The DeepRootLab facility was established at the experimental station of the University of Copenhagen in Taastrup, Denmark (55°40′ N; 12°18′ E). The soil was classified as Agrudalf. A detailed description of the soil physical and chemical condition at the study site is available in Supporting Information Table S1. Weather conditions recorded at the study site are presented in Fig. S1. In total 24 experimental plots (one plot = 19.5 m × 10 m in size) were laid out in four blocks (Fig. 1). The six plots in each block were laid out in a 3 columns × 2 rows layout, and the plot heads of the two rows were facing each other at a distance of 4.5 m. The distance between neighbouring plots was 3 m. The distance between blocks was 5 m in both column and row‐wise. We divided each plot into two kinds of plot layouts, namely intercropping setup (Layout I) and monocropping setup (Layout II). In Layout I, nine subplots were formed with alternating strips: five 1.5 m‐wide strips for perennial crops (1.5 m in width; 10 m in length) and four strips for annual crops next to each other (3 m in width; 10 m in length) for the study of crops in strip intercropping. In Layout II, the main plot was split into two subplots (9.75 m × 10 m), which can be used for individual crops or treatments. We chose and grew a wide range of perennial and annual crop species that are known to be deep‐rooted as shown in Table 2. No data from intercropping treatments are presented here.

We used two types of tubes for deep root study (Fig. 2), both inserted at an angle of 30° from vertical. The minirhizotron tubes are transparent acrylic tubes, 6 m long and have an outer diameter of 70 mm. These tubes were used for nondestructive root measurement with repeated imaging campaigns. The other type is access tubes which are stainless steel tubes (5.85 m in length, 110 mm in outer diameter). Three cut‐out openings were made along the access tubes, which matched soil depths of 0.7–1.2 m, 2.3–2.8 m and 4.0–4.4 m after installation. Ingrowth cores made of stainless steel with a volume of 3931 cm³ were made to be filled with soil and inserted into the access tubes, to allow the study of root growth and uptake processes at the different depths. In total 144 minirhizotron and 144 access tubes were inserted at a 30° angle vertically at the heads of each plot (Figs 1, 2). Inserting the tubes took place from late 2015 to early 2016 via auger drilling.

Time domain reflectometry (TDR) sensors (Acclima 310S; Acclima Inc., Meridian, ID, USA) were installed for monitoring soil volumetric water content (VWC). In selected plots, three replicated water sensors were installed between late 2017 and early 2018 at each of the depths 0.75, 1.5 and 2.5 m, with measurements being recorded on an hourly basis using four CR6 dataloggers (Campbell Scientific Inc., USA). All TDR sensors were located at least 3 m away from the plot edges (Fig. 1). The cables were buried at 0.3 m to allow the use of agricultural machinery.

We developed an AI‐based software for automated root image analysis called RootPainter (Smith et al., 2022). The software features a user‐friendly interface that can be easily operated by researchers without experience in Machine Learning. RootPainter utilizes a convolutional neural network that is built on the U‐Net architecture. After its release, the software was validated for root quantification during destructive sampling (Han et al., 2021b).

We monitored root growth dynamics of lucerne (2016–2019), intermediate wheatgrass (2016–2019), perennial lupine (2016–2019), mugwort (2016–2019), rosinweed (2017–2019), comfrey (2017–2019), curly dock (2017–2019) and tall fescue (2019–2020) by minirhizotron technique (Tables S2). Multispectral images were taken along the minirhizotron tubes at each 0.05 m‐length interval by a Videometer lab instrument (Videometer A/S, Hørsholm, Denmark). Multispectral images were obtained by using Light‐Emitting Diode (LED) light in five different wavelengths ranging from Ultraviolet A (UVA) to violet, amber, red and Near‐Infrared (NIR) was adjusted to 365, 405, 590, 660 and 970 nm, respectively. The camera system was mounted on a pull‐type wagon equipped with an automated deployment mechanism, which allowed the camera to be lowered to the desired soil depth to capture images at predefined depth intervals during retraction. The meta‐data, namely, plot number (PipeID), camera position (Position), Year, Month, Date and Time (hh‐mm‐sec) of the imaging were formulated into the file names for further identification and data processing (e.g. ‘PipeID_61_Position_442_TS_2019‐08‐23 09.15.40’). In the results shown, the images were converted to pseudo‐RGB images before further analysis. The size of each of the images was 40 mm × 50 mm (width × height) consisting of 2048 × 2448 pixels. From the images, root density was determined as planar root length density (pRLD) – root length per image area (cm cm⁻²; Wacker et al., 2025) after training the model using the deep learning software (see the Results section for detailed description). The maximum root depth of each tube was manually determined. For each crop species, we selected one image dataset per year during the crop season. The root depth of each species over accumulative thermal time (°C days) was computed.

From 2017 to 2019, we conducted three experiments (Exp 1, 2 and 3) to compare three crop species for deep root growth and nutrient uptake potential (Table 3). Each experiment included at least two perennial species with relatively earlier sowing dates (lucerne, mugwort and rosinweed) and later sowing dates (chicory, dyers woad and curly dock). In addition, winter cereals or a perennial cereal crop were included for comparisons (winter rye, intermediate wheatgrass and winter wheat).

We designed an experimental workflow called core‐labelling technique (CLT) for measuring deep root uptake (Han et al., 2020). Root‐free soil medium labelled with tracers was packed inside the ingrowth cores and inserted into the access tubes. Ingrowth cores were installed at two or three depths, matching the cut‐out openings of the access tubes. We used various tracers; here we present results obtained with ¹⁵N as a nutrient tracer. For each ingrowth core, 275.24 mg of ¹⁵NH₄Cl was prepared in solution and mixed with a subsoil medium (see Table S1 for physical and chemical soil characteristics) before packing the soil into the ingrowth cores. Tracer concentration determined from the shoot biomass was used as an indicator of root activity.

Ingrowth cores were inserted at two depth levels in Exp 1 and 2 (0.7–1.2 m and 2.2–2.8 m), whereas four depth levels were investigated in Exp 3 (0–0.3 m, 0.7–1.2 m, 2.2–2.8 m and 3.9–4.5 m), where an extra ingrowth core was installed directly in the topsoil to complement the three deeper placements using the access tubes. Shoot biomass was sampled twice, at the fourth and eighth weeks after insertion of the ingrowth cores. The dried and powdered samples were analysed for ¹⁵N enrichment. For each experiment, root growth of each crop species was monitored using minirhizotron technique.

The R software (R Development Core Team, 2019) was used for statistical analysis. Linear regression between segmentation and manual counts was performed to compute the R² values as it was affected by increasing annotation time. Root depth dynamics of the eight key species (lucerne, intermediate wheatgrass, perennial lupine, mugwort, rosinweed, comfrey, curly dock and tall fescue) between the years (2016 to 2020) were analysed using Tukey HSD (P ≤ 0.05). pRLD of the crop species was compared between the growing seasons regardless of the soil depth. More detailed comparisons within seasons in 2017, 2018 and 2019 between the image campaigns were done for four crop species (lucerne, intermediate wheatgrass, rosinweed and tall fescue). For three experiments conducted, pRLD measured by minirhizotrons was compared between the crop species at 0–0.3 m, 0.3–0.7 m, 0.7–1.3 m, 1.3–2.3 m, 2.3–2.8 m and 2.8–4.0 m of soil depth. RLD and δ¹⁵N (‰) were compared between the crop species at soil depths of 0–0.3 m, 0.7–1.3 m, 2.2–2.8 m and 3.9–4.5 m within each experiment. Linear regressions of RLD vs δ¹⁵N (‰) measured from the ingrowth cores, RLD by ingrowth cores vs root length by minirhizotron and root length by minirhizotron vs δ¹⁵N (‰) were computed combining the data from the nine crop species used in the three experiments. R ² values were computed for each crop species.

We determined root depth dynamics of the eight perennial crop species (Fig. 6). Overall, there was a clear trend of increasing average root depth over the years by all measured crop species. The root penetration rate computed by regression analysis between root depth over the thermal time (°C days) was slowest for lucerne (0.07 mm °C d⁻¹) followed by perennial lupine (0.08 mm °C d⁻¹), intermediate wheatgrass (0.1 mm °C dys⁻¹), mugwort (0.2 mm °C d⁻¹), tall fescue (0.24 mm °C d⁻¹), curly dock (0.26 mm °C d⁻¹), rosinweed (0.29 mm °C d⁻¹) and comfrey (0.38 mm °C d⁻¹). Maximum root depth – the deepest point of root observation was shown by curly dock and rosinweed in 2019 at 3.8 m, followed by tall fescue (3.5 m), intermediate wheatgrass and comfrey (2.9 m), lucerne (2.7 m) and mugwort (2.6 m).

We compared the root density (measured as pRLD (cm cm⁻²) of eight perennial species between crop seasons at each 0.5‐m interval to 4.0 m of soil depth (Fig. 7)). Overall, root length generally increased as the crops grew older. The effect was most pronounced below 0.5–1.0 m of soil depth, except for mugwort and tall fescue. Intermediate wheatgrass exhibited greater root length in its second growing season (in 2017), especially in the upper soil layer (0–0.5 m), but its root density decreased over the following 2 yr. Perennial lupine did not show any effect of crop age on root length. All crop species revealed deeper rooting from the first to the final point of observation (note the sowing date and stand age by referring to Table 3).

We compared the root density of lucerne, intermediate wheatgrass and rosinweed at each depth‐level between the two observation times within the crop seasons in 2017, 2018 and 2019 (Fig. 8). Root density of lucerne remained unchanged in 2017 between the observation points, but in the following 2 yr, an increase in root density was shown below 1.0–1.5 m soil depth. In 2019, a significant decrease in root density was observed in the topsoil. Intermediate wheatgrass showed little effect of measurement time, but with a tendency to declining root density later in the season. Rosinweed exhibited an increase in root density within the season in 2017 (0–0.5 m) and 2019 (0.5–1.0, 1.0–1.5, 1.5–2.0 m).

Using minirhizotron technique, we observed a clear difference in root depth among the compared crop species (Fig. 9). In Exp 1, chicory showed the deepest roots at 2.2–2.3 m depth followed by lucerne (2.1–2.2 m) and winter rye (1.4–1.5 m). In general, winter rye had the lowest and chicory the highest root density along the soil depth profile. Root density of chicory was greater than lucerne in the upper soil layers (0–0.7 m).

In Exp 2, dyers woad showed the deepest roots (2.4–2.5 m), followed by mugwort (1.8–1.9 m), and intermediate wheatgrass (1.3–1.4 m). In the topsoil, root density of mugwort was greater than that of other crops. There was some variability with depth, but in general, intermediate wheatgrass showed lower root densities than dyers woad and mugwort.

We found roots of curly dock to 3.8 m depth in Exp 3. Winter wheat and rosinweed grew roots to 2.8–2.9 m and 3.3–3.4 m, respectively. Except in the topsoil layer, curly dock exhibited a higher root density compared with rosinweed and winter wheat.

We compared deep root activity of three crop species by inserting ingrowth cores packed with ¹⁵N‐labelled soil in Exp 1, 2 and 3. Two parameters were measured in all experiments, namely RLD (cm cm⁻³) from the root extraction from the soil in the ingrowth cores (Fig. 10a) and δ¹⁵N (‰) concentration measured from aboveground biomass (Fig. 10b).

In Exp 1, the ingrowth cores were inserted at 0.7–1.2 m and 2.2–2.8 m depth levels. At 0.7–1.2 m, chicory exhibited greater RLD and δ¹⁵N compared with lucerne, while intermediate results were shown by winter rye. No difference in RLD was found between the crop species at 2.2–2.8 m. However, δ¹⁵N of winter rye and chicory was greater than lucerne at this depth.

Ingrowth cores were inserted at the same depth levels in Exp 2. No meaningful difference in RLD and tracer concentration was found in Exp 2. Overall values of RLD and δ¹⁵N were low, especially at 0.7–1.2 m.

In Exp 3, the following four depth levels were investigated: 0–0.3 m, 0.7–1.2 m, 2.2–2.8 m and 4.0–4.5 m. In the topsoil layer (0–0.3 m), RLD of winter wheat was significantly lower than curly dock, but δ¹⁵N of winter wheat was substantially greater than both other crops. No difference in RLD was found at 0.7–1.2 m, yet winter wheat exhibited greater δ¹⁵N compared with rosinweed. At 2.2–2.8 m, greater δ¹⁵N was found for rosinweed and curly dock compared with winter wheat, whereas no difference in RLD was found. RLD of curly dock was greater than rosinweed at 2.2–2.8 m. A similar trend was found for δ¹⁵N at 4.0–4.5 m; however, the difference was not significant.

Regression analysis between the root density found inside the ingrowth cores and tracer uptake (δ¹⁵N) measured from aboveground biomass showed a significant relationship (P ≤ 0.001) with R ² values of 0.67 when all data from the three experiments were combined (Fig. 11a). When the same analysis was performed for each crop species, it revealed strong correlations for chicory (R ² = 0.97; P ≤ 0.001) and winter wheat (R ² = 0.92; P ≤ 0.001). Except for lucerne, dyers woad and mugwort, the relationship between RLD and δ¹⁵N was found to be significant for the rest of the crops.

We performed linear regression analysis between root density found inside the ingrowth cores and root density measured by minirhizotron imaging (Fig. 11b). The relationship was found to be significant (P ≤ 0.01), however, with a small R ² value (0.13). When performed on each crop species, only dyers woad showed a significant relationship with a moderate R ² value (0.6; P ≤ 0.01).

Similarly, a weak but significant relationship (R ² = 0.13; P ≤ 0.05) was found between root density observed by minirhizotron and tracer uptake (δ¹⁵N; Fig. 11c). Only intermediate wheatgrass exhibited a significant relationship between these two measurements (R ² = 0.83; P ≤ 0.05*).

TDR sensors gave us accurate measurements of the soil water content evolution over time. Soil water content followed a wetting and drying cycle, associated with seasonal differences in crop growth and rainfall patterns (Fig. 12). These differences also influenced the groundwater level which is characterized by relatively high ‘standing’ groundwater level in winter and low groundwater level (i.e. below 2.5 m depth) during the summer (Fig. 12e). Over the three growing seasons, all crops were able to deplete water from the 0.75‐m soil layer, with values reaching close to wilting point (i.e. 17% VWC; Fig. 12a–d). However, the rate of water depletion at this depth differs among crops. Perennial lupine, chicory and lucerne quickly depleted this soil layer, which almost all its available water used by early June. By contrast, slower rates of water depletion were observed for intermediate wheatgrass and dyers woad with soil water content reaching a value close to wilting point in July (Fig. 12b,c). Along with these results, perennial lupine, chicory and lucerne depleted substantially more water from the 1.5 m soil depth compared with intermediate wheatgrass and dyers woad. Indeed, at this soil depth, perennial lupine, chicory and lucerne induced c. a 5–6% drop in soil water content, whereas intermediate wheatgrass and dyers woad induced smaller changes of c. 2–3% VWC. However, no crops depleted this soil layer down to a value close to the wilting point. At 2.5 m soil depth, only perennial lupine induced a noticeable decrease of soil water content of 2% VWC, whereas no water use from this layer was shown for any of the other crops.

For both types of tubes, that is minirhizotron and access tubes, we faced challenges for insertion due to relatively high clay content (< 20%; Table S1) at the study site as well as high water table (c. 1 m) at the time of installation during the winter and early Spring in 2016–2017. The long tubes also strongly increase the friction while installing them into the holes after drilling. For minirhizotron tubes, therefore, it was unavoidable to make the holes larger than the tube diameter. This caused a gap between the upper soil surface and the tubes. We attempted to offset this issue by attaching raise‐blocks at the bottom side of each tube to ensure better soil‐tube contact at the upper side of the sloping tubes. The remaining gaps were filled with sand materials. However, this caused an artificial background on many images with the added sand, and for some tubes, the gaps were not filled (Figs 4, 5). This might have caused a low‐resistance path leading to an artificial increase in root growth (Rewald & Ephrath, 2013; and references therein). However, the minirhizotron tubes were inserted under the same protocol across the platform for a short period of time. Therefore, it is unlikely that the differences in deep root density between the crop species and observation timing were artefacts. For the metallic access tubes, we tackled this insertion issue by using sonic‐drilling – in which the resonating energy from a high‐frequency oscillation (e.g. 150 Hz) in the drill head ensures better penetration into the deep soil layers. After installation, we observed occasionally that some soil from upper opening areas fell into the tubes – which may have created an artefact for root growth by reducing root–soil contact.

One of the main advantages of the minirhizotron method is its capacity to measure gross root growth without interfering with the previously measured area/depth. We have further automated image capture by installing a programmable logic controller, which generates file names with useful meta‐data such as the plot number, location of camera (in terms of cable length projected) and timing of image capture as illustrated by Svane et al. (2019a). This enabled us to collect a large amount of data and process the images without much manual effort (e.g. re‐naming of the files).

Integrating the automated image capture with deep learning‐based analysis, we achieved a high efficiency of root data acquisition (c. 12 s per image). This certainly is a substantial improvement compared with manual root counting and annotation in terms of time saving. For example, c. 1 to 1.5 h of annotation time was required for 100 cm² of root image area (Ingram & Leers, 2001), which now can be done in 1 min using our approach (even including the data capture and image processing time).

Our results showing the relationship between annotation time and root length correlation inform the RootPainter community on the trade‐offs of extended annotation for minirhizotron segmentation models. We found a large portion of model improvements to happen within the first 200 min, with extended annotation continuing to provide improvements but with diminishing returns (Fig. 3).

Some of our images captured the new root growth between the measurements taken at relatively short intervals. For example, lateral branching of lucerne roots at 0.35 m within 2 weeks was observed (Fig. 5a,b), which was successfully captured by the trained model showing the increase in root length from 11.21 cm to 15.57 cm per image. It was intriguing to see the rapid colour change of the main root from white to brown upon proliferation. Also, the imaging captured root growth of intermediate wheatgrass into depth, growing below 2 m within the growing season. This new occupancy of roots was observed between the measurements taken in July and September in 2018. Here, we show the images from June and September 2018 which better represent the location of imaging (Fig. 5c,d). This was also accurately determined by the trained model with the root length increasing from 0 cm to 7.3 cm. We were able to capture newly grown roots grown into the area at 1.21 m where dark roots of rosinweed were present from the growth in previous seasons (Fig. 5e,f). As a result, our approach was sufficiently accurate to capture the within‐season variation (May to August in 2017, 2018 and 2019) in root density of crop plants (Fig. 8), which makes consistent observation possible leading to a high‐time resolution dataset.

Although multispectral information was not used in this study, previous work has demonstrated the potential of spectral approaches for root classification. Multispectral imaging, including UV wavelengths has been used for root classification (Bodner et al., 2018; Svane et al., 2019a). The focus here was on automated root segmentation and data generation as a facility; however, future work shall exploit these capacities, with appropriate chemical and genetic validation, to classify root vitality and potentially enable species differentiation (Lombardi et al., 2025).

Preinstallation of access tubes, insertion of ingrowth cores into the tubes containing tracer‐labelled soil made deep root nutrient uptake measurement possible up to 4.5 m soil depth. Two clear advantages of this CLT have been already addressed in detail by (Han et al., 2020). In short, this approach overcomes the disadvantage of the generic tracer injection, which can lead to movement of the injected aliquot depending on soil moisture content and rainfall (Hoekstra et al., 2014). Second, we were able to determine the roots of the plants that had access to the tracers. This ensures and validates that the root activity detected in the aboveground biomass was the result of roots growing into the tracer‐labelled soil of the ingrowth cores. This eradicates the potential discrepancy between root uptake and root density data, which can potentially occur under tracer injection methods (e.g. Da Silva et al., 2011). However, it is unclear how well the RLD within the ingrowth cores represent the RLD in the bulk soil around it. Our validation between aboveground tracer concentration and belowground root density showed a range of variation (R ² values between 0.31 and 0.97) leading to an average of R ² 0.67 (Fig. 11) confirming the reliability of the approach.

However, there was a discrepancy between the two root methods, namely, RLD by ingrowth core and pRLD by minirhizotron methods, despite the significant correlations (P ≤ 0.005). Several factors might have led to this disagreement between the two methods. First, the roots growing into the soil packed into the ingrowth cores may not represent those growing near the inserted minirhizotron tubes well. Several soil conditions differ (e.g. nutrient content; Table S1), and the roots in the bulk soil were typically well established before insertion of the ingrowth cores, while the roots in the ingrowth cores are the young roots which recently colonized the ingrowth‐core soil. Also, the 2‐month incubation time for ingrowth cores in this study might have been long enough for the freshly grown fine roots inside the ingrowth cores to undergo root turnover, whereas the roots visible to minirhizotron tubes might have been less susceptible to decomposition (Higgins et al., 2002), that is not completely exposed to soil microbes. In short, our results show that the two methods are not necessarily in agreement when determining root growth, and caution is required in interpreting the results when independently used.

The experimental setup in the DeepRootLab facility allows differentiation of crop water uptake in deep soil layers (i.e. > 1 m). Perennial lupine, chicory and lucerne were found to substantially use water down to 1.5 m soil depth and even deeper water uptake (i.e. down to 2.5 m) seems to occur under perennial lupine. Such results had also been for older lucerne crops, at the same location (Clément et al., 2022). The slower and more shallow water use by intermediate wheatgrass and dyers woad probably stemmed from different reasons. Dyers woad is a biennial crop, it produced deep roots, but matured early in the second season, terminating its water use in during June. Intermediate wheatgrass, on the other hand, is a perennial crop being more water spending. There are three drought escape strategies, (1) Escaping: usually achieved through shorter crop cycle (2) Water saving: through conservative use of water and (3) Water spending: using water and allocating more biomass to the growth of deep and abundant roots (Bodner et al., 2015; and references therein). Therefore, identifying crop water use strategies is of high importance to build cropping system that suits environmental conditions and maximize soil moisture capture (Blum, 2009). By allowing continuous measurements of root growth, water and nutrient uptake in deep soil layers, DeepRootLab constitutes a unique facility for the study of deep root growth and functions at the field crop and agricultural crop rotation level.

We have successfully established a field facility in which a variety of measurements on deep root growth and function can be made. It allows the relatively easy and repeated study of root growth and function to great depth under field conditions, by using preinstalled minirhizotrons, access tubes and soil water sensors. The extent and design of the facility allow experiments with up to 48 plots, allowing several species or treatments to be studied in well‐replicated studies. The facility is an ideal platform for conducting studies at deep soil layers in the field with a capacity for generating statistically and biologically meaningful results.