Terrain Matters: A Focus+Context Visualization Approach for Landform-Based Remote Sensing Analysis of Agricultural Performance

Highlights

• What is proposed? A focus+context visual analytics tool that effectively facilitates analysis of multiple data layers, including terrain features and field-scale agricultural performance, extracted from remote sensing imagery and elevation models.
• What is found? Relative to depressions and hilltops, baseline areas, which comprise the remainder of the field, demonstrate superior performance and reduced variability.
• What does the tool offer? A practical approach to investigating the relationships between agricultural field performance and terrain features.
• What is the implication? Integrating terrain features into performance analysis supports management zone refinement and more informed agricultural decision-making.

Abstract

Understanding spatial variability is central to precision agriculture, yet terrain features are often overlooked in remote sensing workflows that inform agronomic decision-making. This work introduces a terrain-aware visual analytics approach that integrates landform classification with crop performance analysis to better support field-level decisions. Terrain features are an important contributor to yield variability, alongside environmental conditions, soil properties, and management practices. However, they are rarely integrated systematically into performance analysis and decision-making workflows—limiting the potential for terrain-aware insights in precision agriculture. Addressing this gap requires approaches that incorporate terrain attributes and landform classifications into agricultural performance analysis and management zone (MZ) delineation—ideally through visual analytics that offer interpretable insights beyond the constraints of purely data-driven methods. We introduce an interactive focus+context visualization tool that integrates multiple data layers—including terrain features, vegetation index–based performance metric, and management zones—into a unified, expressive view. The system leverages freely available remote sensing imagery and terrain data derived from Digital Elevation Models (DEMs) to evaluate crop performance and landform characteristics in support of agronomic analysis. The tool was applied to eleven agricultural fields across the Canadian Prairies under diverse environmental conditions. Fields were segmented into depressions, hilltops, and baseline areas, and crop performance was evaluated across these landform groups using the system’s interactive visualization and analytics. Depressions and hilltops consistently showed lower mean performance and higher variability (measured by coefficient of variation) compared to baseline regions, which covered an average of 82% of each field. We also subdivided baseline areas using slope and the Sediment Transport Index (STI) to investigate soil erosion effects, but field-level patterns were inconsistent and no systematic differences emerged across all sites. Expert evaluation confirmed the tool’s usability and its value for field-level decision support. Overall, the method enhances terrain-aware interpretation of remotely sensed data and contributes meaningfully to refining management zone delineation in precision agriculture.

1. Introduction

The delineation of “Management Zones (MZs)” is a key component of precision agriculture, aiming to divide fields into subregions that can be managed based on localized biophysical and agronomic conditions [1]. Most delineation approaches rely on soil potential and productivity metrics such as crop yield or “Vegetation Indices (VIs)” [2]. However, productivity potential is influenced by a complex interplay of factors, including weather variability, precipitation, farm management practices, soil properties, and topographic features. Some of these factors, such as fertilizer deficiencies, are manageable, whereas others, like extreme weather events, are beyond the control of agronomists. Moreover, certain factors directly influence crop growth (e.g., water availability), while others act indirectly (e.g., terrain features affecting water retention) [3]. Isolating the individual contribution of each variable is methodologically challenging [4], as it requires controlling for other influences, which is an impractical condition in real-world field environments. As a result, while indicators like yield or VIs are frequently used for delineating MZs, they reflect aggregate outcomes and often obscure the underlying causes of spatial variability. Integrating terrain features into MZ delineation can help unravel this complexity for topographic features and support more topography-informed, site-specific management strategies [5]. For instance, two zones with similar historical yields may respond differently to inputs if one lies in a depression and the other on an elevated slope, owing to differences in water and nutrient redistribution patterns [6].

Given the necessity of incorporating terrain features into management decisions [5,7], the challenge lies in effectively integrating topographic features into the delineation process of MZs. This challenge raises a key scientific question: How can terrain features be systematically incorporated into agricultural performance analysis and MZ delineation? With the increasing availability of remote sensing data [8,9], various biophysical indicators, such as VIs, growth trends, and surface reflectance, can be utilized to enhance agricultural decision-making [10]. Among the remotely sensed data, “Digital Elevation Models (DEMs)” stand out as a valuable resource for terrain characterization. DEMs are accessible through global datasets, such as SRTM, AW3D30, ASTER GDEM3, and TanDEM-X [11], or high-resolution national products, like HRDEM [12].

Several terrain features can be derived from DEMs, including slope, aspect, curvature measures, “Topographic Position Index (TPI)”, “Topographic Wetness Index (TWI)”, “Sediment Transport Index (STI)”, flow accumulation, and depressions, using tools such as Whitebox [13]. However, the large number of available terrain features [5] introduces the feature selection challenge, determining which features are most relevant to zone delineation. A common approach to address this challenge is dimensionality reduction, such as “Principal Component Analysis (PCA)” [14]. Alternatively, machine learning approaches implicitly learn complex patterns from all available features [15,16]. However, these data-driven techniques often lack interpretability [17] and are highly sensitive to context-specific factors, particularly at the field scale, where localized soil, management history, and topographic interactions can vary substantially [18,19].

To overcome these limitations, integrating human expertise into the analytical process is crucial. A human-in-the-loop approach, supported by expressive visual analytics, allows agronomists and decision-makers to interactively explore terrain features, validate model outputs, and incorporate domain knowledge into decision-making. By leveraging intuitive visual representations, experts can identify meaningful patterns within the initial set of features, whether AI-driven or manually selected, ensuring the most effective refinement of the zonemap to optimize agricultural insights.

Therefore, we propose a novel visual analytics approach that utilizes focus+context visualization, a technique that integrates multiple close-ups of terrain features, dynamically linked to performance metrics. This enables an interactive exploration of the relationships between terrain features and agricultural productivity within the zonemap framework, fostering deeper understanding and more informed decision-making (See Figure 1).

Focus+context visualization has proven effective for managing occlusion, integrating heterogeneous datasets, and supporting scalable, hierarchical exploration in large spatial domains [20,21].

To isolate the effects of terrain on field performance from management practices, we focused on fields with consistent fertilizer application. We applied our visual analytics system to eleven agricultural fields across Canada, which were managed under constant-rate fertilizer regimes during selected years, including two irrigated sites. The system integrates three core datasets: peak green as a proxy for field performance; productivity zonemap; and elevation model (see Figure 2). Performance indicators were derived using the peak green approach [22,23,24], which extracts “Normalized Difference Vegetation Index (NDVI)” images corresponding to maximum vegetative growth (e.g., see Figure 3). MZs were delineated using our prior automated method [23,24]. Terrain features were computed from the TanDEM-X EDEM product [25,26,27,28] and from the higher-resolution HRDEM dataset, where available [12]. WhiteboxTools [13,29] were used to derive terrain features. The pipeline is explained in detail in Section 3.

Using our interactive visualization system with a human-in-the-loop approach, we observed that depressions and hilltops tend to host lower mean performance and higher variability, as measured by the “Coefficient of Variation (CV)”. The greater variability in performance values suggests that such regions are more sensitive to microtopographic and hydrological influences. Across all 11 experimental fields, depressed areas, though covering less than 10% of the field on average, consistently exhibited the lowest mean performance values. Hilltops ranked second lowest in terms of mean performance. In contrast, the remaining areas, hereinafter referred to as baseline, which comprise approximately 82% of the field’s area and do not fall within depressions or hilltops, demonstrated more stable and higher mean performance. These findings underscore the importance of incorporating terrain features and landform classes into spatial performance analysis and field management decisions. Our tool reveals localized patterns that are often obscured when using field-scale aggregate statistics (See Figure 4).

While baseline regions generally appeared to have higher performance and lower variability, they are not uniformly flat. In many fields, baseline areas encompass sloped cropland that may be susceptible to soil erosion. Previous studies have shown that topographic features such as slope, curvature, and elevation strongly influence soil erosion, soil moisture distribution, nutrient cycling, and microclimatic conditions, thereby affecting crop growth and yield [19,30,31]. Moreover, terrain-derived indices such as the STI and TWI provide complementary information on detachment, transport, deposition, and hydrologic pathways, and are widely used to identify erosion-prone positions in agricultural lands [5,7]. However, relationships between terrain metrics and crop yield are highly site- and climate-dependent—for instance, upslope length and curvature correlate more strongly with wheat yield in dry years than in wet years [18]. Collectively, these insights highlight both the promise and the limits of terrain features for guiding management decisions [4], underscoring the need for tools that allow local, case-specific examination of terrain–performance interactions.

In light of these considerations, we further analyzed baseline areas to capture their internal variability. Hence, two complementary refinements were applied to baseline: slope-thresholding to distinguish relatively flat farmland from sloped cropland, and STI-based subdivision into three erosion-sensitivity classes. Results revealed only modest differences between slope- or STI-based subclasses, with STI subdivisions capturing some field-specific heterogeneity but showing no consistent performance gradients across all fields. These outcomes reflect the context-dependent nature of the erosion-related indices examined in this analysis (i.e., slope, STI).

The system was evaluated by a domain expert, whose feedback helped assess its usability and applicability in real-world agronomic workflows (see Section 4.1). Their feedback confirmed the practical utility of the tool in identifying terrain-driven performance variability, particularly in challenging or topographically complex areas. Based on the expert’s input, we have refined our focus+context approach to better align with domain-specific knowledge and feedback, enhancing usability and interpretability. This final design is discussed in Section 3.5.

Overall, our paper presents several key contributions that advance terrain-aware agricultural analysis:

• Introducing a novel focus+context visual analytics tool to support terrain-aware interpretation of crop performance patterns.
• Classifying agricultural land into depressions, hilltops, and baseline using terrain-derived indices, and quantifying and comparing performance across different classes.
• Demonstrate, through spatial analysis, that baseline areas consistently exhibit higher crop performance with lower variance compared to hilltops and depressions.
• Incorporating expert feedback to refine the focus+context approach.

To provide a more comprehensive and structured review of the literature, we have chosen to separate the related work from the Introduction. This allows for a more detailed examination of a broader range of relevant studies, improving the clarity and organization of the paper.

The remainder of the paper is organized as follows: Section 2 reviews relevant literature, Section 3 outlines the methodology and case studies, Section 4 presents experimental results and discussion, and Section 5 concludes the paper and discusses limitations of our work and future research directions.

2. Related Work

Topography has long been recognized as one of the key determinants of crop productivity, directly or indirectly influencing crop growth in agricultural fields [5,32]. This section reviews prior research in four key areas: (1) DEMs and terrain topography, (2) integration of topographic features in agricultural productivity analysis and MZ delineation, (3) landform classification, and (4) visualization approaches.

2.1. Digital Elevation Models (DEMs) and Terrain Topography

DEMs are grid-based datasets of elevation that are essential for capturing the terrain topography. DEMs are widely used across disciplines such as geomorphology, hydrology, and agriculture [33,34]. Mathematically, a topographic surface of the terrain can be described as a continuous function $z=f(x,y)$, where elevation z varies smoothly over the spatial coordinates x and y [34]. This representation allows DEMs to model land surface variations in a structured, two-dimensional raster format and enables an easier computation of terrain features (e.g., slope, curvature, and TWI). These terrain metrics form the basis of most terrain-based analyses in environmental and agricultural studies.

Several publicly available DEM products with a wide spatial range are commonly used in terrain modeling. For instance, the Shuttle Radar Topography Mission (SRTM) is a global DEM available in 3-arc-second (∼90 m) and 1-arc-second (∼30 m) resolutions [35]. The data is acquired via radar interferometry with 16 m spatial and less than 9 m vertical accuracy [36]. TanDEM-X, produced by the German Aerospace Center (DLR) using bistatic radar systems [26], and offers improved vertical accuracy and coverage consistency [28]. The TanDEM-X product is available in 3-arc-second and 1-arc-second resolutions with improved vertical accuracy of ∼2 m in flat areas [26].

Canada’s High-Resolution Digital Elevation Model (HRDEM), produced by Natural Resources Canada (NRCan) and is derived from Light Detection Ranging (LiDAR) or stereo imagery, offering spatial resolutions of 1–2 m [12,37] with less than one meter vertical accuracy [38]. Therefore, it provides detailed topographic information ideal for field-level analysis; however, its spatial coverage is limited, and only three of our eleven study fields are located within areas where HRDEM is available.

HRDEM includes both Digital Terrain Models (DTMs), which represent the bare-earth surface, and Digital Surface Models (DSMs), which capture vegetation, buildings, and other surface features [37,38]. For terrain-based analyses such as slope, curvature, and hydrological flow, the DTM product is typically used to more accurately reflect ground topography.

2.2. Topography in Productivity Analysis and MZ Delineation

Topographic features shape hydrological and erosional processes, which in turn modulate soil moisture, nutrient availability, and plant stress factors affecting crop yield [39,40]. To investigate the relationship between topography and productivity, we need to quantify productivity. Researchers have employed various methods for this purpose. For instance, a study in southern Ontario demonstrated that topographic features explained up to 80% of corn yield variation within a field [41]. They used a yield monitor on the combine harvester to assess the yield amount. While this approach for yield assessment is informative, it requires specialized equipment and extensive post-processing to address sensor errors and harvesting artifacts [42]. As a more scalable alternative, VIs such as NDVI offer practical proxies for estimating yield [43]. For example, Basnyat et al. [44] applied NDVI clustering to delineate productivity zones in a field located in the Canadian Prairies, demonstrating its effectiveness in capturing spatial yield variability. Thus, spatial patterns in NDVI can serve as an indirect yet informative indicator of crop productivity and performance.

Landforms influence productivity by shaping moisture and nutrient distribution [45]. Depressions or concave zones tend to collect and hold water and nutrients, whereas raised or convex surfaces encourage water runoff and increase soil erosion [5]. In a two-year study of farmed prairie potholes in Iowa, Martin et al. [6] demonstrated that cropped depressions can become nutrient sinks, and waterlogging may stress plants. However, the effect of these features on productivity is not uniform, and it can vary widely depending on other factors such as weather conditions, drainage systems, and farming practices. For example, Clare et al. [46] and Boldt et al. [47] demonstrated that field performance in small depressions in Canadian Prairies is unstable and generally lower than in uplands, even in drained wetlands. These studies highlight that while moisture accumulation can benefit crops in arid settings, it sometimes causes stress or delayed planting in poorly drained basins.

Topography impacts crop productivity, making it a valuable input for delineating productivity zones. More commonly, MZs are delineated using spatial variability in crop performance, inferred from yield maps, soil electrical conductivity measurements, or VIs such as NDVI [48]. However, some studies have sought to integrate terrain features directly into the delineation of MZs. For instance, Fraisse et al. [49] utilized an unsupervised classification for defining soil MZs considering topographic features, and the authors in [50] applied fuzzy clustering to elevation and soil data to delineate site-specific zones in Egypt. More recently, Toth et al. [7] demonstrated that elevation-derived zones outperformed traditional soil survey maps in flood-prone areas. In parallel, work in terraced agricultural systems in Mexico has demonstrated that DEM-derived terrain indices, such as STI, TWI, and ”Stream Power Index (SPI)”, quantify water redistribution, erosive power, and sediment-transport potential, thereby helping to identify erosion-prone positions and interpret hydrologic behavior relevant to crop performance [51].

These studies highlight the value of terrain features in productivity MZ delineation but often rely on abstract indices rather than explicit landforms. In this paper, using our visualization tool, we demonstrate how landform classes, such as depressions and hilltops, can be utilized to identify low-performing regions to enhance interpretable decisions.

2.3. Landform Classification

Landform classification has long been a foundational aspect of terrain analysis, with applications in geomorphology, hydrology, ecology, and increasingly, precision agriculture. Classical frameworks such as Hammond’s landform system categorize terrain into plains, hills, and mountains based on local relief, slope percentage, and profile type [52]. The following approaches have sought to automate classification using DEMs. For example, the TPI is widely used to distinguish terrain forms such as valleys, slopes, and hilltops based on elevation differences within a defined window [53]. Similarly, the geomorphic method, introduced by Jasiewicz and Stepinski, identifies landform elements such as peaks, pits, ridges, and hollows through the pattern recognition of local terrain geometry [54].

While terms such as hilltop, valley, and depression are commonly used across landform classification frameworks, their precise definitions may vary based on the computational method, scale of analysis, or application domain [55]. In geomorphological taxonomies such as Murphy’s system, terrain is categorized using a combination of structural, topographic, and process-based criteria, including classes like hills and low tablelands, plains, and depressions or basins [55,56]. In applied fields such as agriculture, simplified terrain classes are often adopted to interpret site-specific productivity patterns [45,49].

Our results using the visual analytics tool reinforce the relevance of landform classification in agricultural analysis. Unlike commonly used terrain features such as slope or TWI, landform classes delineate discrete, interpretable zones that facilitate clearer performance analysis. For instance, performance in depressions varies depending on drainage, which is difficult to detect in raw terrain surfaces alone. By adopting a landform-based visual analytics framework, we improve the interpretability of terrain–performance relationships and enable more targeted refinement of MZs. In this paper, we define three primary landform types, hilltops, depressions, and baseline areas, using a hybrid approach detailed in Section 3.2.

2.4. Visualization

Visualization tools in precision agriculture enable agronomists and decision-makers to interpret complex spatial patterns and interactively explore multiple features. Human-in-the-loop visual analytics is particularly important, as it allows agronomists to validate patterns, integrate domain expertise, and refine decision-making processes. For instance, Wachowiak et al. [57] showed that a participatory visual analytics system helped agricultural producers in Northern Ontario explore multi-layered data, including weather, remote sensing imagery, and yield, through an interactive interface tailored for local decision support. Recent scientific efforts have highlighted the value of coordinated multi-view and integrated 3D visualizations for analyzing agricultural data. For instance, Řezník et al. [58] introduced a novel uncertainty index for yield measurements and demonstrated its synchronized visualization with crop yield and terrain using interactive 3D views. While the system demonstrates the utility of combining spatial and agronomic datasets, it tends to prioritize uncertainty modeling or 3D visual realism over terrain-specific agronomic interpretation.

Standard GIS platforms such as ArcGIS Pro [59] and QGIS [60] support side-by-side or linked map views, which can be manually configured to compare crop performance with terrain features. However, these tools require significant setup and are limited in their ability to dynamically explore multiple terrain metrics in concert. Other tools, such as Swipe and Spyglass [61], enable basic two-layer comparisons but are limited to pairwise interactions and lack support for field-specific, multi-layer terrain analysis.

A growing body of research has demonstrated the value of focus+context visualization techniques for handling spatial occlusion, managing cognitive load, and enabling multiscale comparisons in environmental and geospatial applications [20,21]. Related research emphasizes the Overview/Detail paradigm to support scalable exploration in multiscale visualizations [62]. Techniques such as interactive close-up rendering [63] and coordinated multi-view systems [58] allow users to inspect localized patterns while maintaining global context. Inspired by prior focus+context and multiscale visualization techniques, our system emphasizes morphological interpretation (e.g., hilltops and depressions), localized performance variation, and practical terrain–performance diagnostics at the field scale.

Our work introduces a terrain-aware human-in-the-loop visual analytics system tailored to field-scale agricultural decision-making. Unlike existing GIS tools or general 3D viewers, our system synchronizes multiple localized terrain features with crop performance data and supports intuitive inspection of landform-driven variability.

3. Materials and Methods

This paper investigates how topographic features influence performance and how this relationship can inform management decisions on the farm using remote sensing and terrain features. DEMs and VIs are integrated to visualize and evaluate terrain effects on field performance. The proposed system is useful in decision-making in the agricultural fields and refining MZs based on terrain features and landform classification. The methodology section comprises (Section 3.1) study area, data, and system overview; (Section 3.2) terrain features extraction and landform classification; (Section 3.3) field performance, zonemap, and their relationship with terrain features and landform classification; (Section 3.4) data registration; and (Section 3.5) visual analytics tool.

3.1. Study Area, Data, and System Overview

The study area includes eleven agricultural fields located across Alberta, Manitoba, and Saskatchewan in Canada (Figure 5). These fields were selected to represent a range of topographic settings and moisture regimes commonly found across the Canadian Prairies, which contain approximately 80% of Canada’s cropland [64]. Each field was cultivated with a single crop and managed with uniform fertilizer rates (in the studied years) to minimize variability from input management. Two fields were irrigated, further reducing water deficit effects. These controlled conditions allowed the analysis to further isolate the influence of terrain feature variation on crop performance.

Climatic conditions during the study period (2023–2024) exhibited marked inter-annual variability. The 2023 growing season was drought-affected, with below-average precipitation and above-normal temperatures [65,66]. In 2024, while Manitoba, in general, experienced above-average rainfall, data from the Brandon weather station (the closest to the studied fields) indicated below-average summer precipitation. Northern Alberta was wetter than average, whereas southern Alberta remained drier. Overall, 2024 temperatures were close to the long-term average [67].

Data. Elevation data are sourced from the TanDEM-X Enhanced Digital Elevation Model (EDEM) (available at: https://download.geoservice.dlr.de/TDM30_EDEM; accessed on 3 February 2025), providing global 30 m spatial resolution with improved vertical accuracy for eight fields [25,26] (

Table 1. The field specifications for 30 m resolution DEM data.

Province	ID	Top Left	Bottom Right	Size (km2)	Shape	Year	Crop Type	Irrigated
Saskatchewan	39178	49.37, −109.28	49.35, −109.30	2.49		2023	Chick Peas	No
Manitoba	30955	51.01, −100.75	51.01, −100.76	0.10		2023, 2024	Alfalfa	No
Manitoba	30960	51.01, −100.73	51.00, −100.74	0.29		2023	Alfalfa	No
Manitoba	30967	51.02, −100.79	51.01, −100.81	1.22		2023, 2024	Alfalfa	No
Alberta	88697	49.92, −111.79	49.91, −111.80	0.32		2024	Canola	Yes
Alberta	88699	49.92, −111.77	49.91, −111.78	0.54		2024	Canola	Yes
Alberta	94413	56.89, −117.49	56.88, −117.50	0.59		2023	Spring Wheat	No
Alberta	99553	51.54, −112.50	51.53, −112.51	0.37		2024	Canola	No

). For the remaining three fields located in Manitoba, higher-resolution airborne LiDAR-based HRDEM (available at: https://search.open.canada.ca/openmap/957782bf-847c-4644-a757-e383c0057995; accessed on 25 April 2025) is used, offering a 1 m DTM [12]. These high-resolution fields are denoted by stars in Figure 5, and their detailed specifications are provided in

Table 2. The field specifications for 1 m resolution DEM data.

Province	ID	Top Left	Bottom Right	Size (km2)	Shape	Year	Crop Type
Manitoba	15346	50.37, −100.47	50.36, −100.48	0.48		2023	Soybean
Manitoba	14254	50.09, −99.58	50.08, −99.59	0.60		2023	Peas
Manitoba	5904	49.97, −100.79	49.96, −100.80	1.00		2023	Soybean

. Terrain features and landform classes are extracted from these DEM datasets, as detailed in Section 3.2.

To quantify crop performance, remote sensing provides a practical and scalable approach [10]. We utilized satellite imagery from Sentinel-2 at 10 m spatial resolution (available at: https://docs.sentinel-hub.com/api/latest/data/sentinel-2-l2a/; accessed on 3 February 2025). Two complementary metrics are employed to characterize crop performance and productivity (described in detail in Section 3.3). The annual metric, peak green, is derived from historical NDVI time series within a given year and serves as a proxy for in-season crop performance. Aggregating multi-year peak green values produces a more stable productivity indicator, zonemap.

System Overview. The overall workflow is illustrated in Figure 6. After computing terrain features, peak green images, and productivity zonemap, the next step involves spatially integrating these datasets for terrain–performance analysis. This integration requires a precise data registration step, ensuring all datasets shared the same coordinate reference system and spatial extent despite differences in native resolution (e.g., Sentinel-2 at 10 m vs. DEMs at 1 m or 30 m). Specifically, the latitude–longitude coordinates of each performance pixel are used to retrieve the corresponding DEM-derived features, ensuring one-to-one correspondence between performance and terrain features (see Section 3.4).

Following spatial integration, our visual analytics tool is used to interactively explore the relationship between terrain features (e.g., slope, TWI, landform classes) and performance and productivity metrics. These insights can guide a human-in-the-loop refinement process, in which the initial zonemap, based solely on peak green, is interpreted and adjusted to incorporate the combined effects of terrain variability.

3.2. Terrain Feature Extraction and Landform Classification

Raw DEM data cannot be directly used for terrain–performance analysis because differences in coordinate reference systems, spatial resolution, and boundary effects can introduce significant errors in derived terrain features. To ensure accurate measurements, the DEM is first projected to the UTM coordinate system, which preserves distances and areas necessary for our analysis. In addition, artifacts commonly occur when computing terrain features such as slope or the TWI near field boundaries, as these features rely on neighborhood operations that extend beyond the immediate field boundary. To mitigate these edge effects, a buffered version of the field boundary is created, clipped, and filled using gap-filling algorithms [13,29], ensuring a continuous elevation surface. Terrain features are computed within the buffered extent and then clipped back to the original field boundary, ensuring final analyses are spatially restricted to the exact field while benefiting from improved preprocessing near the edges.

DEM data provide the vertical elevation relative to mean sea level and serve as the basis for several derived terrain features. Here, we focus on the terrain features (i.e., slope, aspect, TWI and RTPI) selected by the domain expert for inclusion in the visual analytics tool, along with an additional metric—STI—which we identified as potentially important through our own analysis. All terrain features are computed with WhiteboxTools Version 2.3.0 (Geomorphometry & Hydrogeomatics Research Group, University of Guelph, Guelph, ON, Canada; manual: https://www.whiteboxgeo.com/manual/wbt_book/intro.html; accessed on 4 March 2025) [13]; the exact command-line commands and options used are listed in Appendix A. Slope measures the maximum rate of elevation change between each pixel and its neighbors and is expressed in degrees. Aspect identifies the slope azimuth, ranging from $0^{\circ }$ to ${360}^{\circ }$, indicating the direction of steepest descent [5,68]. Several curvature metrics are included: profile curvature (curvature in the steepest slope direction, indicating flow acceleration or deceleration), plan curvature (curvature measured perpendicular to the slope direction (i.e., along the contour lines), indicating flow convergence/divergence), total curvature (combined plan and profile), and tangential curvature (slope-weighted plan curvature). TPI describes the relative elevation of a cell within the surrounding terrain, distinguishing local convexity or concavity. We employ a “Relative TPI (RTPI)” formulation [69], following [13]:

$$\mathrm{RTPI}=\left\{\begin{array}{cc}{\displaystyle \frac{z_0-\mu }{\mu -z_{min}}},\hfill & z_0<\mu \hfill \\ {\displaystyle \frac{z_0-\mu }{z_{max}-\mu }},\hfill & z_0\ge \mu \hfill \end{array}\right.$$

(1)

where $z_0$ is the focal cell elevation, $\mu$ is the neighborhood mean, and $z_{min}$, $z_{max}$ are the local extremes. Flow direction defines the steepest downslope path and flow accumulation is the number of upslope cells that drain into a given pixel, based on the flow direction algorithm [5,34]. TWI is a proxy for soil moisture potential based on specific contributing area $A_s$ (the upslope area draining through a unit contour length) and slope angle $\beta$, following [13]:

$$\mathrm{TWI}=ln\left({\displaystyle \frac{A_s}{tan\beta }}\right).$$

STI integrates specific contributing area and slope and is used as a proxy for relative erosion potential, specifically the sediment transport component [13,29,32]:

$$\mathrm{STI}={\left({\displaystyle \frac{A_s}{22.13}}\right)}^m{\left({\displaystyle \frac{sin\left(\beta \right)}{0.0896}}\right)}^n,m=0.4,n=1.3$$

where $A_s$ is specific contributing area and $\beta$ is the slope. This parameterization has been used in recent agricultural terrain studies [51,70].

Hillshade is also computed to enhance terrain visualization, and it represents the intensity of reflected light from terrain surfaces, as influenced by slope orientation and the position of an artificial light source. While hillshade is used for visualization, it is fundamentally a mathematical derivative of terrain geometry and highlights subtle terrain features overlooked by simple peak detection. We use hillshade in hilltop identification to capture subtle topographic features that simple peak detection algorithms miss. To reduce directional bias, a multi-directional hillshading approach [13] is used, combining hillshades generated from eight azimuth angles with a moderate illumination altitude (${25}^{\circ }$). This enhances the identification of subtle convexities typically found in low-relief agricultural fields.

Simplified landform classification. Applying our visual analytics tool allowed us to explore the spatial relationships between terrain features and field performance. Naturally, we sought to verify the agronomic impact of two extreme landforms, hilltops and depressions, on field performance. Although other landforms, such as plains, ridges, valleys, and saddles, exist between these extremes, our objective was to assess the importance of these two key landforms and evaluate the effectiveness of our visual analytics approach. To facilitate this, we consolidated the remaining landforms into a third category, baseline, representing transitional or intermediate terrain (e.g., plains and slopes).

Terrain depressions are the concave areas prone to moisture accumulation, while hilltops denote convex areas, often drier and better drained. All remaining areas not classified as depressions or hilltops are grouped as baseline areas (See Figure 7). To further examine topographic effects within the baseline class, we introduce two distinct refinements that are analyzed in Section 4.3 and Section 4.4, respectively: (i) a flat vs. non-flat split based on a slope threshold [55], and (ii) a three-class grouping of $log\left(STI\right)$ to investigate how relative erosion potential (proxied by STI) relates to performance.

We adopt this three-class framework, depressions, hilltops, and baseline, to simplify landscape complexity while preserving agronomically meaningful distinctions that support terrain–performance analysis and field-level decision-making relevant to site-specific management and balancing interpretability with actionable insight, avoiding complex zoning.

Depression delineation. Depressions are bowl-like areas where surrounding cells are higher in elevation (Figure 7a), and they can be natural or caused by noise and errors in DEM generation. There is no universally accepted method for accurately delineating the terrain’s real depressions [71], and methods try to balance the removal of artifacts with the preservation of actual topographic features. In this paper, depressions are identified using a multi-step approach applied to the preprocessed DEM (Figure 8). First, artificial pits are corrected using breaching and filling operations to ensure hydrologically sound surfaces. Then, depressions are identified and only significant depressed areas are retained by applying depth thresholds to eliminate minor artifacts [72]. The result is binarized to isolate depressions, and a neighborhood filter removes isolated pixels (<3 neighbors).

Hilltop delineation. Hilltops are convex points elevated relative to their neighborhood (Figure 7b). Simple peak detection techniques (highest within a moving window) are highly sensitive to window size [73]. Other approaches using curvature, slope, TPI, shadow boundaries, or pattern recognition [54,55,74] often overfit to noise in low-relief farmland and require extensive site-specific threshold tuning for each field. Multi-directional hillshade has previously been applied in landform classification [74,75] to enhance the detection of subtle convexities. Building on this idea, we adopted a hybrid percentile-based rule: a hilltop pixel must exhibit both high multi-directional hillshade brightness (top 20%) and a positive RTPI (>0) (See Figure 9). We used a percentile threshold to avoid calibrating absolute brightness values. A 20% cutoff was selected as the default after visual assessment across study fields: 10% produced sparse speckling, >30% yielded broad patches and higher sensitivity to RTPI window size, while 20% provided a practical balance for low-relief agricultural terrain. Isolated points were removed with the same neighborhood filter applied to depressions.

Baseline. Remaining pixels not classified as hilltops or depressions are grouped as baseline areas. Figure 7c,d illustrate baseline terrain, representing transitional slopes and plains.

Example terrain features. To illustrate the method, Figure 10 shows terrain features for a Saskatchewan field derived from 30 m TanDEM-X DEM: (a) elevation (931–945 m), (b) slope (0°–20°), (c) aspect (azimuth in degrees [0,360), encoded in grayscale), (d) TWI (4.3–15), (e) RTPI ($-1$ concave → $+1$ convex), (f) multi-directional hillshade, (g) depressions (blue pixels), and (h) hilltops (red pixels).

3.3. Field Performance, Zonemap, and Terrain–Performance Relationship

Peak green. Field performance can be assessed using various approaches, such as yield monitoring, proximal soil sensing (e.g., electrical conductivity, ECa), or VIs derived from remote sensing. While direct yield measurements provide valuable ground truth, applying these approaches is often challenging. In contrast, satellite-based VIs offer a cost-effective and scalable means to characterize within-field variability, providing a stable proxy for long-term productivity potential. In this paper, Sentinel-2 Level-2A imagery (10 m resolution) is used to derive a VI-based performance indicator [76]. NDVI is computed using the near-infrared (band 8, 842 nm) and red (band 4, 665 nm) spectral bands of Sentinel-2 as follows:

$$NDVI={\displaystyle \frac{\mathrm{NIR}-\mathrm{Red}}{\mathrm{NIR}+\mathrm{Red}}}.$$

(2)

For each field and year, n cloud- and haze-free NDVI images from the growing season (typically June–August) are collected ($R_1,R_2,\dots ,R_n$). To characterize maximum crop performance within the growing season, we select the peak green image, defined as the image whose NDVI values are closest to the pixel-wise seasonal maximum. Mathematically, for each pixel $(x,y)$, the seasonal pixel-wise maximum value $R_S(x,y)$ is computed as

$$R_S(x,y)=\underset{j=1,\dots ,n}{max}\left\{NDVI\left(R_j(x,y)\right)\right\},$$

(3)

where $R_j(x,y)$ represents the NDVI value of the j-th image at $(x,y)$. The image $R_p$ corresponding to the acquisition date $p\in [1,n]$ with NDVI values closest to $R_S$ is selected as the representative peak green image for that year (e.g., see Figure 3) [23,24].

The resulting peak green image serves as a proxy for the maximum observed crop performance during the growing season.

Zonemap. To create a stable productivity pattern that reflects the field’s long-term performance potential over multiple years, abnormal years’ data are excluded [24], then peak green images are normalized and averaged. The resulting averaged performance is subsequently classified into discrete MZs by partitioning the distribution of values into statistical bins (e.g., 5 or 6 classes). Lower classes represent consistently underperforming areas, while higher classes represent stable, high-performing zones (See Figure 11). The resulting zonemap provides a spatially explicit representation of the field’s productivity variability, which can guide site-specific management decisions.

Terrain–Performance relationship. Terrain attributes such as slope, curvature, and TWI affect soil moisture, drainage, and fertility, thereby affecting field performance. By integrating performance data with these terrain features, we visualize and quantify how various topographic features correlate with field performance.

While individual terrain features provide valuable insight, landforms often govern more interpretable patterns in performance and productivity. Closed depressions behave as moisture and nutrient accumulation zones, whereas well-drained hilltops are prone to erosion and drought stress. To capture this structural context, we intersected performance data with the simplified three landform classes (depressions, hilltops, and baseline areas) described in Section 3.2. For example, Figure 12 illustrates this integration for a field in Manitoba. Panel (a) shows the peak green (2023). Panels (b–d) depict performance values for each landform class. In panel (b), a depression, denoted by a circle, exhibits higher productivity due to improved moisture retention, whereas some southern depressions show lower values, likely reflecting waterlogging, nutrient leaching, or other depression related issues. These patterns demonstrate how terrain context explains performance variability beyond what zonemaps alone reveal.

3.4. Data Registration

Accurate spatial integration of terrain features and performance metrics requires reconciling differences in spatial resolution and coordinate reference systems (CRS). As illustrated in Figure 6, terrain features are extracted from DEMs with either 1 m or 30 m spatial resolution, while peak green and zonemap are generated from Sentinel-2 imagery at 10 m resolution. At ingestion, the base DEM was reprojected to the Sentinel-2 UTM/WGS84 CRS, and all terrain features were computed in this CRS. To enable pixel-wise correspondence for terrain–performance analysis, each 10 × 10 m performance pixel must be reliably aligned with its corresponding DEM-derived terrain features using a resolution-aware sampling approach.

To achieve this alignment, we implemented a sampling-based alignment procedure that accounts for resolution differences (see Figure 13). For 1 m DEMs, each peak green pixel is subdivided into a 1 × 1 m grid in geographic coordinates. Terrain values are then sampled at these subpixels and averaged to produce a representative terrain feature value for that pixel. For 30 m DEMs, peak green pixels could fall entirely within a single DEM cell or intersect multiple cells. In the former case, the terrain value of that cell is directly assigned. In the latter case, values are computed as a weighted average, where the contribution of each terrain cell is proportional to its area of overlap with the 10 m pixel.

For binary terrain features (e.g., depressions), a 0.5 threshold is applied to the averaged value to assign a class label. This resolution-aware sampling strategy ensures spatial consistency between datasets and enables robust terrain–performance integration across fields with differing DEM sources.

3.5. Visual Analytics for Terrain–Performance Integration

While spatial registration ensures pixel-level alignment between terrain and performance datasets, meaningful interpretation requires thoughtful integration at the visualization level. Presenting multiple spatial features, such as peak green, zonemap, terrain features (e.g., slope, TWI, RTPI), and landform classes within a single interface introduces challenges related to visual complexity, interpretability, and user focus. To address these, we developed a focus+context visual analytics tool that enables simultaneous exploration of terrain features and crop performance at the field scale within a unified, interactive environment.

In our visualization framework, the context panel serves as a spatial anchor for interpretation while surrounding panels (focus) present detailed views of additional features. Although field performance (e.g., peak green, zonemap) is typically used as the default context, any relevant feature, such as hillshade or slope, can serve as context depending on the objective. For example, hillshade may be selected as context when exploring landform structure, while peak green may anchor interpretation when assessing how terrain features influence crop performance. Similarly, the focus panels are not fixed; they may be chosen dynamically to display any combination of terrain features or performance based on the specific analysis. For example, in drainage-sensitive areas, the user may prioritize TWI, flow accumulation, and curvature. In erosion-prone regions, slope, STI, or profile curvature may be more relevant (See Figure 14).

Although additional terrain features, such as RTPI or aspect, can be toggled into the focus panels as needed, prior testing and expert feedback suggest that displaying more than four close-ups concurrently compromises visual clarity and interpretability (Figure 15). The initial set of four candidate layers (e.g., TWI, slope, curvature, RTPI) was proposed by an agronomist and refined iteratively. In the final tool, any of the computed terrain features can be selected into the four focus panels.

In addition to visual inspection, the tool includes an analytics module that computes summary statistics (e.g., range, median, CV) and Pearson correlation coefficients between selected terrain features and peak green, within the defined focus region. These quantitative outputs complement the visual representation and provide evidence-based support for refining management zones.

4. Results and Discussion

We applied the visual analytics tool to eleven agricultural fields to evaluate terrain–performance relationships. Results are presented in five parts: expert evaluation (Section 4.1), analysis across landform classes (Section 4.2), finer subdivision of baseline areas (Section 4.3 and Section 4.4), and synthesis of agronomic implications (Section 4.5). For each landform, performance is summarized using mean and coefficient of variation (CV).

4.1. Expert Evaluation and Practical Implications

To assess the tool’s practical value, we obtained feedback from a crop production and GIS expert with experience across diverse agricultural regions in the Canadian Prairies. Their input provided valuable insights into both the usability of the visualization system and its potential applications in decision-making workflows.

The slope and TWI features were identified as particularly valuable for understanding water redistribution patterns, especially in fields prone to waterlogging or poor drainage. Hillshade was identified as the most intuitive visual layer for grasping terrain structure and was recommended as an alternative for the context. The expert confirmed that displaying up to four layers concurrently is appropriate, provided they are presented separately rather than stacked, which aligns with our final interface design. To guide feature selection, it is recommended to populate the four panels with the highest correlated features while limiting to at most one per process category, such as water redistribution (e.g., TWI, flow accumulation), and terrain structure (e.g., RTPI, hillshade, landform classes), to reduce redundancy and preserve interpretability.

Functionally, the focus+context visualization was considered most beneficial in fields with complex terrain or known management challenges, rather than for all fields. The tool was also recognized as useful for refining MZs in response to atypical climatic conditions (e.g., droughts or excess rainfall), where fertilizer or irrigation strategies may require adjustment. Additional feedback emphasized the need for identifying proximity to gravel roads, which can influence remote sensing imagery due to dust interference.

Several of these recommendations have already been implemented. Hillshade is employed as an alternative for the context image to support interpretation of elevation patterns (Figure 16). Four focused regions are visualized in parallel, each in a separate panel, enabling direct visual comparison without occlusion. A wide range of terrain features, such as slope, TWI, and RTPI, are integrated into the interface (Figure 10). To enhance geographic orientation and environmental context, we incorporated a Bing Maps (Microsoft) basemap [77]. This addition enables visualization of roads, field boundaries, and surrounding land features, supporting interpretation of field conditions and potential sources of image distortion (e.g., dust from gravel roads). However, the basemap does not explicitly distinguish between gravel and paved surfaces. As such, automated identification or quantification of gravel road impacts remains a direction for future enhancement.

Overall, the expert evaluation validated the system’s utility in terrain-aware decision support and highlighted specific directions for future improvements. These include the integration of additional overlays and the potential automation of combined-feature relationship analysis (e.g., slope–TWI) with productivity maps. While this evaluation provided valuable domain-specific insights, it is important to acknowledge that it was based on the feedback of a single expert, and thus may not capture the full range of perspectives that could arise in broader user testing.

4.2. Results and Discussion for Three Landform Classes

This section presents performance metrics across three terrain classes (depressions, hilltops, and baseline areas) for both 30 m and 1 m resolution DEMs. Across all fields, baseline areas, which comprise on average more than 80% of the field area, consistently exhibit the highest mean performance and, in over 90% of cases, the lowest CV. In most fields, baseline performance exceeds that of hilltops and depressions by at least 10%. These patterns are illustrated in Figure 17, which shows the characteristic high-performance, low-variability signature of baseline areas, with detailed results summarized in

Table 3. Summary of mean and variation of performance by landform type in 30 m resolution DEM data.

Field ID	Year	Landform Class	Area%	Mean	CV
39178	2023	Depression	6.70	0.48	0.24
		Hilltop	6.31	0.37	0.24
		Baseline	87	0.61	0.14
30955	2023–	Depression	1.71	0.85	0.01
	2024	Hilltop	6.45	0.81	0.05
		Baseline	91.84	0.90	0.04
30960	2023	Depression	36.77	0.81	0.08
		Hilltop	5.11	0.85	0.02
		Baseline	58.12	0.92	0.03
30967	2023–	Depression	5.23	0.80	0.14
	2024	Hilltop	13.60	0.85	0.07
		Baseline	81.16	0.92	0.03
88697	2023	Depression	0.93	0.79	0.05
		Hilltop	7.70	0.76	0.06
		Baseline	91.37	0.88	0.06
88699	2023	Depression	6.98	0.74	0.08
		Hilltop	10.47	0.73	0.09
		Baseline	82.55	0.85	0.07
94413	2023	Depression	9.65	0.59	0.15
		Hilltop	8.69	0.49	0.14
		Baseline	81.66	0.72	0.10
99553	2023	Depression	0	0	0
		Hilltop	17.93	0.57	0.18
		Baseline	82.07	0.74	0.12

and 

Table 4. Summary of mean and variation of performance by landform type in 1 m resolution DEM data.

Field ID	Year	Landform Class	Area%	Mean	CV
5904	2023	Depression	17.54	0.52	0.31
		Hilltop	3.20	0.57	0.16
		Baseline	79.25	0.74	0.14
14254	2023	Depression	1.71	0.48	0.30
		Hilltop	7.45	0.75	0.05
		Baseline	90.84	0.85	0.07
15346	2023	Depression	7.46	0.63	0.27
		Hilltop	5.72	0.78	0.11
		Baseline	86.82	0.88	0.06

. To visualize these differences more intuitively, Figure 18 shows the distribution of performance across the three landform classes for the HRDEM data. Baseline areas exhibit higher medians and tighter interquartile ranges than depressions and hilltops, indicating better performance with lower variability.

Discussion. These patterns align with prior studies linking concave landforms to water accumulation and potential waterlogging, and convex features to nutrient and water loss via runoff [30,31,46,47]. For example, Clare et al. [46] found that drained depressions in central Alberta still underperformed relative to uplands. Boldt et al. [47] extended this finding by demonstrating that yield loss extends 30–50 m beyond the wetland core. Importantly, as illustrated in Figure 12, not all depressions perform poorly; our tool facilitates the identification of unstable depressions, aiding targeted management interventions.

Two alfalfa fields (30955 and 30967) were managed with uniform fertilizer applications over two consecutive years. This consistent management minimizes residual effects from previous years’ fertilization strategies, providing an opportunity for assessing year-to-year stability (

Table 5. Year-wise performance for fields 30955 and 30967.

		Field 30955		Field 30967
Year	Landform Class	Mean	CV	Mean	CV
2023	Depression	0.86	0.01	0.78	0.17
	Hilltop	0.82	0.05	0.85	0.09
	Baseline	0.91	0.03	0.92	0.04
2024	Depression	0.84	0.01	0.81	0.10
	Hilltop	0.79	0.06	0.85	0.05
	Baseline	0.89	0.05	0.92	0.02

). In both cases, baseline areas consistently outperformed other classes. For instance, in field 30967, baseline areas exhibited a mean performance of 0.92 and a CV of 0.02–0.04 in both years, indicating high productivity and temporal consistency. In contrast, CV values in depressions reached 0.17 in 2023, with slightly elevated variability in the next year, consistent with observations in [78] that uniform fertilizer strategies can produce variable yields across heterogeneous terrain.

In the drought year 2023, fields 39178 and 94413 showed a clear moisture-buffering signal in depressions relative to hilltops: depressions exceeded hilltops by about 0.11 and 0.10 performance mean, respectively, while baseline remained highest in all fields. Coefficients of variation were larger in both depressions and hilltops than in baseline, indicating greater sensitivity to crop growth drivers. While not universal across all fields, in these two fields, the observed contrasts are consistent with drought-year moisture buffering in depressions, narrowing the performance gap with hilltops while not surpassing the generally more stable baseline areas.

In fields 88697 and 88699, which were irrigated using center-pivot systems and cultivated with canola, performance contrasts across terrain classes were partially reduced during the drought year. CVs were also relatively low and similar across classes, reflecting the buffering effect of irrigation on water stress. However, baseline areas still exhibited the highest mean performance. These results suggest that while irrigation reduces climate-driven heterogeneity, it does not fully remove terrain influences, consistent with findings by [19].

Analyses using high-resolution HRDEM data further validate the trends observed in TanDEM-X-based fields. While HRDEM provides more detailed microtopographic information, the spatial patterns in performance were broadly consistent across resolutions. In all three high-resolution fields, baseline areas showed the most stable performance, with the exception of hilltops in field 14254. For example, in field 15346, baseline had a CV of 0.06, while depressions showed more than four times the variability (CV = 0.27). These results align with those of Tóth et al. [7], who reported lower NDVI and higher variability in depressions using 100 m DEMs. This cross-resolution agreement highlights the robustness of terrain–performance relationships across elevation products.

More broadly, our findings underscore that productivity zones derived solely from satellite imagery (e.g., zonemaps generated using the method described in Section 3.3) may not reliably reflect underlying terrain variation. For instance, areas assigned the same zone label may occur in distinct landform conditions, with differing soil, water, or nutrient dynamics and requirements. Therefore, topographic features provide essential context for refining MZs, particularly in complex or hydrologically sensitive areas. This insight aligns with prior work [6,47], which emphasizes that both nutrient cycling and yield variability are tightly coupled to microtopographic structure.

4.3. Results and Discussion for Baseline Classes Based on Slope

In previous sections, landform analysis focused on three classes: depressions, hilltops, and all remaining areas, which were broadly grouped as baseline. This residual category, however, includes a range of microtopographies that may influence crop performance. In this and next experiment (Section 4.4), we investigate whether further differentiation within baseline areas provides useful insights for terrain-informed management.

Rather than applying complex geomorphometric classification (e.g., shoulders, backslopes, ridges), which may introduce unnecessary zoning complexity, we adopted a slope-based categorization. Baseline pixels were split into flats and non-flats using slope thresholds ($\theta =5^{\circ }$ for 30 m DEMs, $\theta =2^{\circ }$ for 1 m DEMs) [55,79]. The analysis was restricted to baseline pixels only, excluding areas previously classified as depressions or hilltops.

Table 6. Summary of mean and variation of performance within the baseline areassubdivided by slope in 30 m resolution DEM data.

Field ID	Landform Class	Area% of Baseline	Mean	CV
39178	Flat	59.25	0.61	0.14
	Non-Flat	40.75	0.61	0.14
30955	Flat	28.08	0.89	0.04
	Non-Flat	71.92	0.90	0.04
30960	Flat	44.18	0.92	0.02
	Non-Flat	55.82	0.91	0.03
30967	Flat	25.76	0.92	0.03
	Non-Flat	74.24	0.92	0.02
88697	Flat	70.23	0.88	0.03
	Non-Flat	29.77	0.85	0.09
88699	Flat	92.28	0.86	0.05
	Non-Flat	7.72	0.77	0.14
94413	Flat	91.93	0.72	0.10
	Non-Flat	8.07	0.75	0.08
99553	Flat	9.19	0.78	0.08
	Non-Flat	90.81	0.74	0.12

and 

Table 7. Summary of mean and variation of performance within the baseline areas subdivided by slope in 1 m resolution DEM data.

Field ID	Landform Class	Area% of Baseline	Mean	CV
5904	Flat	79.21	0.75	0.13
	Non-Flat	20.79	0.70	0.17
14254	Flat	62.11	0.85	0.08
	Non-Flat	37.89	0.85	0.05
15346	Flat	48.91	0.88	0.06
	Non-Flat	51.09	0.89	0.05

summarize the results.

Discussion. At 30 m resolution, flats and non-flats within baseline areas showed similar performance, suggesting limited benefit from slope-based subdivision at coarse scales within the baseline areas. This outcome likely reflects that subtle slope differences are smoothed at 30 m DEM resolution, reducing their influence on performance statistics. In irrigated fields (88697 and 88699), flats dominated and performed marginally better, possibly reflecting more uniform irrigation effects in flat areas. These findings suggest that slope-only subdivision is of limited diagnostic value at coarse scales.

At 1 m resolution, more microtopographic detail was captured, with a larger share of pixels classified as flat. Differences between flats and non-flats were field-specific: in field 5904, flats slightly outperformed non-flats, whereas other fields showed negligible differences. These results suggest that slope-based subdivision within baseline areas may be more informative at high resolution; however, its practical value depends on context. In particular, it is most relevant in fields where subtle slope variations influence water distribution, drainage, localized soil erosion, or other agronomically significant processes, and may not offer actionable insights in all landscapes.

4.4. Results and Discussion for Baseline Classes Based on STI

In addition to slope-based subdivision, baseline areas were further categorized using the STI, which incorporates slope and specific contributing area to capture erosive potential and hydrological redistribution. This makes STI a more meaningful indicator than slope alone, as it accounts for both erosivity and hydrological convergence, extending the analysis from the previous experiment (Section 4.3). To illustrate the role of STI in terrain–performance analysis, Figure 19 shows an example field (ID 5904) where the hillshade is the context image and STI, slope, and relative TPI are displayed alongside the zonemap and summary terrain features statistics. As shown, STI has negative correlation with performance and highlights erosive flow paths and areas of hydrological convergence that are not readily apparent from slope or RTPI alone.

As seen in Figure 19, the STI image appears noticeably darker, reflecting the presence of right-skewed raw STI values. Therefore, we applied a logarithmic transformation prior to classification to improve normality and ensure more statistically balanced partitioning (i.e., to not have few pixels in high-STI and many in low-STI). The transformed $log\left(STI\right)$ values were then grouped into three classes (low, mid, and high) using the Fisher–Jenks natural breaks algorithm [80], which has also been applied in precision agriculture for management zone delineation [81]. Compared to simple quantile thresholds (e.g., 33rd and 66th percentiles), Fisher–Jenks minimizes within-class variance and maximizes between-class separation, thereby adapting to the true distribution of $log\left(STI\right)$ in each field.

In this experiment, we classified baseline areas into three classes based on $log\left(STI\right)$ values. Then we analyzed performance metrics (mean, CV) in each class. The results of this experiment are summarized in

Table 8. Summary of mean and variation of performance within baseline areas subdivided by $log\left(STI\right)$ classes in 30 m resolution DEM data.

Field ID	STI Class	Area% of Baseline	Mean	CV
39178	LowSTI	18.5	0.60	0.14
	MidSTI	48.2	0.61	0.14
	HighSTI	33.2	0.62	0.15
30955	LowSTI	24.8	0.89	0.04
	MidSTI	42.3	0.90	0.04
	HighSTI	33.0	0.90	0.04
30960	LowSTI	24.4	0.92	0.02
	MidSTI	54.3	0.92	0.02
	HighSTI	21.3	0.90	0.04
30967	LowSTI	21.7	0.92	0.04
	MidSTI	47.6	0.92	0.04
	HighSTI	30.7	0.93	0.03
88697	LowSTI	29.1	0.88	0.03
	MidSTI	49.8	0.88	0.05
	HighSTI	21.1	0.86	0.09
88699	LowSTI	39.1	0.86	0.05
	MidSTI	51.8	0.86	0.06
	HighSTI	9.1	0.79	0.13
94413	LowSTI	33.8	0.72	0.10
	MidSTI	55.7	0.72	0.11
	HighSTI	10.5	0.74	0.09
99553	LowSTI	14.5	0.75	0.11
	MidSTI	46.7	0.74	0.12
	HighSTI	38.8	0.74	0.12

and 

Table 9. Summary of mean and variation of performance within baseline areas subdivided by $log\left(STI\right)$ classes in 1 m resolution DEM data.

Field ID	STI Class	Area% of Baseline	Mean	CV
5904	LowSTI	36.7	0.75	0.11
	MidSTI	44.3	0.73	0.15
	HighSTI	19.0	0.71	0.16
14254	LowSTI	25.1	0.84	0.08
	MidSTI	42.9	0.85	0.07
	HighSTI	32.0	0.85	0.05
15346	LowSTI	23.2	0.88	0.06
	MidSTI	48.4	0.88	0.06
	HighSTI	28.4	0.89	0.05

and Figure 20.

Discussion. Patterns across fields showed both consistency and field-specific variability. In 30 m DEM data, in some cases (39178, 30955, 30960, 30967, 94413, and 99553), performance had subtle changes across all $log\left(STI\right)$ classes, indicating limited differentiation within the baseline. In others (88697 and 88699), LowSTI and MidSTI zones’ performance mean outperformed HighSTI with lower CV, consistent with steeper slopes affecting soil erosion potential [70]. At 1 m resolution the contrasts were clearer in 5904, highlighting the added value of high-resolution DEMs in capturing subtle terrain–performance interactions. However, the other two fields’ performance had small changes across STI classes, and performance showed slightly better values in High-STI class during the drought year.

Despite these field-level contrasts, across all 11 fields, we did not observe systematic differences in performance mean or CV between high- and low-STI baseline areas, suggesting that water convergence can sometimes offset erosive disadvantages specifically in drought years (e.g., 2023). This observation is consistent with previous studies that demonstrated relationships between terrain metrics and crop performance are field- and weather- specific [18]. Also, this reflects the conceptual nature of STI, which represents potential erosivity rather than realized soil loss, which may be less relevant during a drought year (2023). Moreover, NDVI values may saturate during peak growth stages, when vegetation cover and biomass are highest, reducing sensitivity to subtle erosion or redistribution effects.

Overall, STI-based subdivision reveals heterogeneity within baseline areas that slope-only approaches may not capture. While not universally decisive, in certain fields it distinguishes regions where performance is either penalized by erosion or enhanced by water redistribution.

4.5. Implications for Terrain-Informed Management

The observed terrain–performance relationships have several practical implications for site-specific management:

• Depressions, while occasionally exhibiting high productivity, were the most unstable due to excess moisture, shallow rooting depth, soil compaction, or limited aeration, which can promote denitrification and yield loss [6,82]. Effective treatment may include drainage improvements, split nitrogen applications, or edge-of-field practices to mitigate nutrient leaching [6,83]. The developed tool enables precise identification of low-performing depressions to support such interventions.
• Hilltops outperformed depressions in most cases, likely due to better drainage or soil characteristics. However, they often underperformed compared to baseline, possibly due to erosion, evaporation, or limited water retention. Targeted conservation practices, such as organic mulching or contour-based farming, may improve sustainability in these areas [84]. The tool allows for proactive detection and treatment of underperforming hilltop zones.
• Baseline areas exhibited the highest productivity and lowest variability across nearly all cases. These zones serve as stable reference regions for assessing treatment impacts, calibrating models, or estimating spatial yield. Their consistent performance makes them highly reliable for predictive modeling and management benchmarking.
• Baseline subdivisions, when further stratified by slope or STI, revealed that subtle microtopographic variation can still influence performance stability. While differences were not systematic across all fields, some high-STI zones underperformed due to erosion risk, whereas others showed equal or better productivity, likely benefiting from lateral water convergence. Using high-resolution DEM data, these contrasts were clearer, underscoring the potential of fine-scale DEMs to guide targeted interventions such as erosion control or water redistribution practices within otherwise stable baseline regions.

5. Conclusions and Future Works

This paper presents a terrain-based analysis of crop performance variability using an introduced novel visual analytics tool. We classified agricultural fields into depressions, hilltops, and baseline areas by integrating DEMs with satellite-derived performance metric. The expert’s feedback shows this tool enables a deeper understanding of terrain influences on crop productivity. The tool was applied to eleven fields across Canada under constant-rate fertilizer application but varying environmental conditions and crops. The results show that the introduced focus+context visualization approach enhances interpretability by presenting close-ups of terrain features alongside performance maps. This tool enables validating zonemaps derived from satellite imagery and supports more informed decisions in field treatment and MZ refinement. Our method was also applied to fields with high-resolution HRDEM data, demonstrating its adaptability and robustness across spatial scales.

Results show that baseline areas exhibited higher performance and lower variability. In contrast, depressions and hilltops were associated with more instability that could be due to nutrient- or water-related stressors. Even in irrigated fields, topographic influence is present. The findings confirm the value of incorporating terrain structure into field-scale management, particularly in refining MZs in fields with heterogeneous or highly variable topography, and identifying areas requiring specific treatments (e.g., drainage or soil conservation interventions).

While this study demonstrates the value of terrain-informed performance analysis, several limitations should be acknowledged. First, the analysis relied primarily on peak green (computed based on NDVI) as a performance proxy, which can saturate under high biomass conditions. Alternative performance metrics or radar-based VIs may provide complementary information. Second, thresholds used for terrain classification (e.g., depressions, hilltops) are methodologically transparent but may be subjective or field-dependent, and their universality across agroecological regions requires further testing. Finally, expert validation was based on a single agronomist, providing detailed domain knowledge but limiting the breadth of perspectives. These limitations highlight opportunities for refining the approach and testing its applicability across broader datasets and contexts.

Future research will focus on developing and evaluating strategies for zonemap refinement, with an emphasis on integrating terrain knowledge into management decisions. Based on expert feedback, expanding the analysis to include additional and combined terrain features is a key direction. In addition, future work will involve validation with multiple experts and structured user studies to broaden the evaluation of the tool across diverse agronomic contexts. More broadly, adopting terrain-aware approaches in other aspects of agricultural management (e.g., crop growth monitoring, soil surveys) holds significant potential. The focus+context visualization framework could also be extended to support multi-feature exploration in other field management aspects, such as detecting anomalies by jointly visualizing multiple features, such as drought indices and growth metrics, or enhancing soil surveys by displaying multiple soil characteristics simultaneously. We also plan to apply the approach to higher-resolution performance datasets (e.g., 3 m satellite imagery or yield maps). Furthermore, applying the approach across variable-rate fertilizer strategies to assess its generalizability and operational scalability.