新疆农业大学低成本快速构建田间棉花3D点云评估冠层性状:UAV航拍141张影像约20分钟重建,效率优于SfM-MVS和Instant-NGP

1. Introduction

Cotton, as one of the most economically important crops worldwide [1], not only provides the core natural fiber raw material for the textile industry, but also serves as a significant oilseed and cash crop. Its industrial chain plays an irreplaceable supporting role in global rural employment and economic stability [2,3,4]. Xinjiang, as the major cotton-producing region in China, holds a strategically irreplaceable role in ensuring the supply security of the national cotton industry [5]. With the rapid advancement of agricultural mechanization and smart farming technologies, Xinjiang cotton has achieved a high degree of mechanization in key production processes such as sowing and harvesting, greatly improving production management efficiency and reducing labor dependence. However, the current analysis of canopy characteristics for field-grown cotton still relies primarily on traditional manual field survey methods, which suffer from high labor intensity, strong subjective bias in observation, limited sampling coverage, and insufficient spatial representativeness [6,7], making it difficult to achieve large-scale, refined, and time-series dynamic monitoring and quantitative evaluation of canopy characteristics at the field scale [8,9]. Therefore, it is necessary to develop efficient and scalable technical approaches for quantitative analysis and comprehensive evaluation of canopy structure and growth status in field-grown cotton populations, thereby providing scientific data support and technical foundations for growth diagnosis, cultivation regulation, and population optimization [10,11].

To overcome the limitations of the aforementioned manual surveys, the rapid development of UAV remote sensing technology and three-dimensional (3D) reconstruction methods in recent years has brought new technical opportunities for crop canopy analysis. By equipping UAVs with RGB cameras, multispectral sensors, or even LiDAR, researchers can acquire both two-dimensional spectral information and three-dimensional spatial structural information of the canopy, and subsequently estimate key phenotypic parameters such as plant height, leaf area index, biomass, and canopy porosity [12,13,14,15]. Among these, 3D point cloud data, owing to their ability to faithfully reconstruct the stereoscopic spatial configuration of the canopy, exhibit unique advantages in canopy vertical stratification analysis, light radiation interception assessment, population ventilation and light transmission diagnosis, and planting density optimisation [16]. In recent years, deep learning-based 3D reconstruction methods have made significant progress, among which Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have become research hotspots in this field [17,18]. NeRF can generate high-quality, photorealistic 3D reconstructions from sparse-view 2D images, demonstrating unique advantages in fine-scale plant structure modelling. 3DGS, on the other hand, explicitly represents scene geometry via Gaussian primitives, showing remarkable potential in reconstruction efficiency and scalability. These two types of methods have been applied in phenotypic studies of various crops, including wheat [19] and pepper [20], and have demonstrated promising application prospects in tasks such as plant organ segmentation, yield-related trait prediction, and large-scale field reconstruction [21]. However, current mainstream methods for field 3D data acquisition still face significant technical bottlenecks and engineering obstacles. On one hand, although LiDAR systems based on ground or UAV platforms can directly acquire high-precision 3D point clouds, they suffer from high equipment costs, large data volumes, and complex pre-processing workflows [22,23,24]. Moreover, under field conditions, the multiple-return signals of LiDAR are susceptible to canopy occlusion and wind-induced disturbances, leading to data redundancy and local information loss [25]. On the other hand, traditional point cloud generation methods based on Structure from Motion (SfM) require a large number of UAV aerial images and high overlap rates—typically hundreds or even thousands of images per experimental plot—resulting in lengthy reconstruction times and heavy computational resource consumption. Furthermore, under complex field illumination, mutual plant occlusion, and repetitive leaf texture conditions, these methods are prone to point cloud voids, excessive noise, and reduced geometric accuracy. These technical bottlenecks make it difficult for existing 3D canopy monitoring methods to achieve a balance among cost, efficiency, and accuracy, and thus they cannot truly meet the practical demands of rapid field monitoring, real-time diagnosis, and large-scale application [20,26,27].

To address the above issues, this study aims to explore a low-cost, high-efficiency, and few-image-input method for constructing 3D point clouds of field-grown cotton, and to carry out quantitative evaluation of canopy characteristics and population structure diagnosis based on high-quality point clouds. Using UAV aerial images as the sole data source, this study innovatively integrates the Splatfacto [28,29] reconstruction algorithm with the 3DGS-to-PC combined point cloud generation technique [30,31,32], while introducing ground control points (GCPs) placed in the field as geometric scale references. This effectively corrects scale drift during reconstruction, restores the generated cotton canopy point clouds to true physical dimensions, and ensures the accuracy and comparability of subsequent phenotypic parameter calculations. Based on this technical pathway, we systematically conduct research on the low-cost and rapid construction of 3D point clouds for field-grown cotton [33]. Compared with traditional methods, this technical pathway significantly reduces the reliance on the number of images and computational resources, achieving a technical breakthrough in reconstructing with fewer input images and high efficiency while ensuring point cloud reconstruction accuracy and completeness, greatly improving the acquisition efficiency and engineering practicality of 3D data for field cotton. On the basis of obtaining high-fidelity, true-scale 3D point clouds, we further carry out systematic quantitative evaluation of the canopy characteristics of cotton populations at the harvest stage. By extracting key canopy phenotypic indicators from point cloud data and performing quantitative calculations and comparative analyses, we systematically reveal the current structural characteristics, spatial heterogeneity patterns, and weak links of the field population canopy structure [34], and identify the core shortcomings and optimisation directions of the population structure. This provides scientific and reliable data support for precision cultivation management, canopy structure optimisation, and mechanisation-adapted planting in cotton [35]. Ultimately, the findings of this study will contribute to the sustainable development of the cotton industry towards high quality, large scale, and intelligence, providing a novel, scalable, and readily deployable technical paradigm for field crop phenotyping and population optimisation in the era of smart agriculture. Nonetheless, the present validation has been carried out only in a single season, one growth stage, and two experimental plots, and further validation involving multiple seasons and varieties is already underway to strengthen the applicability of these results.

2. Materials and Methods

2.1. 3D Reconstruction Workflow for Field-Grown Cotton

2.1.1. Data Acquisition

The data acquisition area was located at Huaxing Farm, Changji City, Changji Hui Autonomous Prefecture, Xinjiang Uygur Autonomous Region, with geographic coordinates of 44°02′ N, 87°30′ E, as shown in Figure 1. The acquisition area is located in a large-scale commercial field planting zone for industrial crops. The cotton variety used in this study was Zhongmian 113GK (National Certified Cotton Variety 20253014), bred by the Institute of Cotton Research, Chinese Academy of Agricultural Sciences. This variety is a conventional cultivar of upland cotton (Gossypium hirsutum L.) with a well-defined genetic background and stable transgenic insect-resistant traits. It exhibits a relatively open plant architecture, with clear spatial distribution patterns of branching configuration and leaf area. These genotype-determined morphological characteristics directly shape the canopy spatial structure and height uniformity, providing a validation scenario with a clear biological benchmark for 3D reconstruction algorithms. Irrigation, fertilisation, and other field management practices were implemented uniformly in accordance with local standardised large-scale production protocols to ensure the reproducibility of data acquisition conditions. In this study, two plots with dimensions of 5.5 m × 5.5 m each were established, and each plot contained six standardized cotton rows with a row spacing of 0.6 m and a plant spacing of 0.1 m. The data acquisition campaign was conducted on 25 October 2025, during the cotton harvest period under clear skies, no precipitation, and stable illumination conditions, effectively ruling out the interference of meteorological variability on data quality. The average wind speed from ground level to take-off altitude was approximately 1 m/s (corresponding to Beaufort Force 1), falling within the calm to light air range. Despite the low wind speed, the cotton canopy leaves and bolls exhibited continuous slight oscillation, and since the 3D Gaussian Splatting (3DGS) reconstruction algorithm strictly relies on the static scene assumption, wind-induced deformation constituted a potential factor affecting reconstruction accuracy. The current algorithm does not explicitly model wind-induced deformation; instead, the acquisition strategy relied on two mitigation measures. First, the extremely short single exposure time (1/2000 s) ensured that leaf displacement within this integration period was far smaller than the ground sampling distance, rendering motion blur within individual frames negligible. Second, the high overlap rate and short shooting intervals resulted in a baseline distance of approximately 0.5 m between adjacent along-track images, with each spatial point being covered by multiple overlapping views. The shooting interval between adjacent images was 1 s. If leaf positions exhibited observable changes due to wind during this interval, the 60% overlap rate ensured that every spatial point was recorded from a large number of views at different instants. The random positional deviations among views caused by wind were effectively averaged out through multi-constraint joint adjustment during the estimation of 3D Gaussian parameters, thereby mitigating their systematic impact on canopy structure reconstruction. A total of 141 RGB images were acquired in this campaign, all collected within a single day, ensuring the timeliness and spatial consistency of the dataset.

Figure 1. Overview of image data acquisition for field-grown cotton. (A) Location map of the sampling site. (B) Photograph of the field experimental acquisition area. (C) Multi-view images of field-grown cotton acquired by the UAV.

2.1.2. UAV Flight Route Planning and Design

In this study, a DJI Mavic 3M UAV (DJI, Shenzhen, China) was employed for field image acquisition of field-grown cotton. The flight height was set to 5 m above ground level, and the flight speed was set to 0.5 m/s, which served as the core operational parameters for flight route planning. All acquired images were saved in JPG format, with a minimum resolution of 5280 × 3956 pixels per image. Each image was accompanied by GPS positioning, flight attitude, camera parameters, and absolute flight altitude information, providing comprehensive data support for high-precision 3D reconstruction of cotton from multiple viewpoints. A grid-based flight route was specifically designed for data acquisition in this study, and the route layout is illustrated in Figure 2. Although this grid route does not possess the capability of the CCO (crop canopy-oriented) route to penetrate the occlusion of upper canopy leaves and fully capture the detailed images of middle and lower plant organs [36,37], our study aimed to construct point clouds of field-grown cotton rapidly and efficiently. Therefore, the grid route allowed for the rapid acquisition of images from four viewing angles with substantially higher collection efficiency. The camera inclination angle was set to 45°, with both heading and side overlap ratios set to 60%, and the shooting interval was 1 s, thereby enabling automatic temporal continuity and precise spatial alignment.

Figure 2. Design of the UAV-based flight route for data acquisition. (A) Flight trajectory of the grid route. (B) Distribution of image capture points along the grid route.

2.1.3. GCP-Based Georeferencing and Scale Calibration for 3D Canopy Reconstruction

The acquired aerial images were imported into Agisoft Metashape Professional Software (Version 2.2.2) [38] to complete the full photogrammetric processing workflow. This software, based on the Structure from Motion (SfM) algorithm, performs image stitching and registration to construct a sparse 3D point cloud model of field-grown cotton. In this study, ground control points (GCPs) were placed at the four corners of the acquisition area as georeferencing references for photogrammetric processing. The GCPs provide absolute control for the UAV images, effectively correcting geometric distortions caused by flight attitude variations, camera lens distortion, and GPS positioning errors, eliminating cumulative drift errors generated during the aerial triangulation process of 3D reconstruction, and significantly improving the absolute positioning accuracy of the model in both planar and elevational directions. More critically, the GCPs also serve a core function in establishing the true physical dimensions for the reconstructed model. During the SfM processing stage, the GCPs deployed in this study have already enabled georeferencing and absolute scale calibration of the imagery. Through bundle adjustment optimisation, a sparse point cloud model with true physical dimensions and precise camera pose information was obtained. However, the subsequent Splatfacto reconstruction model, whose input consists of normalised camera poses and image features, outputs the scene’s 3D structure in a normalised representation, the reconstructed result is constrained within a normalised coordinate space, where geometric parameters including distances and volumes are expressed in relative proportions, inherently lacking a scale correspondence to the real physical world. Consequently, although the dense point cloud produced by Splatfacto reconstruction exhibits superior local geometric details and completeness compared to the SfM sparse point cloud, its spatial scale has deviated from true physical dimensions. To address this scale loss problem, this study leverages the precisely measured GCP coordinates obtained during the SfM stage to establish a correspondence between the GCPs in the normalised point cloud space and their true physical dimensions. Specifically, based on the known physical size of the GCP markers (80 cm × 80 cm) and the distances between their corresponding corner points in the Splatfacto-reconstructed point cloud, a global scale factor between the normalised space and the real physical space is calculated. This scale factor is then applied to the entire point cloud, transforming it from normalised coordinates to true physical coordinates, thereby restoring the geometric metrics such as distances, areas, and volumes to their actual field values. Through this GCP-based post-calibration strategy, the final cotton canopy 3D point clouds not only preserve the high-precision geometric details offered by Splatfacto, but also possess a dimensionally consistent basis that matches the actual field scale, providing a reliable scale reference for subsequent accurate quantitative analyses of phenotypic parameters such as canopy height, volume fraction, and porosity [39,40,41]. Moreover, the large 80 cm × 80 cm size significantly enhances the detectability of GCPs in UAV aerial imagery, ensuring stable and repeatable corner detection even under challenging conditions such as high flight altitudes, varying illumination, or partial canopy occlusion. This effectively reduces uncertainties in scale recovery and georeferencing, further improving the robustness and engineering practicality of the overall technical workflow. The GCP coordinates were collected using a Qianxun SR3 Pro RTK receiver (Qianxun Spatial Intelligence Inc., Shanghai, China), and the layout of the control points is shown in [Figure 3](https://www.mdpi.com/2073-4395/16/17/1