Using Drones to Predict Degradation of Surface Drainage on Agricultural Fields: A Case Study of the Atlantic Dykelands

Abstract

Excess water in agricultural fields can significantly limit crop productivity. Drone technology offers solutions for identifying and predicting drainage degradation. This study utilized drone-based photogrammetry to create high-resolution elevation models, multispectral imagery for vegetation indices, and flood simulations models to identify zones at risk of poor surface drainage. These models, collected from 2021 to 2023, were used to assess the relationship between poor drainage and corn productivity. The findings revealed a substantial decline in productivity in poorly maintained surface drainage areas, notably a decrease in mean plant height from 1.43 m in 2022 to 0.26 m in flood-prone areas in 2023. Flood-prone zones expanded significantly, from 37% to 61% of the field area between 2022 and 2023, emphasizing the negative cumulative impacts of pre-existing drainage issues. Conversely, fields receiving regular annual maintenance showed an increase in mean plant heights (from 2.23 m to 2.54 m) and NDVI values, reflecting improved drainage conditions. This research demonstrates the effectiveness of drone-derived elevation models for proactively identifying problematic drainage areas, allowing farmers to make informed decisions regarding field maintenance. Implementing these technologies can optimize drainage management practices, enhance agricultural productivity, and increase economic viability in regions that rely on surface drainage.

1. Introduction

The province of Nova Scotia, Canada is a coastal region that experiences high rainfall, receiving an average of over 1300 mm of precipitation each year [1]. Much of the precipitation occurs in the Spring and Fall, which can delay agricultural fieldwork, especially if field drainage is insufficient. The inability to maintain adequate drainage on agricultural fields impedes the aeration of the plant root zone required for crop production [2]. In Nova Scotia, field drainage is often the limiting factor of the type of crops that can be successfully grown, especially on the dykelands, which are more prone to these challenges [3]. Dykelands are low-lying agricultural land, reclaimed from the sea, and protected by dykes. These fields are contained within multiple dyke systems along the Bay of Fundy. In Nova Scotia, the total surface area of these dyke systems is 17,401 hectares, of which 9272 hectares are actively used for agriculture [4].

The soils on the dykelands are characterized by their naturally low permeability, which slows both the downward and lateral movement of water through these soils [5]. Combined with a flat landscape, this causes water to accumulate on the surface and water-log the fields. However, the agricultural productivity of these soils can be significantly improved through proper surface drainage [2,3]—the orderly removal of excess water from the land’s surface. On the dykelands, this is accomplished by shallow ditches, which discharge into larger and deeper collector drains that eventually evacuate excess water into the ocean [6]. To facilitate the flow of excess water toward the drains, the field is given an artificial slope through land forming [7], which is the process of mechanically moving soil to change the topography of the field and thereby improving surface drainage [8]. The soil excavated from the drain is pushed into hills called “crowns”. When these crowns have a uniform slope, surface water from heavy precipitation can effectively run off into the adjoining ditches [7]. Land forming is the principal drainage technique used on the dykelands. In 2020, it was estimated that 6924 hectares of fields actively used for agriculture in Nova Scotia were land formed [4].

Assessing the degradation of surface drainage on dykelands remains a complex task due to the unique characteristics of the soil. Traditional assessment methods have relied on visual observations conducted by farmers after rainfall events. These observations typically involved walking the fields to manually identify and mark ponding areas [2,9]. Basic surveying tools, such as optical or hand levels, have also been used to measure elevation and map depressions. More recently, surveying-grade Global Navigation Satellite System (GNSS) receivers have been incorporated into the process to improve accuracy [3]. Despite these advancements, the methods remain labour-intensive, time-consuming, and often too imprecise for detailed assessments. As a result, farmers have traditionally relied on periodic land-forming operations, such as land grading, across entire fields to maintain effective drainage rather than adopting a more targeted approach.

Recent studies have shown that drones equipped with RGB (red, green, blue) cameras can be used to measure plant height in corn fields. Che et al. [10] and Gilliot et al. [11] demonstrated that 3D surface models generated from high resolution drone RGB imagery can be used to accurately predict plant height in corn fields. Other studies, such as Bending et al. [12] and Barrero Farfan et al. [13], showed that plant height is a good indicator for evaluating plant growth and grain yield. Further studies have shown that drones can effectively assess plant health in agricultural fields using vegetation indices. In Janousek et al. [14], they examined the effectiveness of various vegetation indices (NDRE, NDVI, GNDVI) derived from drone-mounted multispectral cameras in estimating the quantity of dry matter in corn. Their findings revealed significant correlations between these indices and the nutritional values of dry matter, which are important for yield estimates. Similarly, Tsakmakis et al. [15] investigated the correlation between Normalized Difference Vegetation Index (NDVI) values and corn yield, finding strong positive correlations (r > 0.8). Macedo et al. [16] also explored the use of the NDVI to estimate productivity and above-ground biomass in corn. They obtained similar results, demonstrating strong correlations between these indices and corn productivity, underscoring their value in predicting yields. The use of UAV imagery has also been demonstrated for mapping subsurface drainage systems in agricultural fields. Koganti et al. [17] compared the effectiveness of visible-color, multispectral, and thermal infrared cameras for drainage mapping purposes. Li et al. [18] utilized Unmanned Aerial Vehicle (UAV)-based LiDAR to enhance flood modelling accuracy by capturing micro terrain features. Chidi et al. [19] highlighted the importance of high-resolution DEMs in soil erosion estimation, demonstrating significant sensitivity to (Digital Elevation Model) DEM resolution.

The improvement of water movement in the fields, by means of land forming, extends the growing season and enhances the field trafficability early in the Fall [8]. Over time, however, field topography changes after plowing, and dykeland fields must be reformed—a process is known as “recrowning”. Additionally, if the ditch drains are not properly maintained and become obstructed by sediments, the water table will tend to rise. This results in poor drainage of gravitational water towards the side of the fields, thus reducing crop growth [7]. The cost associated with recrowning dykelands is estimated to range between $300 and $500 an acre and is performed, on average, every ten years [20]. To date, the negative impacts of poor surface drainage from one season to another have not been well documented on the dykelands, although they have been reported in the literature [3,21,22].

Despite the proven effectiveness of drones and remote sensing in agricultural assessments, there is a lack of studies specifically addressing the evaluation of seasonal changes in surface drainage conditions on dykeland fields in Nova Scotia. The unique characteristics of these fields—such as their low-lying nature, susceptibility to waterlogging, and the practice of land forming—present specific challenges that have not been adequately explored using remote sensing technologies. Unlike previous research focusing primarily on yield estimation, this study uniquely addresses the seasonal monitoring of subtle topographical changes influencing surface drainage. This aspect, which is critical for managing agricultural fields that rely on surface drainage, remains largely underexplored.

The goal of this research is to bridge this knowledge gap by evaluating the accuracy of drones and remote sensing techniques in identifying seasonal changes across agricultural dykeland fields by comparing remote sensing data with ground-truth measurements. Specifically, this study seeks to determine whether these technologies can help farmers identify low-lying areas with poor surface drainage, which may not be evident through traditional methods. By providing accurate data, these assessments aim to support farmers in making informed decisions regarding the frequency and necessity of recrowning dykeland fields. This approach has the potential to optimize field management practices, improve crop yields, and ultimately enhance the economic viability of agricultural operations on the dykelands.

2. Materials and Methods

Figure 1 illustrates the overall workflow process used in this research. Aerial images were collected to generate elevation data and compute vegetation indices. The elevation data was then used to carry out a flood simulation and predict surface drainage patterns. The aerial images were processed to develop plant height models. These models and vegetation indices were analyzed to identify low-productivity zones on the dykelands. The outcomes from both analyses were combined to assess the relationship between poor surface drainage and crop productivity.

2.1. Study Area

Four dykeland fields were evaluated in both Truro and Grand-Pré, Nova Scotia, Canada (Figure 2). Fields A, B, C, and D are located along the Bay of Fundy and are parts of the Central Onslow, Cobequid and Grand-Pré and dyke systems.

Table 1. Field size, location, and crop production of the study areas.

Field Name	Location	Coordinates	Area (ha)	Field Production
				2021	2022	2023
Field A	Truro	45.3735° N, −63.3305° W	13.5	Corn	Corn	Alfalfa-Grass
Field B	Truro	45.3761° N, −63.2870° W	6.2	Corn	Soybeans	Corn
Field C	Grand-Pré	45.1109° N, −64.3208° W	3.2	Corn	Corn	Corn
Field D	Grand-Pré	45.1086° N, −64.3202° W	5.6	Corn	Corn	Corn

details the size of the fields along with the field production through the years. Crops and forage grown on the study sites included soybeans, corn, and alfalfa-grass, which are all commonly grown on the dykelands [23].

The four fields differed in management practices, soil types, and drainage systems. Fields A and B were actively managed by the landowner to minimize surface damage, including erosion from water runoff, compaction, and ruts caused by heavy machinery, and ponding or waterlogging due to inadequate drainage. A central component of the management strategy was the annual maintenance of field ditches. Each fall, the landowner cleared and re-ditched these channels to improve surface drainage and enhance field usability.

In addition to ditch maintenance, Fields A and B also employed a combination of grassed waterways and land-forming techniques. Field elevation was adjusted using a rotary ditcher, which redistributed ditch spoil to level uneven areas. Culverts were strategically placed to support efficient water movement and maximize land use.

In contrast, none of these maintenance activities were carried out on a yearly basis in Fields C and D. As a result, Fields A and B benefited from consistent ditch maintenance, deliberate soil management to reduce compaction, and the use of grassed waterways to support effective surface drainage (

Table 2. Drainage characteristics of fields investigated in this study.

Field Name	Drainage Technique *	Open Ditches Spacing	Last Recrowning	Soil Type
Field A	GW, LF	92–183 m	Yearly maintenance	Minas lowlands
Field B	LF	46–61 m	Yearly maintenance	Minas lowlands
Field C	GW, LF	46 m	>10 years	Annapolis valley soil
Field D	LF	51–87 m	>10 years	Annapolis valley soil

* GW: Grassed waterways, LF: Land forming.

).

2.2. Aerial Surveys

Aerial images were acquired with a DJI Matrice 300 RTK (SZ DJI Technology Co., Ltd., Shenzhen, China) (

Table 3. Characteristics of the DJI Matrice 300 RTK drone used for mounting the payloads.

Matrice 300	Characteristics
Dimensions	81 cm × 67 cm × 43 cm
Weight	6.3 kg
GNSS	GPS + GLONASS + BeiDou + Galileo
Maximum payload	2.72 kg
Maximum flight time	55 min
Maximum speed	22.78 m/s

). All surveys were planned and flown using the DJI Pilot app (SZ DJI Technology Co., Ltd., Shenzhen, China) at an altitude of 106 m. A D-RTK2 base station (SZ DJI Technology Co., Ltd., Shenzhen, China) was used during each survey to minimize positioning errors. Across all fields and years, Ground Control Points (GCPs) and checkpoints were surveyed across the study areas and used during the post-processing to align the aerial images with GCP locations [24].

On Fields A, C, and D, ten GCPs and seven checkpoints were used, while Field B had six GCPs and four checkpoints for each survey date. The GCPs were evenly distributed around the edges and center of each field, placed on flat ground with an unobstructed view of the sky and positioned away from obstacles. Each GCP consisted of a 45 cm by 45 cm checkerboard marker with a clearly marked center intersection. For 2021, coordinates of the GCPs and checkpoints were measured using a Topcon HiPer SR RTK-GPS (Topcon Positioning Systems Inc., Livermore, CA, USA) in a base station and rover configuration. In 2022 and 2023, an Emlid Reach RS2 (Emlid Inc., Hong Kong, China) multi-frequency GNSS receiver, linked to a networked transport of RTCM via internet protocol (NTRIP) correction service, was used. Both receivers are capable of centimetre-level accuracy in Real-Time Kinematic (RTK) modes and have similar horizontal and vertical accuracy [25,26]. All data collected from the surveys were converted to a local coordinate system NAD 83 (CSRS)/UTM zone 20N. During data collection, each GCP and checkpoint were observed for one minute under a fixed RTK solution and a position dilution of precision (PDOP) below two. The mean root mean square (RMS) errors for all measured points were, on average, 1.1 cm in easting, 1.5 cm in northing, and 1.3 cm in elevation.

2.3. Payloads

Two successive survey missions were conducted on each of the survey dates. The fields were first surveyed with the Matrice 300, equipped with a MicaSence Altum multispectral sensors (MicaSense, Inc., Seattle, WA, USA), and secondly with a DJI Zenmuse P1 (SZ DJI Technology Co., Ltd., Shenzhen, China). The Altum captures images in five spectral ranges, including spectral bands in blue (Band 1—459–491 nm), green (Band 2—546–573 nm), red (Band 3—661–675 nm), red-edge (Band 4—710–723 nm), and near-infrared (Band 5—813–870 nm). The Zenmuse P1 is a 45-megapixel optical camera, which was fitted with a 35 mm lens. Due to the unavailability of the P1 camera in 2021, all images of the surveys were conducted with the Altum camera.

Table 4. Basic characteristics of the DJI Zenmuse P1 optical camera and MicaSense Altum multispectral camera used to obtain aerial images of dykelands.

Characteristics	Zenmuse P1	MicaSense Altum
Dimensions (cm)	19.8 × 16.6 × 12.9	8.2 × 6.7 × 6.45
Weight	787 g	357 g
Spectral Bands EO	Blue, Breen, Red	Blue, Breen, Red, Red-edge, NIR
Spectral Bands LWIR	N/A	Thermal Infrared 8–14 um
Sensor Resolution	8192 × 5460 pi	2064 × 1544 pi (MSI), 160 × 120 pi (TIR)
Capture Rate	1 capture every 0.7 s	1 capture per second (all bands), 12-bit RAW
Field of View	53.63° × 36.96°	48° × 37° (MSI), 57° × 44° (TIR)
Pixel Size	4 µm	3.45 µm (MSI), 12 µm (TIR)
Effective Pixels	45 MP	3.2 MP per EO band

TIR = thermal infrared, MSI = multispectral imaging.

contains the characteristics of the two payloads used to obtain aerial images. Surveys were conducted with the purpose of generating three sets of data, namely Digital Terrain Models (DTMs), Digital Surface Models (DSMs), and orthomosaics.

2.4. Survey Frequency

The survey missions were carried out from May 2021 through August 2023, with two surveys being conducted annually across the study sites (Figure 3). The scheduling of these surveys was designed to ensure align with the agronomic milestones of corn development. The first survey of the year took place in the Spring, prior to seeding, when the soil was exposed. The second survey was conducted in early August, before the corn harvesting period and approximately three weeks after the corn had reached physiological maturity.

All surveys were deliberately planned for days with clear sky conditions to ensure optimal data collection.

Table 5. Aerial survey parameters used with the Zenmuse P1 and MicaSense Altum. The table outlines differences in Ground Sampling Distance (GSD), overlap percentage for image stitching, and the operational speed of each system used to conduct the survey missions.

Characteristics	Zenmuse P1	MicaSense Altum
Altitude AGL (m)	106	106
GSD (cm/px)	1.33	4.57
Front Overlap	75%	80%
Side Overlap	75%	75%
Speed (m/s)	15	10

AGL = Above Ground Level, GSD = Ground Sampling Distance.

contains details of the survey mission parameters used to collect the aerial images.

2.5. Data Processing

Aerial photos of the study areas were processed using the Agisoft PhotoScan 1.8.2 software (Agisoft LLC Inc., St. Petersburg, Russia). The processing of the images followed a standard Structure from Motion processing workflow [24,27]. Figure 4 details the processing steps taken to generate DTMs, DSMs and orthomosaics from the RGB and the multispectral images.

Table 6. Summary of the processing parameters used in Agisoft PhotoScan.

Category	Parameter	Value
Alignment Parameters	Accuracy	High
	Generic preselection	Yes
	Reference preselection	Source
	Key point limit	40,000
	Key point limit per Mpx	1000
	Tie point limit	4000
	Exclude stationary tie points	Yes
	Guided image matching	No
	Adaptive camera model fitting	No
Optimization Parameters	Parameters	f, cx, cy, k1-k3, p1, p2
	Adaptive camera model fitting	Yes
Depth Maps	Quality	High
	Filtering mode	Mild
	Max neighbors	16
DEM	Coordinate system	NAD83(CSRS)/UTM zone 20N
	Source data	Point cloud
	Interpolation	Enabled
Orthomosaic	Blending mode	Mosaic
	Surface	DEM
	Enable hole filling	Yes
	Enable ghosting filter	No

summarizes the processing parameters used for image alignment, DEM creation, and orthomosaic production.

DTMs were created in Agisoft (v1.8.5) using dense 3D point clouds from bare soil surveys conducted in the spring, while DSMs were generated in the CloudCompare (v.2.12) software (GPL software, cloudcompare.org, URL accessed on 14 June 2024) from the dense 3D point clouds exported from Agisoft during the August surveys (Figure 3). The 3D point clouds were filtered using the statistical outlier removal tool in CloudCompare and transformed into a 2.5D model using the rasterize tool.

Table 7. Digital elevation model accuracy report.

Year	Survey ID	Field Coverage	Control Points RMSE (Z, cm)	Ground Resolution (cm/pix)	Point Density (Points/m2)
2021	TR_A_DTM_2021	A	0.91	4.07	605
	TR_B_DTM_2021	B	0.12	4.18	572
	GP_CD_DTM_2021	C, D	0.69	4.05	609
2022	TR_A_DTM_2022	A	6.26	1.11	834
	TR_B_DTM_2022	B	5.63	4.43	508
	GP_CD_DTM_2022	C, D	2.80	2.23	201
2023	TR_A_DTM_2023	A	3.62	2.19	208
	TR_B_DTM_2023	B	5.17	2.22	203
	GP_CD_DTM_2023	C, D	4.12	4.46	503

contains information on the accuracy of the elevation models used in this research.

2.5.1. Corrections of Multispectral Images

The Altum was calibrated using a Calibrated Reflectance Panel (CPR) to prevent banding and patchiness in the orthophotos and to enable more accurate compensation for incident light conditions. This process ensures that multispectral images can be used for accurate analysis across multiple datasets [29]. The CPR was helpful to improve the radiometric quality of the images since it provides a definite reflectance value during the corrections process. The surface of the calibrated reflectance panel has been measured at numerous wavelengths using a spectrometer. Panel captures were acquired just before and after each flight as well as during each battery swap.

2.5.2. Vegetation Index

The NDVI is a standardized index used to quantify the health and density of vegetation. This index leverages the contrast between the characteristics and chlorophyll pigment absorption of the red and near-infrared bands (NIR) [30]. The following formula was used to compute the NDVI:

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{(NIR-Red)}{(NIR+Red)}}$$

(1)

Areas with low NDVI values typically represent conditions with little to no vegetation, like rocky terrains or bare earth. In contrast, moderate NDVI values suggest the presence of grasslands and shrubbery, whereas high values are indicative of dense forests and thriving vegetation [31,32].

2.5.3. Plant Height Estimation

A Corn Height Model (CHM) was generated by calculating the difference between the corn surface (Z_s), taken from the DSM and the bare soil elevation (Z_e), calculated from the DTM as follows:

$$CHM=Z_s-Z_e$$

(2)

The calculation was performed in ArcGIS Pro 3.5 (ESRI, Redlands, CA, USA) using the raster calculator, and results were expressed in meters. To reduce errors and increase accuracy while calculating the CHM, the boundaries of all the field ditches in the study areas were digitized in ArcGIS Pro and extracted from the DTMs and DSMs datasets using the erase tool. To validate the photogrammetric estimate of the CHM, 12 Corn Height Reference (CH_ref) per field were acquired each year. To ensure an unbiased randomized selection of the sampling areas, the sampling points were selected using the inverted W pattern sampling technique [33,34,35].

Sampling areas consisted of a 1 m × 1 m square, oriented parallel to the corn rows. All the plants within the sampling box were measured. Corn fields in the study areas were sown 15 cm apart in rows distanced by 75 cm; therefore, 10 to 12 plants were captured within each sampling area (Figure 5). The geolocation of each sampling area was digitized in ArcGIS Pro and imported to the Emlid Flow (Emlid Inc., Hong Kong, China) application on an IOS device. The Emlid Reach RS2 was then used to navigate to the location of the sampling areas, allowing for the collection of CH_ref measurements. The CH_ref measurements were acquired one week after the second survey (Figure 3).

The plant height was measured at three locations from the ground base: (1) to the base of the flower apex of the stem end, (2) to the first leaf and (3) to the second leaf. Figure 6 offers a visual representation of the different measurements taken from the plant used to generate CH_ref. Measurements were acquired with a four-meter telescoping grade rod, equipped with a spirit level and positioned vertically along the corn stalk.

2.5.4. Corn Height Analysis

The accuracy of the CHM was assessed by comparing it to the CH_ref. The same sampling boxes that were used to acquire CH_ref data were also employed to filter cell values within the CHM. The zonal statistics tool in ArcGIS Pro was used to calculate the 95th percentile of these cell values and estimate plant height per plot. This particular percentile was chosen based on findings from Malambo et al. [36], which indicated that the 90th, 95th, and 99th percentile height metrics from the CHM correlated better with field measurements than the maximum height metric. Tests on the dataset confirmed that the 95th percentile was the most accurate metric.

The degree of agreement between CH_ref and predicted plant heights from the CHM was evaluated using Coefficient of Determination (R²), Lin’s Concordance Correlation Coefficient (ρ_c) and Root Mean Square Error (RMSE). R² was used to assess how well the aerial-based model captures the variability in ground-measured plant heights. A high R² indicates that the model is successful in capturing the relationship between the aerial images and the actual plant heights, whereas a low R² suggests that the model is not accurately representing this relation. However, R² alone does not account for systematic overestimation or underestimation in the predictions. It measures only the strength of the linear relationship without considering how closely the predicted values align with the 1:1 line of perfect agreement.

Therefore, the values of ρ_c were also employed since it combines measures of both precision and accuracy to evaluate how well the predicted plant heights agree with the CH_ref values along the line of perfect concordance (1:1 line). Lin’s concordance correlation coefficient assesses the degree to which pairs of observations fall on the 1:1 line, accounting for any systematic deviations from this line [37]. It includes a Bias Correction Factor (C_b) that adjusts for over or underestimation, ensuring that both the slope and location differences between the predicted and observed values are considered [38]. This is particularly important in this analysis since any consistent underestimation or overestimation could indicate a bias in the model predictions [37]. An accurate model should have predictions that align closely with the 1:1 line, indicating both high correlation and agreement. A high ρ_c value suggests that the model not only captures the variability in plant heights (as indicated by R²) but also provides unbiased predictions that closely match the CH_ref values [39].

Lasty, the RMSE was used to quantify the average error in the plant height predictions. Lower RMSE values indicate that the height of corn plants predicted from aerial images is very close to the ground measurements, indicating high model accuracy. Calculations of both R² and RMSE were made using MATLAB (v2024) (MathWorks, Inc., Natick, MA, USA) while ρ_c values were calculated in IBM SPSS Statistics (v29) (IBM Corp., Armonk, NY, USA). These statistical measures were calculated using the following formulas:

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{\left(y_i-\widehat{y_i}\right)}^2}{{\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}}$$

(3)

$${\rho }_c={\displaystyle \frac{2\rho {\sigma }_x{\sigma }_y}{{\sigma }_x^2+{\sigma }_y^2+{\left({\mu }_x-{\mu }_y\right)}^2}}$$

(4)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\left(y_i-\widehat{y_i}\right)}^2}$$

(5)

2.5.5. Modelling Flooded Zones

A flood simulation technique called the Arc-Malstrøm hydrologic screening method [40] was used to detect the extent and depth of landscape depressions, termed “bluespots”, within the fields. This method assumes Hortonian flow conditions, identifying bluespots and categorizing the surrounding surface into their respective watersheds [41]. The model identifies the discharge points of each bluespots and delineates the interconnecting streams that facilitate overflow [40].

The method calculates the runoff volume for each watershed (termed RainVolume) during a uniform rainfall scenario across the basin. If the capacity of any bluespot is exceeded, it results in a spillage (SpillOverOut) at the pour point, initiating a water flow path downstream (Figure 7). Additionally, any bluespot receiving overflow from upstream will count as a SpillOverIn, helping estimate the water balance for each bluespot. The overall downstream overflow volumes are compiled from these individual water balances during a specified rainfall event [42].

The Arc-Malstrom method was used on the DTMs from the study areas and executed using ArcGIS Pro. The model excluded bluespots > 5 cm and > 1 m³ to respect the accuracy limit of the DTMs. The model was tested against a uniform precipitation scenario of 25 mm. This threshold was selected due to the frequency of these rain events, as recorded by the closest weather station to the study sites Truro (Debert) and Grand-Pré (Kentville; Figure 8).

Due to the presence of dense vegetation, such as bushes, tall grasses, or small trees along some of the ditches, the DTMs were hydrologically conditioned by flattening the ditches to a depth of 91 cm, using the surrounding terrain’s elevation [43]. By utilizing a Python (v3.11.8) script in the ArcGIS API environment, the elevation values along the ditch lines were systematically lowered to achieve this specified depth, effectively modifying the DTMs for hydrological analysis. This depth was chosen based on field measurements using the GNSS receivers, common practices for ditch depth on dykelands, as well as guidelines from the literature [2,3,8].

The DTMs were not hydro-enforced when simulating water flow through the culverts that separated the ditches. The goal was to assess the impact of clogged or poorly functioning drains on the fields and to identify areas where water accumulated during a specific rainfall scenario. Additionally, to achieve this objective, the main collector drain for the fields was blocked to prevent water from exiting the watershed of the field. These decisions were aimed at simulating a moderately common rainfall event and evaluate the response of the drainage system under such conditions. Figure 9 illustrates the original and hydro-flattened elevation models resulting from these alterations, which were used to compute the flood model.

2.5.6. Flood Risk and Dry Zones

Areas of the fields located within the flood zones were classified as “flood risk,” while the remaining areas were classified as “dry” zones. As noted in Section 2.5.5, thresholds for the flood simulation and the definition of flood risk zones were based on a 25 mm precipitation scenario and a minimum depression depth of 5 cm. These parameters reflect both the frequency of local rainfall events (Figure 8) and the vertical accuracy limits of the DTMs, in accordance with the Arc-Maelström guidelines [40,42].

To avoid misrepresenting the different zones, areas with planting gaps caused by planter malfunctions were manually digitized and excluded from the analysis by changing the raster values on the CHM to null. The same approach was applied to the gap typically found between the edge of the field and the first row of corn. If not removed, these areas, often composed of grass or bare ground, could misrepresent the height of the corn plants in these areas.

To statistically assess the observed relationship between flood risk zone size and mean plant height, a Spearman correlation analysis ($\rho )$ and R² were calculated. The Spearman correlation was selected due to its suitability for analyzing relationships in non-normally distributed datasets and small sample sizes, which aligns with the nature of the data in this study. The following formula was used to calculate the Spearman correlation:

$$\rho =1-{\displaystyle \frac{6\sum d_i^2}{n(n^2-1)}}$$

(6)

2.5.7. Plant Height Data and NDVI Analysis

Flood risk zones generated by the hydrologic model were exported as vector polygons and overlaid on the CHM. The cell values beneath these zones were then extracted and compared to those in the rest of the fields to evaluate dry areas. For each year and field, the cell values on the CHM and the size of each zone were calculated and compiled into a table. A similar approach was applied to the NDVI maps, where the same vector polygons of flood risk zones were used to extract values from the NDVI data. These tasks were automated using the ArcGIS API for Python.

3. Results

3.1. Plant Height Model

Figure 10 displays the results from the CHM compared to the CH_ref. The trend in R² values from 2021 to 2023 indicates good model performance, with regression lines closely fitting the data points and explaining a substantial part of the variance in corn plant heights. R² values ranged between 0.62 and 0.82 between the three years period. Additionally, the RMSE ranged between 0.224 m and 0.333 m, indicating that the model accurately predicted corn plant heights with reasonable precision, although accuracy slightly decreased in 2023.

Values of ρ_c increased from 0.784 in 2021 (95% CI: 0.694–0.850) to 0.882 in 2022 (95% CI: 0.823–0.922) and remained high at 0.883 in 2023 (95% CI: 0.823–0.924), indicating an overall improvement in the model accuracy over the years. The bias correction factors were 0.989, 0.974, and 0.988 in 2021, 2022, and 2023, respectively, suggesting minimal systematic bias between the measurement methods each year.

3.2. Surface Drainage Analysis

3.2.1. Changes in the Size of the Flood Risk Zones

During the study period, Fields A and B exhibited a decreasing trend in the size of flood risk zones (

Table 8. Comparison of plant height and NDVI values across flood risk and dry zones in Fields A, B, C, and D from 2021 to 2023.

Flood Risk Zones							Dry Zones
			Plant H (m)		NDVI			Plant H (m)		NDVI
Field	Year	Area (%)	$\overline{\mathbf{x}}$	σ	$\overline{\mathbf{x}}$	σ	Area (%)	$\overline{\mathbf{x}}$	σ	$\overline{\mathbf{x}}$	σ
A	2021	16	2.13	0.29	0.58	0.08	84	2.23	0.23	0.57	0.08
	2022	6	2.38	0.49	0.68	0.08	94	2.54	0.29	0.68	0.07
	2023	N/A	N/A	N/A	0.76	0.10	N/A	N/A	N/A	0.80	0.06
B	2021	25	1.50	0.53	0.63	0.09	74	1.58	0.46	0.60	0.09
	2022	N/A	N/A	N/A	0.70	0.13	N/A	N/A	N/A	0.74	0.09
	2023	18	1.90	0.51	0.67	0.08	82	2.13	0.34	0.67	0.07
C	2021	37	1.38	0.50	0.71	0.09	63	1.99	0.35	0.69	0.08
	2022	37	1.43	0.53	0.65	0.14	63	1.79	0.37	0.68	0.10
	2023	61	0.26	0.42	0.46	0.14	39	1.07	0.66	0.55	0.13
D	2021	21	2.19	0.55	0.64	0.14	79	2.47	0.31	0.66	0.09
	2022	25	1.77	0.56	0.64	0.14	75	2.17	0.26	0.68	0.10
	2023	35	1.11	0.64	0.62	0.11	65	1.74	0.54	0.67	0.07

). In 2021, 16% of Field A was classified as a flood risk zone, which decreased to 5% the following year. In contrast, Fields C and D showed an increasing trend in the size of the flood risk zones. Notably, the proportion of land at risk in Field C increased from 37% in 2022 to 61% of the field in 2023.

3.2.2. Mean Plant Height in Flood Risk Zones and Dry Areas

The analysis of plant height across the four fields over three years under a water-logged scenario revealed important differences between flood risk and dry zones. Dry areas consistently exhibited greater mean plant height than flood risk zones across all fields. In Field A, the mean plant height in dry areas increased from 2.23 m in 2021 to 2.54 m in 2022 and from 1.58 m in 2021 to 2.13 m in 2023 for Field B (Table 8). A similar trend was observed in Fields C and D, where plant height in dry areas remained higher than flood risk zones despite a noticeable overall decrease in the mean plant height from one year to the other, which indicated worsening growing conditions (Figure 11).

Conversely, flood risk areas often displayed more variability and a trend of declining plant height over time. This was particularly evident in Field C, where the mean plant height in flood-risk areas dropped sharply from 1.43 m in 2022 to 0.26 m in 2023. These results indicated that the productivity classification strongly influenced plant height, with dry areas generally supporting taller plants as shown in Figure A1 and Figure A2 of Appendix A.

The standard deviation analysis in mean plant height across fields also revealed interesting patterns, particularly in 2023, where several fields, especially Field C and Field D, exhibited increased variability in plant height, indicating less uniform growth (Figure A3 and Figure A4). Dry zones generally maintained lower and more stable standard deviations, reflecting more consistent plant growth. However, Field C was an exception in 2023, with a significant spike in variability in both productivity areas, suggesting unusual factors affecting plant growth that year.

Lastly, results from the Spearman’s rho indicated a statistically significant negative correlation between the size of the flood zone and plant height ($\rho$ = –0.921, p < 0.01), confirming that increases in flood-prone areas are associated with decreases in plant height (Figure 12). A linear regression model further supported this finding (R² = 0.923, p < 0.01), with each 10% increase in the flooded area linked to an approximate decrease of 0.4 m in mean plant height. In contrast, the analysis of drier zones revealed a strong positive correlation ($\rho$ = 0.827, R² = 0.775), suggesting that plant height tends to increase with decreasing flood extent.

3.2.3. NDVI Values in Flood Risk Zones and Dry Areas

The NDVI values provided insights into the differences between flood risk and dry areas, although these differences were more nuanced than initially expected. Dry areas generally had values similar to flood risk zones in Fields A and B, while dry areas had slightly higher NDVI values than those in Fields C and D. Overall, NDVI values in fields located in Truro (A & B) and Grand-Pré (C & D) followed a similar pattern to the mean plant height, where values progressively increased over the years in Truro and decreased in Grand-Pré. For example, in Field A, NDVI in dry areas increased from 0.57 in 2021 to 0.68 in 2022, while Field C showed a decrease in mean NDVI values in dry areas, ranging from 0.69 to 0.55 over the three-year period, which correlated with a decrease in mean plant height. This field also had the highest number of short plants and bare ground spots, which negatively impacted NDVI values.

4. Discussion

The results from the UAV-derived CHM demonstrated that the model provided accurate and reliable estimates of corn plant heights. The combination of high R², high ρ_c, and low RMSE values were all strong indicators of the performance of the CHM. R² and RMSE results were consistent with those found in the literature on similar studies [11,36,44,45]. While the ρ_c values demonstrated moderate to strong agreement with ground-measured plant heights it is important to acknowledge some of the inherent limitations of the ρ_c metric [46,47]. Specifically, the ρ_c conflates both precision and bias, making it challenging to distinguish whether the observed agreement stems from high correlation or low bias. For instance, although the ρ_c values suggest improved model accuracy in 2022 and 2023, the corresponding increase in RMSE highlights a slight decline in predictive accuracy in 2023, possibly indicating systematic errors [47].

This decline can be attributed to challenges specific to the growing season of 2023, particularly related to the prevalence of shorter corn plants. Issues with canopy cover and gaps could have created sparser and uneven canopy cover, resulting from shorter plants, which may have led to underestimations of height. In this case, drone imagery would have captured more ground surface than the canopy top. Additionally, less reliable field measurements for short plants are likely to introduce errors, as small variations in measurement can produce larger relative inaccuracies compared to taller plants since RMSE is an absolute measure. For instance, A 10 cm error on a 20 cm plant is a 50% relative error, while the same error on a 100 cm plant is only 10%.

Moreover, since ρ_c is sensitive to the variance of the data, the higher values observed in later years may reflect changes in the variability of the measured plant heights rather than true improvements in model performance. Therefore, while ρ_c provides a useful indicator of model agreement, it should be interpreted alongside other metrics such as RMSE and R² to obtain a more nuanced understanding of model precision and bias [47].

The results of the study also revealed a relationship between the size of flood risk areas and mean plant height, with Fields C and D showing a negative association and Fields A and B showing a positive one. The reduction in the size of the flood risk zones in Fields A and B can be attributed to the proactive maintenance efforts by the landowner, who addressed minor drainage issues annually (Table 2). On the other hand, the increasing trends in the size of the flood risk zone for Fields C and D resulted from the minimal slope on land, which hindered adequate surface drainage and the absence of maintenance work on these fields (Table 2). In these conditions, even minor changes in soil surface caused by farming operations lead to significant variations in yearly productivity. Overall, these findings emphasize that while dry zones typically showed more uniformity, certain years and fields experienced greater inconsistency in plant height, especially in flood-risk zones.

Lastly, results showed that the relationship between NDVI and plant health was inconsistent across all fields. In Field D, NDVI values remained stable throughout the years, which suggested that while NDVI helped identify areas with very short plants and bare spots, particularly in flood risk zones, it was less effective in consistently reflecting subtle differences in plant health that could significantly impact yields. In contrast, Fields A, B, and D exhibited more stable NDVI values across low and high-productivity areas, regardless of plant height, indicating that NDVI alone may not fully capture the variability in plant health across different conditions.

The observed inconsistencies between NDVI values and plant height may have stemmed from the sensitivity of the NDVI to vegetation greenness and canopy density, rather than plant physiological stress directly related to waterlogging. Indeed, corn planted in the flooded areas might have been shorter than those in the rest of the field due to waterlogging stress in early spring. If water had accumulated near the ditches, it could have created anaerobic soil conditions, impairing root development and nutrient uptake during the vegetative stages of corn [48,49]. Consequently, the affected plants may have remained shorter, though they might have survived due to the temporary and partial nature of the flooding [48,50]. This could have influenced plant height more than leaf greenness, potentially explaining inconsistencies between NDVI values and plant height. Since NDVI primarily reflects leaf health and canopy cover rather than height, plants in flooded areas might have appeared healthy on NDVI maps despite being less productive overall.

4.1. Other Remote Sensing-Based Approaches

While various remote sensing technologies have been employed for drainage assessment in agriculture, this study demonstrates that drone-based photogrammetry offers a particularly effective balance between accuracy, spatial resolution, and operational feasibility. Compared to UAV-LiDAR, which provides high-resolution elevation data suitable for detailed topographic analysis [18], photogrammetry is significantly more cost-effective and accessible to end-users with limited technical expertise. Although UAVs equipped with LiDAR sensors may perform better in areas with dense vegetation, the higher operational costs and complexity of these systems may limit their adoption for routine field monitoring. In contrast, satellite-based multispectral imagery has been used to detect moisture stress over larger areas [15,16]. However, the coarse spatial resolution of these datasets limits their ability to detect subtle micro-topographic changes that influence surface drainage, especially in flat terrains like the dykelands.

Non-remote sensing methods such as ground-based LiDAR and soil moisture sensors also offer valuable insights but face scalability issues. Ground-based LiDAR systems can capture highly accurate elevation data but are time-intensive and expensive to deploy across large agricultural fields [51]. Similarly, soil moisture sensors provide precise point-based measurements of field conditions but lack the spatial coverage necessary to inform field-scale decisions [52]. Consequently, while these tools may serve well in validation or targeted monitoring roles, they are less suited for proactive, large-area drainage assessment.

4.2. Field Management Practices & Application to Other Sectors

Differences in field management practices, such as annual ditch maintenance and periodic recrowning, substantially influenced drainage performance and productivity across the fields. Specifically, Fields A and B, characterized by yearly ditch clearing, careful soil management to reduce compaction, and the use of grassed waterways, exhibited improved drainage efficiency and higher productivity. Conversely, Fields C and D, lacking regular ditch maintenance and recrowning over the past decade, experienced increasing flood risk zones and reduced crop productivity.

Beyond dykelands, the techniques developed in this study for assessing and improving surface drainage have potential applications in other sectors facing similar challenges. For example, the peat moss industry in Canada often struggles with surface drainage issues [53]. Peatlands used for peat moss extraction must be drained to improve agricultural productivity and support heavy machinery [54]. However, identifying problematic areas for drainage remains a significant challenge. The solution presented in this study could offer a cost-effective alternative for the peat moss industry. Adapting these methodologies could enable peat producers to better identify areas with poor drainage and implement targeted interventions.

4.3. Limitations & Recommendations

One limitation of this study is the assumption that shorter corn plants correlate with reduced yields. While there is a generally positive correlation between plant height and yield, particularly in later growth stages, as demonstrated by Yin et al. [55] and Kelly et al. [56], this relationship is not universally consistent. Factors, such as genetics and environmental conditions, can influence yield—as demonstrated by Suazo et al. [57]. However, Sammis et al. [58] suggested that corn plant height was a good indicator of water stress levels, which aligns with the findings of this study. These complexities underscore the need to consider multiple factors when assessing plant growth and yield outcomes.

Building on these considerations, the hydrological modelling approach is another key limitation. While the Arc-Malstrøm method efficiently identifies potential flood-prone areas, its reliance on simplified Hortonian flow assumptions and lack of infiltration modelling may limit its accuracy, particularly on soils with variable permeability. Alternative models that incorporate infiltration and evapotranspiration processes, such as physically based hydrological models, could enhance predictions but require significantly more data inputs and computational resources [59]. Moreover, the Arc-Malstrøm approach is most effective in fields with a relatively flat topography. Indeed, if a field has significant elevation differences, the model may inaccurately represent surface drainage by pooling all water into the lowest elevation. Segmenting the elevation data into smaller sections could enhance accuracy, especially when farmers have previously identified waterlogged and damaged drains.

A third limitation is the limited representability of the fields examined, as dykelands in Nova Scotia are vast, and soil type and field layout across different regions can vary. Such diversity in field characteristics makes it essential to adapt the method before generalizing its effectiveness on all dykelands in Atlantic Canada.

Climatic variations also play an important role in surface water accumulation. Although Bootsma et al. [60] showed that Crop Heat Units (CHU) differ only slightly between Truro (2500 CHU) and Grand-Pré (2700 CHU), even small shifts in temperature can affect crop growth [61]. More importantly, variability in rainfall distribution and intensity between years may directly impact water accumulation (Figure 8). These differences potentially affect seasonal flooding and water retention, which could influence the results of any surface drainage assessment.

Despite these limitations, several practical recommendations emerge from this study.

Based on the findings, farmers are encouraged to proactively conduct annual drone-based assessments to identify areas prone to drainage degradation. These aerial surveys could help locate water-holding depressions caused by soil settlement, allowing targeted land levelling and filling.

In addition to standard drainage maintenance practices, such as removing debris from open ditches, mowing grassed waterways, and maintaining adequate side slopes to prevent erosion, farmers should also focus on minimizing siltation. One effective approach is preserving a three-meter-wide grass buffer strip along ditches [62]. Regularly inspecting and repairing damaged drainage lines could also ensure efficient water movement [3]. Furthermore, farmers should prioritize targeted annual ditch cleaning to prevent blockages. These measures, guided by drone-derived insights, could help mitigate waterlogging and improve yields.

5. Conclusions

This study addressed the seasonal degradation of surface drainage on Nova Scotia’s agricultural dykelands using drone-derived elevation models and vegetation indices. By evaluating changes in field topography and crop development from 2021 to 2023, the results demonstrated that drone-based photogrammetry and multispectral imagery offer an effective means of identifying poorly drained areas that may not be visible through traditional methods. The findings revealed a strong negative correlation between the extent of flood risk zones and mean plant height, supporting the use of plant height as a useful indicator of drainage performance. While NDVI values provided additional context, they were less consistent than height measurements in capturing productivity declines in flood-prone areas.

Despite the demonstrated potential of this approach, several limitations should be acknowledged. First, although plant height is a commonly used proxy for productivity, direct yield measurements were not included in this study. While statistical analyses confirmed a significant relationship between plant height and flood zone extent, future work should seek to validate these findings against actual yield data. Second, the flood modelling approach used in this study, while efficient and practical for field-scale applications, is based on simplified hydrologic assumptions and does not account for infiltration or subsurface processes. As a result, flood extent predictions may be influenced by factors such as soil type and rainfall intensity, which were not explicitly modelled.

In addition, the methodology was applied to a limited number of fields within Nova Scotia, and caution should be exercised before generalizing the results to other regions without further validation. Climatic and topographic differences may influence both drainage behavior and the effectiveness of drone-based assessment techniques. Future studies should explore the scalability of this approach across diverse agricultural regions and under varying environmental conditions. Integrating additional remote sensing metrics, such as thermal imagery, could also improve predictions of crop stress.

Lastly, while the methods used in this study relied on commercial software and technical expertise, future work should prioritize the development of user-friendly tools that can be more readily adopted by farmers. A simplified application that allows users to upload a DTM and automatically generate drainage maps could help translate this research into practical on-farm decision-making tools. Improving the accessibility of these technologies may support more proactive drainage management strategies.

Overall, this research highlights the practical value of integrating drone technologies with geospatial analysis to support improved field management on agricultural dykelands. By identifying early signs of drainage degradation and linking them to crop development patterns, this approach can assist farmers in determining when and where recrowning or drainage maintenance may be required, ultimately contributing to improved productivity and long-term field performance.