).

3.3. Spatio-Temporal Distribution of Soil Nitrate

The measured soil nitrate concentrations varied markedly from 0.1 to 158 mg L⁻¹ (Figure 7). Their lower quartile, median and upper quartile were 7.2, 14.01 and 28.8 mg L⁻¹, respectively, with few occasions above 100 mg L⁻¹. Shown spatio-temporally, their dynamics clearly showed peaks, i.e., hotspots in autumn and winter, and this was more evident in the grass compared to the tree zone, except for the S treatment where high soil nitrate concentrations were measured in the tree zone in autumn–winter of 2024.

Figure 7. Spatio-temporal distribution of soil nitrate measured at 1 m depth in experimental paddocks with sows and piglets and their resources (feed and hut) positioned either on the same (S) or opposite side (O) with trees in between, and trees were either tall (not pruned, T) or pruned (P). Bright line rectangles depict soil mineral nitrogen sampling [43]).

3.4. Spatio-Temporal Maps of Nitrate Leaching

Spatio-temporal delineation (kriging) of nitrate leaching depicted the observed soil nitrate, which overall ranged similarly between the two years, i.e., 9 to 860 kg N ha⁻¹ in year 1, and 7 to 779 kg N ha⁻¹ in year 2 (Figure 8). Also, the range of the semivariogram, i.e., the maximum distance over which nitrate leaching is spatially autocorrelated, was about 20 m. On average from these results, nitrate leaching was higher in year 2 (206–242 kg N ha⁻¹) during paddocks’ animal occupation than in year 1 (144–194 kg N ha⁻¹) representing crop husbandry. Paddock averages in year 2 were highest for the PO (242 kg N ha⁻¹). This was contrary to expectations for the O treatment, which depicted, on average, lower nitrate leaching compared to the S treatment. Also, according to the GLM, the effect of zones depended on the two treatment factors and the observation year as well (Table A3 in Appendix A). Model-estimated accumulated leaching averages for the paddocks in year 2 (the year with pig occupation) were higher with pruned trees (234 kg N ha⁻¹) than tall trees (177 kg N ha⁻¹) and slightly higher with the hut on the opposite side (200 kg N ha⁻¹) than on the same side (192 kg N ha⁻¹). For year 1 (the year with crop husbandry), the values were 15–20% lower compared to year 2.

Figure 8. Spatial distribution of annual (accumulated) nitrate leaching from 1 m soil depth across paddocks with trees either tall (not pruned, T) or pruned (P) and the hut (boxes with indication of batch no.) positioned either on the same (S) or opposite side (O) of feed with trees in between.

The statistical comparison for each zone in each paddock revealed a significant increase in average annual nitrate leaching in the tree zone of pruned trees (P) compared to tall trees (T) in both years (Figure 9). This could indicate that pruning reduced the soil nitrate uptake by the poplar trees, but it could also be caused by differences in pig behaviour not studied here. The only significant difference between resource position treatments appeared in the year with crop husbandry (year 1) with higher leaching in O compared to S in grass zone A. This result is not readily explained as all four paddocks had undergone identical management in the years leading up to pig occupation in year 2. In the year with pig occupation (year 2), in grass zone A, the average annual leaching tended (P = 0.087) to be higher when the hut was placed on the same side as feed (S) instead of the opposite side (O). In contrary, in grass zone C, the average annual leaching tended (P = 0.054) to be higher in O compared to S. When analyzing leaching at higher temporal resolution with the GAM, weekly nitrate leaching data revealed comparable dynamics between the treatments in year 1 when the magnitude was larger in O compared to S, in the grass zone with the feed and water (Figure A4 in Appendix A). The effect of tree management was less evident on the weekly soil nitrate leaching, but values in the tree zone were clearly larger for P compared to T in the leaching season 2023–2024.

Figure 9. Effect of vegetation zones, pruning (not pruned, T, or pruned, P) and resource (hut, feed) position (on the same, S, or opposite sides (O) with trees in between) on annual nitrate leaching from 1 m soil depth. Boxplot midlines show mean, hinges show first and third quartiles, and whiskers extend to the most extreme values but not exceeding 1.5 times the inter-quartile range. Points on the right side of the boxplots show least-square means with error bars being 95% confidence intervals. Points on the left side are the raw data. p-values for contrasting zones (line colour showing treatment level) or treatments (black lines) are given on top.

3.5. Nitrogen Mass Balance for the Experimental Paddocks

The surface N balances are shown in Table 1, which refer to the entire area occupied by the four paddocks since the input mass flows of N did not differ between paddocks. Annual differences were however present, and in year 1 (before the experiment), barley production with grass–clover atmospheric N fixation was the largest input and harvested grains were the main output, leaving about −45 kg N ha⁻¹ surplus at the soil surface. In year 2 (the year with sows), feed contributed 88% to the total input of 778 kg N ha⁻¹, and the output of 404 kg N ha⁻¹ was contributed by weaned piglets, resulting in a surface surplus of 374 kg N ha⁻¹. Nitrate leaching was the largest mass flow contributing to the soil N balance, although direct N emissions from animal manure compared to crop residues were also substantial (Table 2). In year 1, the soil N balance was similar between all treatments, whereas in year 2, it was favourable (closer to zero) for the paddocks with unmanaged (T) compared to pruned trees (P), whereas differences by resource position had little effect.

Table 1. Annual surface nitrogen balances (kg N ha⁻¹ year⁻¹) for four paddocks (total area ca. 0.4 ha) with a third covered by poplar trees. Year 1 (April 2022–March 2023) with barley undersown grass–clover; year 2 (April 2023–March 2024) with pasture for 12 sows supplied in three subsequent batches.

Mass Flow	Year 1—Barley	Year 2—Sows	Average Year
Feed, sow	-	480	240
Feed, piglets	-	210	105
Sow weight loss	-	28	14
Straw	-	29	15
Atm. deposition	11	11	11
Clover fixation	20	20	20
Seeds	5	-	2
Total input	36	778	407
Weaned piglets	-	348	174
Dead piglets	-	15	8
Disappeared piglets	-	41	21
Harvest barley grain	66	-	33
Harvest barley straw	15	-	8
Total output	81	404	243
Surface N balance	−45	374	165

Table 2. Annual soil nitrogen balances (kg N ha⁻¹ year⁻¹) for four paddocks (total area ca. 0.4 ha) with about a third covered by poplar trees. Year 1 (April 2022–March 2023) with barley undersown grass–clover; year 2 (April 2023–March 2024) with pasture for 12 sows supplied in three subsequent batches. Trees were either pruned (P) with woodchip mulch or unmanaged (tall, T), and resources (feed, hut) were placed either on the same (S) or opposite side (O) with trees in between. Nitrate leaching was obtained from a generalized linear model; different letters indicate significant differences within the year (confidence level p = 0.05).

Mass Flow	Year 1—Barley				Year 2—Sows				Average Year
	P O	P S	T O	T S	P O	P S	T O	T S	P O	P S	T O	T S
Surface N balance (Table 1)	−45				374				165
Direct emission of ammonia (NH3) volatilization	2				53				27
Indirect emission by denitrification	10	10	2	2	23	23	15	15	16	16	9	9
-manure	-				13				7
-grass–clover residues	0				0				0
-woodchips	8	8	-	-	8	8	-	-	8	8	-	-
-poplar leaves	1				1				1
-nitrate leaching	1				1				1
Actual nitrate leaching	122 ᵃᵇ	122 ᵃᵇ	152 ᵃ	113 ᵇ	231 ᵃ	214 ᵃᵇ	170 ᵇ	163 ᵇ	168 ᵃ	162 ᵃᵇ	161 ᵃᵇ	136 ᵇ
Total mass outflows	134	134	156	117	314	297	245	238	215	209	201	176
Soil N balance	−179	−179	−201	−162	51	68	120	127	−120	−114	−106	−81

4. Discussion

4.1. Remote Sensing Support of High-Resolution Spatio-Temporal Modelling of Nitrate Leaching

An added value of optical remote sensing is to depict explicitly the spatial variability to support modelling of soil water transport-related phenomena such as nitrate leaching. In silvopastoral systems with N hotspots, such knowledge provides an opportunity to design and implement tools for distribution of the hotspot over a larger area to reduce the hotspot magnitude. There are well-known weather datasets obtained from remote sensing via temperature, radiation and anemometer sensors. For studies at high spatial scales in the order of metres, such as this study, it is fair to assume a homogeneous weather distribution over the field occupied with paddocks of 0.4 ha, i.e., about 2 ha in total, housing 16 paddocks (Figure 1). For larger areas with tree hedges or undulated rolling terrains, weather inputs depicting variability to feed models like Daisy are invaluable; otherwise, large errors would propagate and accumulate in the simulated water balance and nitrate leaching [55].

Noteworthy is the micrometeorology between pruned and tall poplar trees, which might differ in relation to more surface exposed to radiation under pruned conditions [56], which in turn would increase evapotranspiration and lower percolation. However, studies show that this effect is temporary and pruned trees eventually dissipate less energy due to the smaller total leaf area [57], which in the context of this study fits the observation of increased nitrate leaching.

The spatial variability of the soil properties such as clay content is another potential role of remote sensing, i.e., proximal sensing, to aid nitrate leaching spatio-temporal depiction [24]. Over time, these properties do not vary significantly. However, they are a major determinant of percolation and nitrate leaching in agricultural fields, especially clay content and bulk density [22,24,55]. The soil physical properties used in this study turned out to be homogeneous across the paddocks, based on the best available data at 10 m spatial resolution. Otherwise, soil physical properties must be included in geostatistical modelling of nitrate leaching from heterogeneous soils, and increasing availability of proximal sensing makes these data available [58,59].

In this study, LAI was conveniently estimated from CC using the Beer–Lambert Law. Field measurements of LAI were not conducted; otherwise, the retrieved values could have been compared and the retrieval improved; however, the poplar canopies in the two treatments and the soil conditions were rather uniform. Also, models like Daisy are not that sensitive to subtle LAI variations, even less so for seasonal or annual water balances. It is, nevertheless, important for the retrieved values to be within an observed range of variation as increasing the accuracy of the LAI data could improve the quality of water balance modelling at higher temporal scales and also in studies where root density in soil layers is not measured and root uptake is assumed, which is often the case. The LAI development in this study (Figure 5) corroborated the LAI dynamics reported by other studies also for coppiced poplar, increasing up to the third growing season after coppice to values more than 5 [60]. Tripathi [61] reported a long-term maximum LAI of 9.5 for coppiced poplar (larger LAI due to intense branching and numerous leaves), which is comparable to the maximum LAI in the P treatment in this study of 7.5 two years after coppice (Figure 5). The values were higher than the 0.8–2.5 range for coppiced poplar derived from Sentinel-2 imagery by Cañete-Salinas [62] likely due to signal dilution by edge effects in the satellite data compared to the UAV data; also, the poplars were pruned annually in their study under a Mediterranean climate, compared to 3-year pruning in this study under a temperate wet climate, much more suitable for poplar growth.

Noteworthy is that a k-value of 0.5 is common across many plants irrespective of canopy architecture. Although k can vary depending on growth stage, it is not notably different than k-values of 0.56–0.59 reported for plants in shrubland and broadleaf forests [42]. Hence, we show a simple, reliable and efficient method of UAV-based RGB remote sensing to support the modelling of belowground processes. In Daisy, part of the precipitation (and overhead irrigation, e.g., sprinkler irrigation, if given) reaching the top of the crop is intercepted by the crop canopy, which acts as an interception storage. The direct water throughfall is assumed to be a function of LAI and the water intercepted by the canopy may be evaporated, stored or flow to the ground as canopy spill-off. Therefore, and especially at seasonal and annual scales, LAI plays a role in correct simulation of evapotranspiration and percolation used to estimate nitrate leaching (Figure 5 and Figure A1). With increasing computational power and machine learning-based retrieval of parameters [11,63,64], LAI data fed to process-based models offers an excellent opportunity for high-precision mapping of soil nitrate leaching.

4.2. Nitrate Leaching Remains a Problem for Silvopastoral Agroecosystems Under Humid Climates

The integration of geostatistical data with UAV-retrieved LAI and the Daisy model provided explicit spatio-temporal variation in nitrate leaching—a strong spatial heterogeneity otherwise difficult to capture [20]. Distinct localized hotspots occurred—especially in the grass zone before the study year and the tree area during the study—in areas where leaching exceeds 700–800 kg ha⁻¹. Notably, the patterns shifted between the two years, with many zones exhibiting lower or more diffuse leaching during the study year with pigs compared with the more concentrated high-leaching patches observed before the study year without pigs. It can only be assumed that the hotspots could intensify in the second year if there was no poplar zone. Also, studies reporting geostatistical outcomes of soil nitrate are few, reporting a semivariogram range (distance beyond which spatial correlation between sample values effectively disappears and observations become statistically independent) of 5–50 m, e.g., [65,66], as in this study, but 150–200 m can also be seen for larger or more homogeneous areas (e.g., clayey soils which tend to retain more water and nutrients, increasing spatial autocorrelation over larger areas) or coarser sampling schemes, e.g., [67].

As in previous agro-environmental studies in temperate climates, considerable soil nitrate leaching was found in this study at annual and seasonal scales (Figure 9 and Figure A2). Soil nitrate-N concentrations in the soil water were frequently above 20 mg L⁻¹. Nitrate concentrations as high as 200 mg L⁻¹ have also been reported by Manevski, Jakobsen [25], at a hotspot close to the hut. In the present study, a concentration this high occurred only once when one out of the three pooled suction cups did not collect water due to freezing. In the tree zone of PO treatment, high concentrations above 140 mg L⁻¹ were also sampled on several occasions (Figure 7) and similar responses have been reported earlier [25]. Leaching was similar to other farrowing paddocks with poplar trees (ca. 170 kg N ha⁻¹, Manevski et al. [25]) and farrowing paddocks without trees (126–276 kg N ha⁻¹ in Eriksen et al. [30]).

Earlier studies hypothesized that poplar trees reduce nitrate leaching, where leaching was indeed lower in the tree zone compared to grass [25]. However, it has also been reported that fertilization rates beyond 50 kg N ha⁻¹ exceeded the N retention of younger poplar trees (7 years old) in Florida [68]. Additionally, an effect of wood chips immobilizing mineral N was expected. Instead, nitrate leaching across paddocks and within the tree zone was the lowest in treatments with tall trees, whereas pruned trees with wood chips beneath showed peaks in soil nitrate, and hence, nitrate leaching (Figure 6 and Figure A3). Soil mineral N showed high values (March 2023, before pigs; October 2023, after 2nd batch; March 2024, after 3rd batch; [69]), hence, it is likely for nitrate leaching to have reduced soil mineral N before soil sampling, or the leaching was caused by continuous mineralization and less by excretion by pigs. Altogether, it indicated that wood chips could not immobilize mineral N in the current study, but added organic N which was mineralized. Additionally, pigs have been shown to root the soil with wood chips at a young age. This aerates the soil and thus provides necessary oxygen for nitrification. Soil moisture seemed to be higher in soil with wood chips [69], which would be another factor promoting the mineralization of organic N [70]. High leaching under pruned trees might also be caused by less developed root systems due to intensified investment in aboveground biomass, although this would also increase N demand. Pruned trees were infected with rust on 90% of their leaves in September 2023, which must have slowed their development. It is important to emphasize that soil water was sampled at a depth of 1 m, but roots of some poplar species can be as deep as 2.5 m, especially on sandy soils [71], and might uptake nitrate before it is leached to deeper layers. Thus, no clear conclusion on the leaching leaving the system can be drawn. Noteworthy are the legacy effects of outdoor pig activity even on sandy soils. The experimental paddock area has a long history with outdoor pigs in varying paddock designs, but always with the feed located in grass zone A. This might be a contributing reason for the high-leaching hotspots seen in year 1 (grass zone A, Figure 8).

4.3. Profiling Nitrogen with Empirical Data, Remote Sensing and Process-Based Model

Previous studies involved coupling in situ soil N data and remote sensing data to map soil N balance using machine learning, e.g., random forest [72,73]. Studies seldom employ process-based models to quantify soil N balance which is also tightly linked with the water balance, and this is especially true in silvopasture where evapotranspiration is considerable. Moreover, livestock could promote nitrate leaching and ammonia volatilization as grazing provides urine returns [74]. López-Díaz et al. [75] showed similar profiles of soil nitrate for grazed (low stocking rate of 3 sheep ha⁻¹) and mowed trees in silvopastoral systems with walnut and pasture in Spain. Thus, efficient control of the risk of nitrate leaching could be tested in other studies involving sheep instead of pigs and tree species other than poplar or willow, but evergreen.

Nitrate leaching was the dominant factor in the soil N balance in terms of magnitude, as also shown in other studies [20,25,26,76], and this remains a problem for all agricultural systems receiving ample N from feed or fertilizer in wet climates with surplus percolation. Annual nitrate leaching was on average lower in year 1 without pig occupation, but with more intense hotspots, compared to year 2 (Figure 8). This might be linked to the sows occupying the paddocks the year before the first observation year (2021/2022). Agronomic and engineering methods therefore need to be studied for optimization during- and post-animal husbandry over several years.

Spatial variability in N inputs from fertilizer or feed and outputs from harvest could be determined using sensor- and satellite-supported methods [77,78]. The annual surface N balances were substantial in the year with animal husbandry (year 2; >200 kg N ha⁻¹; Table 1). Such a situation has also been reported by other studies under comparable pedo-climatic and agronomic conditions [25]. Of the different treatments investigated, tree management had the strongest impact, with unmanaged trees showing soil N balances close to 0, as opposed to pruned trees showing highly negative soil N balances (Table 2). Therefore, the findings suggest that avoiding poplar pruning before inclusion of animals in silvopastoral systems under a temperate wet climate could be a beneficial management strategy. However, given the complex nitrogen dynamics in outdoor pig systems and site-specific factors, further studies are also needed to confirm or augment these effects.

5. Conclusions

This study showed how affordable and simple remote sensing, with UAVs providing true-colour data for LAI retrieval, can be integrated into detailed process-based modelling to support the depiction of explicit spatio-temporal patterns of nitrate leaching from the subsoil. The approach can be improved with larger involvement of machine and deep learning in the LAI estimation, although the choice of algorithm remains an open question. Also, measured LAI was not included but the effect on the annual time scale was deemed low; however, it should not be overlooked. Future studies should intensify the link between UAV optical data and process-based modelling to improve the knowledge on the spatio-temporal distribution of water transport in nitrate leaching from soils. Also, the study was conducted in silvopastoral settings, which are among the most complex and hence the principles should be applicable and tested across other agroecosystems. From an agronomic perspective, the study also revealed that tree management influences spatial and temporal patterns of nitrate leaching in outdoor pig systems. Unmanaged poplar trees showed lower leaching compared to trees harvested and mulched with wood chips, especially during the leaching season in autumn and winter. Placing resources on opposite sides should redistribute soil mineral N but does not necessarily reduce nitrate leaching when total surface balance is excessive. Hence, silvopastoral designs should carefully consider optimizing the uptake potential of tree vegetation and the use of wood chip mulch in addition to pig behaviour when aiming to mitigate nutrient losses.