Graphics-Guided Interactive Farmland Layout Design

Abstract

The spatial layout of farmland involves coordinated planning across diverse functional zones. Irregular land boundaries and functional demands pose challenges to traditional CAD-based workflows and general optimization algorithms. To address these limitations, we propose an interactive farmland layout system based on the Graphic-Guided Evolutionary Layout (GGEL) algorithm. GGEL not only introduces a graph-based spatial pruning and encoding strategy to improve search efficiency, but also performs real-time spatial overlap detection based on functional region boundaries to ensure layout feasibility. Additionally, an interactive module enables real-time user customization to accommodate specific planning needs. Experimental results demonstrate that the system can efficiently generate complete multi-region layouts, significantly enhancing design productivity. A user study with 20 agricultural park experts confirms the system’s usability and effectiveness. This study highlights the potential of integrating evolutionary algorithms with topological graph representations to address the complex spatial planning requirements of digital agricultural parks.

1. Introduction

The automation of farmland layout design plays a critical role in enhancing land-use efficiency, resource allocation, and digital agricultural management, making it a cornerstone of smart agriculture. Traditional design processes are often labor-intensive and repetitive, hindering both productivity and flexibility in spatial planning [1]. To address these limitations, the integration of artificial intelligence (AI) offers a promising pathway toward more efficient and intelligent farmland design. Existing AI-based approaches, including deep learning models [2,3,4,5], typically require extensive training datasets, while heuristic algorithms [6,7,8,9,10,11] often treat layout units as geometric planes, resulting in increased computational cost when modifying their sizes and positions. However, these methods offer limited opportunities for expert intervention and fail to leverage the valuable domain knowledge of agricultural professionals.

This paper specifically addresses this gap by proposing an interactive approach that integrates human expertise into the layout optimization process. Recent research highlights the potential of using simple 2D graphical and topological representations to streamline complex design workflows [12], reducing computational complexity and enhancing user interaction. Building on this insight, we propose a phased, interactive farmland layout system that incorporates a user-friendly interface and an advanced 3D visual editing environment. The system begins by preprocessing freehand boundary sketches provided by the user, then applies the Graphics-Guided Evolutionary Layout (GGEL) algorithm to efficiently generate high-quality 2D layouts while ensuring spatial feasibility. It further supports real-time user adjustments and offers procedural tools for detailed scene refinement, enabling flexible and efficient farmland planning.

Our key contributions are as follows:

• We introduce a topologically aware optimization algorithm, Graphics Guided Evolutionary Layout (GGEL), which ensures fast convergence and spatial feasibility. GGEL significantly improves computational efficiency, reducing optimization time while maintaining high-quality layout solutions.
• We develop an interactive layout module that allows users to dynamically modify farmland layouts. This module provides real-time feedback to ensure the final design meets specific agricultural requirements and spatial constraints.
• By combining GGEL and the interactive layout module, we design a user-friendly system that enables efficient design and refinement of farmland layouts. With 95% support for creative flexibility and 90% support for usability, the system demonstrates its effectiveness in real-world applications.

Qualitative and quantitative evaluations show that our system combines topological optimization with interactive design, significantly improving the efficiency of farmland layout planning and supporting users in creating customized farmland designs. This work advances the application of artificial intelligence in smart agriculture and provides a scalable and user-friendly solution to modern agricultural challenges.

2. Related Work

Overall, our work belongs to the task of layout generation based on structured arrangements, aiming to use graph structure and heuristic algorithms to optimize scene layout, with a particular focus on automatic design of farmland layout. Significant progress has been made in the research of graph-guided generation models and heuristic algorithms, especially in urban layout [13,14,15,16], architectural layout [5,17,18,19], and other fields [20,21,22,23]. However, the research on farmland layout remains in the preliminary stage, facing challenges such as complex land shapes and dynamic demands.

2.1. Graph-Guided Generative Models

A graph is a structured data model that can represent the spatial relationships and semantic information of objects in a scene. In recent years, graph structure has been widely used in 2D layout design, which effectively captures the user’s layout intention and simplifies the input. Zhang et al. [24] review recent work on graphic visualization and computational efficiency optimization and propose concepts of pre-layout, in-layout, and post-layout. The pre-layout method focuses on graphic data compression, the in-layout method optimizes the layout process from the computational architecture and algorithm, and the visual mapping and interactive layout adjustment are the post-layout optimization techniques. Sepulveda et al. [25] design a Unity component that allows users to edit graphical structures on a 2D plane to generate reasonable building layouts based on room size and tightness constraints. Wu et al. [26] introduce a deep network for generating residential building plans. Given an architectural outline as input, the position of rooms and walls is predicted via the network and then converted into a vector graphic format. The network is trained on massive, densely annotated datasets of real residential buildings. The limitations of this approach are that it is difficult for the user to control the generated layout, and the cost of data acquisition and annotation is also high due to the large dataset. Hu et al. propose a learning framework for automatic floor plan generation, Graph2Plan [5], which takes the layout diagram provided by users as a sparse design constraint and combines it with a deep neural network to generate and model indoor room layout. However, Graph2Plan is an end-to-end model, which does not allow users to make adjustments during the design process. To enhance interaction between the model and designers at different levels, He et al. design a human-in-the-loop staged generation model called iPlan [19], which can automatically generate layouts while interacting with users, offering a high degree of flexibility. However, both Graph2Plan and iPlan require a large amount of data for training, and their rule boundaries are highly dependent. Compared with the method combining graphs and deep learning, this study focuses on the graph-guided heuristic algorithm, which can efficiently generate layouts that meet users’ needs without large amounts of data, providing greater flexibility and controllability.

2.2. Heuristic Algorithm-Based Scene Layout

The application of a heuristic algorithm in layout optimization has achieved remarkable results, especially in reducing computation overhead and improving generation efficiency. Hu et al. [27] optimize the layout of e-commerce warehouses with the aim of reducing picking costs. They construct attributes such as center point, length, and width for each layout entity, and then stretch and translate these layout entities as faces to complete the layout. By comparing Genetic Algorithms (GA) [9,10], Simulated Annealing (SA) [8], and Particle Swarm Optimization (PSO) [6,7], they validate the advantages of GA in layout optimization problems. Jalayer et al. [28] use a GA with multiple constraints as the optimization framework to automatically design cage layouts within animal shelters, minimizing the number of cages facing each other to reduce stress and ensure accessibility for staff and visitors. Garcia et al. [29] propose a multi-objective interactive genetic algorithm, which is effective in application scenarios such as sheep slaughterhouses and carton recycling plants. In this approach, the decision-maker’s knowledge contribution to the method guides a complex search process that is adjusted to the decision-maker’s preferences. Kumar et al. propose an Entity Planning Genetic Algorithm (EPGA) [10] that optimizes the layout design of buildings and houses by leveraging the entities structure topological relationships extracted from a Multi-Population Genetic Algorithm (MPGA). de Almeida et al. [30] propose an evolutionary approach that enables the automated design of mass custom production of modular homes and is not constrained by size and location constraints on the search space. Coelho et al. [11] optimize the number of light sources in indoor environments using GA to achieve adequate light intensity with the minimum number of sources. This study introduces topology awareness on the basis of heuristic algorithms, demonstrating higher computational efficiency, especially for layout problems with complex and irregular shapes.

2.3. Layout Optimization in Agriculture

The proper layout of farmland is the key to ensuring sustainable agricultural production [31]. Ju et al. [32] propose a genetic algorithm guided by support vector regression to solve the problem of layout optimization of farm stroke motors. Gao et al. [33] propose a supplementary sampling layout optimization method using the spatial simulated annealing algorithm to improve farmland management and obtain the optimal sampling layout. Ye et al. [34] propose the layout optimization of urban space and agricultural space based on spatial decision, determine the boundaries of different functional zones in agricultural space, and optimize the boundaries based on the principle of contradiction and coordination. Dadi et al. [35] believe that planning farm layout before starting agriculture can provide higher quality services, and adopt three facility planning methods to develop sustainable layout, namely, system layout design (SLP), pin-pair exchange method (PEM), and Graph-based theory (GBT). Satake et al. [36] compare the simulated annealing algorithm and the ramp climbing algorithm (HC) to optimize the layout of functional partitions such as cold storage, warehouse, packaging plant, and grading area in the agricultural facility site. In conclusion, optimizing the layout of functional zones in agriculture has always been a key issue, but it often faces challenges such as insufficient time efficiency and lack of interactivity during the design process. To address these issues, this study proposes a topology-aware heuristic optimization strategy and designs an interactive layout optimization system, aiming to provide a more intuitive and convenient user experience.

3. System Overview

Figure 1 presents an overview of our system. In the following, we provide an overview of each component in the system.

User Strategy. The system allows the user to draw and adjust the boundary sketch of the farmland scene. For functional zones, users can specify their location, quantity, and type. Our predefined functional zones include six types: planting area, seed planting area, storage area, office area, research area, and greenhouse.

2D Layout Generation and Optimization. The GGEL algorithm can quickly design the layout of the farmland scene through the topological perception of the graph. In addition, we design a layout adjustment module that allows users to visually modify the size and location of functional zones to meet customized needs.

3D Farmland Scene Generation and Edit. Based on the 2D layout, we also design a series of 3D procedural brushes, including plant brushes, construction brushes, and pond brushes, allowing users to adjust the details of the scene. Brushes can control the creation or destruction of single or multiple entities.

Agricultural Requirement and Spatial Constrains. The system’s design and optimization processes are governed by a comprehensive set of constraints, ensuring that generated layouts are spatially feasible and agriculturally viable. These constraints are derived from agronomic practices and spatial planning principles, acting as hard rules and soft objectives within our GGEL algorithm, specifically as follows: (1) Minimum Area Requirements: Each zone must satisfy a predefined minimum area to be functional; (2) the research area should be proximate to both the greenhouse and planting areas for monitoring and experimentation; (3) the greenhouse may require proximity to a water source; (4) Connectivity: All zones must be connected via a network of pathways for personnel and machinery access; (5) Shape Rules: Different areas within the boundaries hand-drawn by users all adopt regular rectangular shapes to promote efficient mechanized agricultural operations.

User Interface.Figure 2 presents the user interface we designed. The Scene Editor module in the main panel enables users to draw boundary sketches. The Key Parameter module displays the predefined functional areas and the coordinates of road nodes. The core module of the interface is Graph Mode, which presents the freehand boundary sketches drawn by users on the UI canvas. In the brush control panel, we offer multiple procedural 3D brushes with adjustable parameters (b).

The rest of this article is arranged as follows: Section 4 describes user input. Section 5 discusses 2D layout generation and optimization of farmland. Section 6 describes the 3D detail shaping of the farmland scene. Section 7 analyzes experimental results and user evaluations, and Section 8 presents conclusions and future work.

4. User Input

4.1. Freehand Boundary Sketch

The boundary drawn by users is represented as a series of continuous points $P=\{p_1,p_2,\dots ,p_n\}$, which are recorded sequentially when users draw. To address the inefficiency of directly using a large number of points, we use a sampling process to extract key points from the boundary, while effectively preserving the overall shape of the boundary. Our sampling algorithm is based on the principle that more points are sampled on segments with sharp bends and fewer points are sampled on smooth segments with low curvature. Therefore, we use curvature as the sampling weight. Given a boundary stroke composed of dense points P, we first compute the curvature ${\kappa }_i$ at each point $p_i$ through polynomial fitting. The curvature at a point $p_i$ is calculated as follows in Equation (1):

$${\kappa }_i={\displaystyle \frac{|f^{″}\left(p_i\right)|}{{\left(1+{\left(f^{\prime }\left(p_i\right)\right)}^2\right)}^{3/2}}}$$

(1)

where $f\left(p_i\right)$ is the polynomial function obtained by fitting at the points $p_i$. Then, from the starting point $p_1$ of the point set P, we use a variable move step to access the point set P in order to collect the key points. The moving step at point $p_i$ is calculated as follows in Equation (2):

$$\mathrm{\Delta }\left(p_i\right)=⌊k·{\displaystyle \frac{1}{\sqrt{1+\frac{1}{\rho }}}}+2⌋$$

(2)

where $\rho$ = $max({\kappa }_i,0.01)$. The default value of the constant k is set to 10, which keeps the step size within the range of [2, 11]. Finally, we use Hermite interpolation to fit the sampled boundary points, creating a smooth boundary curve connected end to end.

Based on the sampled vertices $P=\{p_1,p_2,\dots ,p_n\}$, we design a planar boundary optimization strategy using Laplacian deformation method, which allows users to adjust the freehand boundary by updating the position of sampling points in real time. We define Laplacian coordinates for each point, capturing the differential characteristics of the sampling point with respect to its neighbors. For each sampled point $p_i$, its Laplacian coordinate is specified in Equation (3):

$$\mathrm{\Delta }{p}_i=p_i-{\displaystyle \frac{1}{\left|\mathcal{N}\right(i\left)\right|}}\sum _{j\in \mathcal{N}\left(i\right)}{p}_j$$

(3)

where $\mathrm{\Delta }{p}_i$ is the Laplacian coordinate of point $p_i$, $\mathcal{N}\left(i\right)$ is the set of points adjacent to $p_i$, and $\left|\mathcal{N}\right(i\left)\right|$ is the number of neighbors.

Additionally, we design the following objective function to minimize the difference between the coordinates of the sampled points before and after optimization. The objective function is defined in Equation (4):

$$E\left(P^{\prime }\right)=\sum _{i=1}^n{∥\mathrm{\Delta }{p}_i^{\prime }-\mathrm{\Delta }{p}_i∥}^2+\sum _{k\in \mathcal{C}}{∥p_k^{\prime }-{\widehat{p}}_k∥}^2$$

(4)

where $\mathcal{C}$ denotes the subset of control points subject to user-defined adjustments, and ${\widehat{p}}_k$ corresponds to the designated new position for each control point $p_k$. The first term of the objective function, ${\sum }_{i=1}^n{∥\mathrm{\Delta }{p}_i^{\prime }-\mathrm{\Delta }{p}_i∥}^2$, is crucial for preserving the integrity of the Laplacian coordinates, thereby maintaining the intrinsic geometric properties of the original curve. The second term, ${\sum }_{k\in \mathcal{C}}{∥p_k^{\prime }-{\widehat{p}}_k∥}^2$, ensures that the modified positions of the control points align with the user-specified positions ${\widehat{p}}_k$, facilitating a controlled and precise optimization process.

4.2. Node Representation of Functional Zones

In order to simplify the layout design, each function zone is represented by a node. Users only need to mark nodes in a reasonable position to describe the distribution of function zones in the form of a graph and complete the preliminary layout planning. For each node, we encode two pieces of information: the zone type and the location. At the same time, we preset the zong constraints of each area type and the types of objects that may appear in the 3D scene to ensure the rationality and authenticity of the farmland scene.

5. 2D Layout Generation and Optimization

Since most of the zones in farmland have rectangular characteristics, we design a Graphics Guided Evolutionary Layout (GGEL) algorithm to better conform to the actual structure of the farmland scene. The algorithm flow is shown in Figure 3.

5.1. Node Encoding and Graph Construction

In order to clearly describe the spatial position of nodes, as shown in Figure 4, we use the convex hull algorithm to classify nodes into external nodes and internal nodes, and then encode the orientation of the external points.

We first encapsulate the freehand boundary and nodes in the bounding box and extract the four corner coordinates of the bounding box. Then, the Euclidean distance between the external node and each corner point is calculated separately, and the node closest to each corner point is selected as the reference point. After that, the lower left reference point is taken as the starting point, and the convex hull boundary is traversed clockwise: if the external node is the reference point, the node is coded directly, otherwise the coding of the previous reference point is inherited.

Finally, we use the Delaunay triangulation to connect the nodes and construct a graph structure. At the same time, we build an adjacency matrix to record the adjacency relationship between nodes.

5.2. Layout Spatial Pruning

As shown in Figure 5, we design the layout space pruning strategy according to the topological relationship of the graph nodes.

We first select two neighboring nodes of the target node and compute the normalized vector (red) ${\widehat{V}}_i={\displaystyle \frac{V_i}{\parallel V_i\parallel }}$ starting from the target node, respectively. The resultant vector is then used as an attraction (blue) ${\widehat{V}}_{\mathit{Guidance}}={\widehat{V}}_i+{\widehat{V}}_j$ to continuously expand the activity space of the target node until it touches other nodes, reaching the farthest range of activity. Combine all the active ranges to get the layout range of the target node (green). Compared with the whole inner region of the freehand boundary, the layout range after pruning is greatly reduced, which not only improves the calculation efficiency, but also reduces the problem of overlapping functional zones.

5.3. Conditional Constraint

In order to avoid gaps or overlaps between functional zones, we design two constraint modules and one optimization module to ensure the rationality of layout generation.

Coordinate Coupling. In the process of layout generation, we detect the point coordinates in the generated layout in real time, according to the orientation of the nodes being coded, so that the last node in each direction has the same x or y coordinates as the first node in the next direction, ensuring that there is no gap between the outside area and its neighbors.

Overlap Detection. We use geometric conditions to compare the corner coordinates of any two zones, preserve the layout without overlapping regions, and further optimize the layout structure.

Zone Alignment. In order to optimize the layout, we traverse the boundary corners of the function zone in turn. If the difference between different x or y coordinate values is less than the threshold value, the corresponding coordinate values are set to be the same, effectively eliminating the gap between the regions.

5.4. Fitness Functions

We design a suitable fitness function to evaluate whether the generated layout is reasonable. In the desired optimal solution, all zones should avoid overlapping, and the total area of each zone should be equal to the area of the entire region. The fitness function is designed to minimize the area error, and the closer the value is to 0, the better the quality of the solution. The fitness function is defined as shown in Equation (5):

$$Fitness={\omega }_1\times F_{OA}\left(B_i\right)+{\omega }_2\times F_{TA}\left(B_i\right)$$

(5)

where $\left\{B_i\right\}$ is the set of zones. $F_{OA}\left(\left\{B_i\right\}\right)$ is defined as the area where i blocks overlap each other. $F_{TA}\left(\left\{B_i\right\}\right)$ is the total area of i blocks. ${\omega }_1$ and ${\omega }_2$ are weight coefficients used to adjust the impact of overlap area and total area on fitness.

OverlapArea Fitness. Functional zones should not overlap, and the fitness function of the overlap area is defined in Equations (6)–(8):

$$F_{OA}\left(\left\{B_i\right\}\right)=\sum _{i=0}^n\sum _{j=i+1}^n{\displaystyle \frac{\mathit{Overlap}(B_i,B_j)}{\mathit{TotalArea}}}$$

(6)

$${\mathit{Overlap}}_1(B_i,B_j)=\left\{\begin{array}{c}(B_{j,\mathit{right}}-B_{i,\mathit{left}})(B_{i,\mathit{top}}-B_{j,\mathit{bottom}})\hfill \\ (B_{i,\mathit{right}}-B_{j,\mathit{left}})(B_{i,\mathit{top}}-B_{j,\mathit{bottom}})\hfill \\ (B_{i,\mathit{right}}-B_{j,\mathit{left}})(B_{j,\mathit{top}}-B_{i,\mathit{bottom}})\hfill \\ (B_{j,\mathit{right}}-B_{i,\mathit{left}})(B_{j,\mathit{top}}-B_{i,\mathit{bottom}})\hfill \end{array}\right.$$

(7)

$${\mathit{Overlap}}_2(B_i,B_j)=\left\{\begin{array}{c}(B_{i,\mathit{right}}-B_{i,\mathit{left}})(B_{i,\mathit{top}}-B_{j,\mathit{bottom}})\hfill \\ (B_{i,\mathit{right}}-B_{j,\mathit{left}})(B_{i,\mathit{top}}-B_{i,\mathit{bottom}})\hfill \\ (B_{i,\mathit{right}}-B_{i,\mathit{left}})(B_{i,\mathit{top}}-B_{j,\mathit{bottom}})\hfill \\ (B_{j,\mathit{right}}-B_{i,\mathit{left}})(B_{i,\mathit{top}}-B_{i,\mathit{bottom}})\hfill \end{array}\right.$$

(8)

where $\mathit{TotalArea}$ represents the area within the freehand boundary, $P_i$ represents the vertex of $B_i$, judging the overlap according to whether $P_i$ is inside $B_j$. ${\mathit{Overlap}}_1(B_i,B_j)$ means there is one vertex in $B_j$, and ${\mathit{Overlap}}_2(B_i,B_j)$ means there are two vertices in $B_j$.

TotalArea Fitness. Zones should be distributed as tightly as possible within the campus. Therefore, the total area fitness is defined as Equation (9):

$$F_{TA}\left(\left\{B_i\right\}\right)={\displaystyle \frac{\left|{\sum }_{i=0}^n\mathit{Area}\left(B_i\right)-\mathit{TotalArea}\right|}{\mathit{TotalArea}}}$$

(9)

5.5. Interactive Layout Edit

We design an interactive editing module, which can adjust the layout according to the programmatic generated results to further optimize the overall layout. As shown in Figure 6, users can edit road nodes independently to adjust the area. Specifically, we construct coupling matrices $C_x$ and $C_y$ for the x and y coordinates of road nodes, respectively, and assign the same labels to nodes with the same coordinate values. When a coordinate value changes, other coordinate values coupled with it in the x or y direction will also adapt to maintain the consistency of the layout. In addition, users can also edit the degree of curvature of regional boundaries to improve the authenticity of farmland scenes.

5.6. Experiments of EEGL

In order to prove the layout capability of EEGL in the farmland scene, we design experiments on the algorithm convergence speed and running time.

Contrast experiment for running time. We compare the operation time of EEGL, GA, and PSO under different numbers of zones. As shown in Figure 7a, when the number of zones is small, the operation time of EEGL and PSO is similar and lower than GA. With the increase of the number of zones, the running time of EEGL increases gently and stays below 10 s, while the running time of GA and PSO increases rapidly, indicating that EEGL has the best time efficiency in large-scale farmland scene layout tasks.

Ablation experiment for convergence speed. As shown in Figure 7b, we conduct ablation experiments on two modules of EEGL: overlap detection and spatial pruning. After removing the overlap detection module, the algorithm presents a downward trend similar to that of the complete EEGL in the initial stage, but in the later convergence process, the fitness value fails to reach the low point of the complete model, and the curve finally stabilizes at a higher fitness value. After the spatial pruning module is removed, the convergence speed of the algorithm in the early stage is significantly slowed down. Although the final fitness value can still reach a level similar to that of the complete EEGL, the convergence process is significantly prolonged. The ablation experiment results show that in the EEGL algorithm, the overlap detection module can ensure the rationality of the generated solution and reduce the unreasonable overlap between zones. The spatial pruning module greatly improves the computational efficiency and convergence speed of the algorithm. Multiple modules cooperate with each other and play an important role in the farmland scene layout task.

6. 3D Farmland Scene Generation

For different functional areas, we map the preset model to the actual layout to generate a preliminary farmland scene model. In addition, we have designed some procedural brushes that users can use to edit scene details and enhance the visualization of the model.

Plant Brush. We add four tree models to the system, allowing users to adjust the brush size and density according to their design needs, precisely controlling the distribution of vegetation.

Building Brush. There are three main buildings in the system, and users can adjust the position and orientation of individual buildings compared to plant brushes.

Pond Brush. The role of the pond brush is to change the height of the local terrain, and the user can freely adjust the size and force of the brush. In addition, we also provide a smoothness parameter to smooth terrain of different heights.

7. User Evaluation

7.1. Research Approach

The user study lasts approximately 1 h and is divided into three main parts.

System introduction. We introduce our system to each participant, and at the same time show some of the scene layouts generated by the system to give participants a clear visualization.

Task assignment. We assign each participant two tasks. Task 1: We present a farmland scene on Google Maps and ask participants to design a digital farmland scene based on the map scene. Task 2: We provide a farmland scene layout for participants to make creative designs, including but not limited to modifying boundaries, layouts, scene models, etc.

System feedback. Finally, we ask participants to complete a survey to collect their experience and feedback on using the system to assess the effectiveness of the system and user satisfaction.

We use a questionnaire to collect users’ subjective evaluations on three questions with 5-point Likert scale [where 1 = strongly disagree, 5 = strongly agree ] to evaluate the following: (1) users’ ability to successfully edit farmland scenes, (2) users’ satisfaction with the farmland scenes they created, and (3) users’ satisfaction with the real-time editing experience provided by our system.

7.2. Participant

This test involves 20 participants, aged between 20 and 50, including 13 males and 7 females. The participants come from diverse backgrounds, covering multiple professional fields. Specifically, 5 participants are graduate students in the field of computer graphics, while 4 are from non-graphics fields (including 2 from evolutionary computing and 2 from computer vision), 4 are experts in agricultural park planning, 5 are experts who are not only proficient in modeling but also possess a deep understanding of agriculture, and 2 are undergraduate students in computer science and technology. In the self-assessment of 3D modeling skills, 6 participants consider themselves novices, 6 are relatively familiar with 3D modeling, and 8 have extensive experience in 3D modeling, previously working on Unity Engine and Unreal Engine 4. Regarding the participants’ cognition of farmland scenes, 3 participants indicate that they have never seen farmland scenes, 6 have only learned about farmland scenes through videos and images, 7 have personally visited farmland scenes, and 4 are in charge of a farmland.

7.3. Research Result

In the first task, all participants successfully generated a digital farmland scene model based on a real scenario (Figure 8). In the second task, all participants used the system to edit their own field scenes (Figure 9). After all users complete the task, we collect the time spent by each user on each task. As first-time users of the system, they took slightly longer to complete, mainly due to differences in the level of refinement of the scene model. The results showed that the participants completed the first task in an average of 14 min and the second in an average of 6 min.

In the course of user research, we do not encounter any failures or unacceptable results. The qualitative evaluation shows that participants generally believe that our system has important practical value in the field scene design. Users also note that the system has a user-friendly interface and there is no delay during use. To further illustrate, we ask the following three questions and record participants’ responses: (1) “I can easily complete the task” (score: $\mu$ = 4.50); (2) “I am satisfied with the modeling results of the farmland scene” (score: $\mu$ = 4.45); (3) “I feel very creative” (score: $\mu$ = 4.75).

Table 1. User information survey and the time spent on each task.

User	Modeling	Agricultural	Task 1	Task 2	Q1	Q2	Q3
ID	Experience	Cognition	Time [min]	Time [min]
#1	1	1	14.8	8.1	4	4	4
#2	0	0	11.3	6.9	5	3	4
#3	1	2	12.7	5.4	5	4	5
#4	1	2	13.6	7.5	4	5	5
#5	0	0	18.3	6.3	4	3	5
#6	0	2	16.4	6.2	4	5	5
#7	1	2	17.9	7.2	5	5	5
#8	0	1	15.1	6.8	4	4	5
#9	0	0	17.6	4.9	3	3	4
#10	2	2	12.5	6.9	5	5	5
#11	0	1	13.6	4.1	4	5	4
#12	2	3	12.3	5.9	5	5	5
#13	3	1	11.8	4.9	5	5	5
#14	3	2	10.3	6.1	5	4	5
#15	2	3	13.7	6.4	5	5	5
#16	3	1	13.6	5.4	4	5	4
#17	2	1	12.3	5.3	5	4	5
#18	1	3	16.8	5.9	5	5	5
#19	1	3	14.1	7.1	4	5	5
#20	2	3	15.3	6.8	5	5	5
Avg	1.25	1.65	14.20	6.21	4.50	4.45	4.75
SD	1.07	1.04	2.28	1.00	0.61	0.76	0.44

also provides a detailed summary of each participant’s information and the amount of time they spent on each task.

Participants also share some additional tips. Agricultural park planning experts believe that our system effectively solves practical problems in current farmland scene planning in terms of efficiency and visualization, and are highly satisfied with the layout generation results and time efficiency. In addition, two graduate students from the field of evolutionary computing showed great interest in our point mapping mechanism. Although they have no exposure to computer graphics and 3D modeling, they believe that the graph guidance has significant innovation potential and application value, and suggest further exploration of more application scenarios in the field of farmland scene designing. These recommendations are expected to significantly enhance the user experience. Based on this valuable feedback, we plan to further improve the system’s flexibility and user satisfaction.

8. Conclusions

In this paper, a farmland scene layout design system is proposed which successfully realizes the scene generation by user input and graph guidance. The experimental results of several key performance indicators show that the proposed EEGL is superior to the mainstream heuristic algorithm in convergence speed and running time, and interactive editing modules shorten the user’s time from design to implementation and enhance the convenience of the overall workflow. In addition, the scene generation quality and user experience also received positive feedback, providing a high standard and diversified digital modeling solution for modern farmland scene.

Looking forward, future work will include conducting longitudinal real-world case studies, combining Internet of Things (IoT) and unmanned aerial vehicle (UAV) technology to quantify the economic return on investment (ROI), such as evaluating the reduction in design time, improvements in space utilization and potential yield increases, and the lowering of planning costs. We will also expand the model to incorporate key agronomic factors including soil properties, fertility conditions, and crop rotation patterns, thereby enhancing the practical relevance and decision-support capability of the system in diverse agricultural settings.