Super-Resolution Point Cloud Completion for Large-Scale Missing Data in Cotton Leaves

Abstract

Point cloud completion for cotton leaves is critical for accurately reconstructing complete shapes from sparse and significantly incomplete data. Traditional methods typically assume small missing ratios (≤25%), which limits their effectiveness for morphologically complex cotton leaves with severe sparsity (50–75%), large geometric distortions, and extensive point loss. To overcome these challenges, we introduce an end-to-end neural network that combines PF-Net and PointNet++ to effectively reconstruct dense, uniform point clouds from incomplete inputs. The model initially uses a multiresolution encoder to extract multiscale features from locally incomplete point clouds at different resolutions. By capturing both low-level and high-level attributes, these features significantly enhance the network’s ability to represent semantic content and geometric structure. Next, a point pyramid decoder generates missing point clouds hierarchically from layers at different depths, effectively reconstructing the fine details of the original structure. PointNet++ is then used to fuse and reshape the incomplete input point clouds with the generated missing points, yielding a fully reconstructed and uniformly distributed point cloud. To ensure effective task completion at different training stages, a loss function freezing strategy is employed, optimizing the network’s performance throughout the training process. Experimental evaluation on the cotton leaf dataset demonstrated that the proposed model outperformed PF-Net, reducing the Chamfer distance by 80.15% and the Earth Mover distance by 54.35%. These improvements underscore the model’s robustness in reconstructing sparse point clouds for precise agricultural phenotyping.

1. Introduction

Xinjiang cotton is prized for its exceptional fibre purity, strength, softness, and breathability, making the region China’s foremost cotton-producing area [1]. Owing to its extensive cultivation, high yield, and premium quality, cotton has become a pillar of Xinjiang’s economy and one of the nation’s most valuable cash crops [2,3]. However, data from agricultural research institutions revealed a 17.6% decline in China’s cotton yarn exports in 2023, indicating a broader drop in domestic lint production and a weakening of global competitiveness [4,5]. Because growth conditions and phenotypic traits directly determine lint yield and quality, real-time monitoring of leaf size and automated phenotypic analysis via 3D point clouds are essential for improving production outcomes [6,7]. Nevertheless, occlusion, specular reflection, limited sensor resolution, and constrained viewpoints often produce incomplete point clouds that omit critical geometric and semantic information, highlighting the urgent need for robust completion techniques tailored to cotton phenotyping [8].

These challenges have prompted researchers in plant phenotyping to seek effective solutions; consequently, considerable efforts have focused on developing point-cloud completion and super-resolution techniques to overcome incomplete and sparse data. Initial research in this area was dominated by geometric algorithms that utilized hand-crafted features, symmetry assumptions, and surface-smoothness constraints to interpolate missing regions or retrieve similar patches from 3D databases [9,10,11,12]. Jiang et al. [13] restored stem topological continuity through L1-median skeleton extraction. Xie et al. [14] enhanced oilseed rape silique reconstruction via composite feature point registration. Chen et al. [15] filled missing leaf surfaces via Harris corner detection. Although these classical methods, which are primarily based on composite feature registration and corner detection, facilitate 3D reconstruction in certain plant species, they remain inadequate for addressing extensive sparsity or irregular structural variation due to their dependence on manually designed features.

These limitations have catalyzed a shift toward deep learning approaches, which learn permutation-invariant representations directly from raw point sets and can model complex structures end to end. The introduction of neural architectures capable of learning permutation-invariant features directly from unordered point sets [16] laid the groundwork for data-driven point cloud analysis. Building on this foundation, researchers have developed various deep learning-based completion methods tailored to plant morphology. Multiscale pyramidal decoding approaches have been shown to restore complete leaf geometry and enhance the global topological structure, although they often fail to fully recover fine boundary details [17]. Transformer-based semantic encoding methods increase organ-level reconstruction accuracy but exhibit heightened sensitivity when confronted with extremely sparse data [18]. Encoder–decoder frameworks for biomass estimation are effective in reconstructing multiplant scenes; however, they still struggle to achieve high-fidelity restoration of fine local structures [19]. These advances collectively enhanced the reconstruction of both the global topology and local details, overcoming the limitations of traditional hand-crafted methods.

Recent advances in point cloud completion have yielded a series of influential deep learning algorithms, each introducing distinct strategies to address specific limitations of incomplete data. GRNet [20] leverages a gridding structure and 3D convolutional operations to learn global geometric features, which enhances its ability to recover the overall coarse structure of objects with substantial missing regions. However, this global approach often leads to uneven point distributions and limited restoration of fine details in complex morphologies. PF-Net [21] introduces a progressive folding mechanism to iteratively refine point locations, aiming to improve both the accuracy and continuity of completed shapes. Despite better surface refinement, PF-Net may still suffer from oversmoothing and insufficient boundary fidelity, especially when the input data are highly sparse. To address the need for detailed boundary preservation, PMP-Net [22] and its improved version, PMP-Net++ [23], implement point-mixing and progressive point generation strategies. These models excel at capturing local contextual information and reconstructing sharp geometric features, which is particularly beneficial for restoring the intricate contours of plant leaves. Nevertheless, their focus on boundary detail can introduce issues such as internal voids, point clustering, or instability in severely incomplete regions. SnowflakeNet [24] further improves completion performance through the use of a hierarchical snowflake-shaped expansion module, which employs point expansion from coarse to fine to achieve a balance between structural completeness and detail. However, it is prone to layering artefacts and sampling irregularities, which can compromise its morphological fidelity. These methods demonstrate clear progress in point cloud completion, but each retains characteristic limitations in global uniformity, local detail preservation, or robustness to severe sparsity. This underscores the necessity for integrated approaches capable of jointly optimizing geometric completeness, distribution uniformity, and boundary fidelity in complex plant phenotyping scenarios.

In addition to geometric completion, point clouds require effective densification, as severe sparsity can degrade phenotypic measurements even when shapes are fully recovered. Deep learning-based super-resolution addresses this challenge by converting low-resolution point sets into dense, structurally consistent outputs, thereby increasing geometric fidelity, visual realism, and downstream analytic accuracy [25]. Traditional super-resolution techniques have employed interpolation-based strategies [26], statistical filtering methods [27], and geometric or topological algorithms [28] to upsample sparse point sets. While approaches such as Gaussian filtering, kriging interpolation, Laplacian smoothing [29], and IsoMap [30] have achieved varying degrees of success, they often suffer from excessive computational cost or loss of fine-scale structural fidelity in complex, large-scale scenes.

In contrast, recent learning-based super-resolution frameworks leverage hierarchical feature encoding to infer latent high-resolution representations, enabling robust point-density enhancement while preserving local morphology. Inspired by progress in image super-resolution, neural architectures such as the SRCNN [31], VDSR [32], and ESPCN [33] have been extended to 3D point clouds, enabling more accurate recovery of dense structures from sparse data. These models typically employ end-to-end frameworks with hierarchical encoders to capture both global shape information and local geometric features. Current deep learning-based super-resolution methods can be broadly categorized into three technical directions. Feature-enriched encoders, including multibranch MLPs and PointNet++ pyramids, enhance multiscale information extraction and support high-resolution point set reconstruction [34,35]. Adversarial and graph-based frameworks integrate graph convolution, generative adversarial training, and residual connections to accelerate convergence and improve structural realism [36]. Attention-driven and dilated-convolution modules utilize self-attention, dilated kernels, or local CNN blocks to refine the point distribution and detail fidelity, particularly under extreme sparsity [37,38,39]. However, these approaches remain fragmented, and no single framework has yet achieved a comprehensive balance among density enhancement, geometric completeness, and point distribution uniformity, all of which are essential criteria for reliable cotton leaf phenotyping.

Unlike interpolation or voxel-based densification techniques, deep learning models are uniquely capable of modelling complex spatial relationships and generating semantically coherent distributions, even in the presence of extreme sparsity or occlusion. Most existing super-resolution models continue to focus on isolated tasks, often prioritizing apparent density or overall plausibility while neglecting fine-scale morphological accuracy or distribution regularity. The absence of unified frameworks that jointly optimize all critical aspects underscores a persistent gap in current research and highlights the urgent need for more robust and comprehensive solutions for point cloud densification in plant phenotyping applications.

In response to these challenges, this study introduces a super-resolution point cloud completion framework tailored for highly sparse, nonuniform, and geometrically complex cotton leaf data. The proposed method incorporates five key modules within a unified end-to-end architecture. The design begins with a PointNet++ encoder that extracts global geometric features to capture the overall shape information. A decoder then integrates bilinear interpolation with the PF-Net algorithm to reconstruct coarse geometry, followed by an upsampling module that generates additional points in regions of data loss, thereby enhancing local density and detail recovery. The feature-fusion module subsequently merges local and global representations into deconvolutional feature maps, further improving the structural detail of the reconstructed point cloud. A convolutional completion network refines the output, whereas farthest-point sampling ensures that the final point cloud is of high quality and uniformly distributed, meeting the requirements of downstream phenotypic analysis. The entire training pipeline is further optimized via a multiloss freezing scheme, and the inclusion of an attention-augmented PointNet++ module enhances point cloud uniformity throughout the reconstruction process. The main contributions of this paper are as follows:

I. Super-resolution completion for sparse data: A novel completion network reconstructs highly sparse cotton-leaf point clouds. Bilinear-interpolation upsampling improves local feature capture, restores missing regions, and yields accurate leaf-area representations.

II. Uniformity module based on PointNet++: A PointNet++ feature-fusion block merges incomplete inputs with reconstructed points, equalizes the point distribution and improves the geometric consistency.

III. Stagewise loss-freezing strategy: During training, early layers are frozen under CD loss to stabilize global reconstruction; later layers optimize a feature-fusion loss to refine local detail and maintain uniform density, leading to faster convergence and higher fidelity.

2. Materials and Methods

2.1. Cotton Leaf Dataset

Owing to the absence of point cloud data for cotton plant leaves in public datasets, this study collected data from 96 cotton plants over one year in Shier Tuan, Alar city, Xinjiang. The dataset spans five growth stages: sowing and seedling, vegetative, bud, flowering and boll formation, and boll opening. We adopt a stratified split by plant ID and growth stage: 8:1:1 for train: validation: test. Sampling was conducted in adjacent cotton fields with consistent water and fertilizer conditions, providing a robust foundation for future studies on yield and quality estimation. To minimize the effects of uneven lighting and wind-induced motion on image quality and 3D reconstruction, all the experiments were conducted indoors under controlled conditions. The laboratory setup included four soft light lamps to ensure uniform lighting, a black backdrop beneath the plants to eliminate background interference, and curtains to block external light sources. Each plant was photographed sequentially from three angles: approximately 45° downwards, horizontal, and 75° upwards, resulting in approximately 260 images per plant.

The 260 images for each sample are organized by plant ID to maintain image-to-sample correspondence. All plants were photographed with a Canon EOS 200D II camera (Canon Inc., Urumqi, China) at a native resolution of 4000 × 6000 pixels (24 MP). As shown in Figure 1, we employ multiview stereo using RealityCapture (version 1.2.0.11881, Capturing Reality, Bratislava, Slovakia) to align the images, generate a sparse point cloud, and subsequently perform dense reconstruction. This process yields high-fidelity 3D models of entire cotton plants. Although CloudCompare (https://cloudcompare.org/index.html, accessed on 18 September 2025) supports purely manual selection, we adopted a semi-automatic workflow with manual verification following dense plant reconstruction. The plant cloud was first denoised using statistical outlier removal (SOR) and smoothed with moving least squares (MLS). We then computed per-point normals and curvature/roughness scalar fields to guide segmentation. Leaves were coarsely isolated by thresholding these fields and applying connected components; overlapping organs were further separated via region growing (using normal/curvature cues) and interactively refined with the Segmentation tool, consulting RGB textures where boundaries were ambiguous. Verified leaves were exported as individual PLY files.

To construct the dataset required for this study, 3D point cloud data of cotton leaves were extracted and uniformly sampled via farthest point sampling (FPS). Each model was sampled to contain 2048 points. Through leaf segmentation, a total of 6000 cotton leaf point clouds were obtained. Point cloud data were extracted from multiview images, and each leaf was downsampled to 2048 points via FPS. For point clouds containing fewer than 2048 points, upsampling was performed by duplicating neighboring points to achieve the required point count. CloudCompare was used for segmentation and annotation. For training and evaluation, each leaf cloud was denoised, normalized to [−1,1], and converted to a canonical ground truth (GT) of 2048 points using deterministic FPS (fixed seed for reproducibility): iteratively select the point with the largest minimum Euclidean distance to the already chosen set, ensuring uniform coverage. Rare leaves with fewer than 2048 points were first completed to 2048 by nearest-neighbor duplication and then FPS-refined. The resulting 2048-point GT serves as supervision for all losses and as the reference for metrics (CD/EMD/F-score).

To emulate acquisition failures, we derive incomplete inputs from the GT leaves at 50% and 75% missing using two mask families. (i) Spherical removal (local-region erasure): one or several balls are sampled with centers on the leaf point set; radii are adjusted so that the retained-point ratio matches the target. Points whose Euclidean distance to any center is below its radius are removed and overlapping balls form contiguous holes. (ii) Axial cut (one-sided half-space): a cutting plane, whose normal is aligned with the leaf’s principal/midrib axis, removes all points on one side (“one cut”). For reproducibility, we store all mask parameters (centers/radii or plane coefficients) and random seeds. When a method requires a fixed input size, sparse inputs are upsampled to 2048 by nearest-neighbor duplication; at evaluation, both prediction and GT are FPS-downsampled to 2048 before computing CD/EMD/F-score under a unified protocol.

2.2. Super-Resolution Point Cloud Completion Network

Owing to the inherent sparsity and incompleteness of the obtained cotton point cloud data, this model focuses on upsampling and feature representation of cotton leaf point clouds as key technical research. This project is based on the PF-Net and PointNet++ models, focusing on addressing the issue of reduced density accuracy caused by point cloud incompleteness and constructing a network model with both completion and densification capabilities. Figure 2 illustrates the overall framework of the project. To address the insufficient feature extraction from sparse point clouds, bilinear interpolation is used to upsample the original data. The point cloud reconstruction process involves the fusion of complementary and missing point clouds, while also achieving a uniform distribution of the point cloud. Finally, by freezing the loss function, the complementary and plasticity tasks are completed simultaneously during the end-to-end network training.

2.2.1. Point Cloud Completion

To achieve accurate and uniform reconstruction of sparse cotton leaf point clouds, this study proposes a PF-Net-based completion framework that integrates multiresolution feature extraction and hierarchical decoding, as illustrated in Figure 3. The primary objective is to address limitations in traditional feature extraction and enhance the uniformity and geometric fidelity of the reconstructed point clouds.

Bilinear Interpolation and Iterative Upsampling: Let $X\in R^{B\times N\times 3}$ denote a batch of sparse cotton-leaf point clouds (coordinates normalized to [−1,1]). To provide a dense scaffold for completion, we perform bilinear-interpolation upsampling on local tangent patches and obtain a seed set $X_k\in R^{B\times K\times 3}$ with $K=kN$ (we fix $K=2048$ in all experiments), as shown in Figure 3a. In local parametric coordinates $(x,y)\in {\left[0,1\right]}^2$, bilinear interpolation is given by:

$$f(x,y)=(1-x)(1-y)f_{00}+x(1-y)f_{10}+(1-x)y{f}_{01}+xy{f}_{11}$$

(1)

Here $f_{00}$, $f_{10}$, $f_{01}$, $f_{11}$ are the values at the four corners $(0,0)$, $(1,0)$, $(0,1)$, $(1,1)$; the weights sum to 1.

Specifically, during the upsampling process, a nearest-neighbor search identifies a set of surrounding points for each reference point. These neighbors are used as both references and supplements, ultimately improving the precision of the upsampling step. Let $D_i$ represent the set of distances from point $P_i$ to all other points. The nearest neighbors of $P_i$, denoted as $N_i$, are selected from $D_i$ based on the minimum distance criterion. For each point $P_i$, the average position ${\overline{F}}_i$ of its neighboring points is computed, reflecting the latent structural information of the point cloud. Subsequently, a directional vector ${\overline{\mathrm{v}}}_i$ is calculated from the reference point $P_i$ to the mean position ${\overline{F}}_i$, which describes the trend or direction of the variation in the surrounding point cloud. This vector is then utilized to generate new points $P_i^{\prime }$ along the direction of ${\overline{\mathrm{v}}}_i$, with one new point being the default, which are sequentially added on the basis of the directional vector. This procedure is iterated for each point, continuously identifying neighboring point sets and generating new points. The entire point cloud is processed in this manner, resulting in a modified set of points. Through the iterative adjustment of each point, analogous modifications and enhancements are applied to the entire point cloud.

Encoding and Multiresolution Feature Extraction: In the encoding stage, the MRE processes an unordered input point cloud of size $N$, as illustrated in Figure 3b. Bilinear interpolation is applied to upsample the initial data, generating additional scales with $(k-1)N$ and $kN$ points. These multiscale representations enable the encoder to focus on both global and local features. Three distinct MLPs are subsequently utilized to fuse these scales into a comprehensive latent feature vector, facilitating the extraction of rich geometric and semantic information.

Hierarchical Decoding and Point Cloud Refinement: The latent feature vector is then input into the PPD, which hierarchically upsamples and reconstructs the missing data through three levels of linear-convolutional layers, as illustrated in Figure 3c. At each decoding stage, the output from the previous layer is refined and upsampled, progressively generating new points and restoring the detailed structure of the original cloud. This multistage decoding mechanism enhances both the global topology and local geometric details, such as sharp edges and intricate surfaces. The outputs $\alpha$, $\beta$, and $\gamma$ from each level are fused to produce the final, high-resolution point cloud of size $kN$.

Multiscale fusion at the output layer: Let the sparse input be $X\in R^{B\times K\times 3}$ (coordinates normalized to [−1,1]). After bilinear-interpolation upsampling, we obtain the seed set $X_K\in R^{B\times K\times 3}$ with $K=kN$ (fixed $K=2048$ in all experiments). Three parallel multiscale branches operate on the same $K$ seeds and do not generate independent point clouds; instead, each branch predicts an index-aligned residual offset $\Delta Y^{(s)}\in R^{B\times K\times 3}$ for $s\in \{1,2,3\}$. For every seed point we also learn non-negative weights $\alpha ,\beta ,\gamma \in R^{B\times K\times 1}$, normalized by softmax so that $\alpha +\beta +\gamma =1$. The final reconstruction is obtained by residual fusion:

$$Y=X_k+\alpha \odot \Delta Y^{(1)}+\beta \odot \Delta Y^{(2)}+\gamma \odot \Delta Y^{(3)}\hspace{1em}\hspace{1em}(Y\in R^{B\times K\times 3})$$

(2)

where $\odot$ denotes element-wise (per-point) scaling and broadcasting. Because $\Delta Y^{(s)}$ are index-aligned with $X_k$, the fusion is computed point-wise rather than by concatenating/unioning three sets. When $K\ne 2048$, we FPS-downsample $Y$ to 2048 before computing CD/EMD/F-score; when $K=2048$, no further resizing is required.

2.2.2. Point Cloud Reshaping

The primary objective of this stage is to achieve a complete and uniformly distributed 3D point cloud model suitable for accurate phenotypic analysis. While the preceding completion module reconstructs missing regions within the point cloud, the distribution of points may still be uneven, which can hinder downstream processing and analysis.

To address this, the output of the completion algorithm is first integrated with the original point cloud, forming a unified point cloud set. PointNet++ is employed at this stage to facilitate effective fusion of the generated and original data, leveraging its strong capability to extract both local and global geometric features. Precise alignment strategies are then applied to ensure spatial consistency between the two point clouds, preventing any overlap or misalignment. We term the subsequent post-fusion distribution normalization Reshape. After completion, we denote the prediction by $Y$ and the original input by $X$ (both normalized). Reshape produces a uniformly distributed, boundary-preserving point set used for evaluation and analysis: $\tilde{Y}=R(Y,X)$. Concretely, the merged cloud is first deduplicated using a radius criterion and then rigidly aligned to ensure fine spatial consistency; local densities are equalized by iteratively dispersing overly clustered points. When necessary, small gaps are interpolated, isolated outliers are lightly denoised, and the final set is brought to a fixed cardinality ($K=2048$) via deterministic farthest-point sampling. The resulting cloud achieves uniform spatial sampling while preserving boundaries and details, providing a robust foundation for subsequent phenotypic analysis.

To rigorously evaluate the model’s ability to handle incomplete data, additional experiments are performed on synthetic partially missing point clouds. Specifically, the input data are first aligned to the center of the coordinate system and normalized to fit within the range of [−1, 1]. Each shape is then uniformly sampled with 1024 points to capture its intrinsic geometric features. To simulate missing regions, a random viewpoint is selected as the center, and all points within a specified radius are removed. This controlled generation of incomplete point clouds enables systematic assessment of the completion and Reshape modules under realistic conditions, highlighting their effectiveness in restoring both structure and uniformity. The degree of missing data can be controlled by adjusting the removal radius. PointNet++ is employed to achieve the fusion and uniformization of the missing point cloud $P_m$ and the complementary point cloud $P_c$. The fusion process is defined as the union of these two points, described as follows:

$$P_f=P_m\mathrm{U}{P}_c$$

(3)

PointNet++ processes $P_f$ by extracting features from each point $P_i$ via an MLP. It then aggregates local and global features to generate a unified representation, denoted as $f(P_f)$. The process can be described as follows:

$$\left\{\begin{array}{c}{f}_i=MLP(P_i) ,\\ f(P_f)=PointNet++(P_f)\end{array}\right.$$

(4)

Here, $Po\mathrm{int}Net++(P_f)$ represents the PointNet++ network used for fusing the point cloud $f(P_f)$. The resulting feature vector $f(P_f)$ encapsulates a homogenized representation of the fused point cloud, facilitating subsequent analysis or processing tasks.

2.2.3. Loss Freeze Program

This study employs a loss function freezing strategy to address the critical tasks of point cloud completion and reshaping for cotton leaf data. The loss functions consist of the CD loss ($L_{CD}$) and the reshaping loss ($L_{Res}$). $L_{CD}$ ensures the uniform distribution of upsampled points across the target surface, while $L_{Res}$ maintains consistent local structures for points within the same region. To ensure training accuracy, distinct freezing schemes are applied to these loss functions. During the first 15 h of training, $L_{Res}$ remains constant to enhance learning for point cloud completion and super-resolution. Between the 15th and 80th epochs, the weight of $L_{CD}$ is reduced to 0.5, while $L_{Res}$ is assigned an equal weight of 0.5. After 80 epochs, $L_{CD}$ is frozen, allowing the network to focus entirely on fusion and morphological reshaping tasks. This adaptive weighting strategy ensures efficient optimization and precise task-specific learning throughout the training process.

The CD loss function quantifies the difference between two points by calculating the average distance from each point in one cloud to its nearest point in the other cloud. This effectively guides the network in learning features, ensuring the precision and uniformity of the generated point clouds. The CD loss function formula used in this study is defined as follows:

$$L_{CD}(P,S)={\displaystyle \frac{1}{\left|P\right|}}{\displaystyle \sum _{p\in P}\underset{s\in S}{\min }}{‖p-s‖}_2^2+{\displaystyle \frac{1}{\left|S\right|}}{\displaystyle \sum _{s\in S}\underset{p\in P}{\min }}{‖p-s‖}_2^2$$

(5)

To improve the control of point cloud fusion and uniform distribution, this study introduces a reshaping loss to constrain the distribution process. The reshaping step in the model builds upon the method used in PointNet++ and incorporates the application of a feature transformation matrix. To ensure that this matrix approximates an orthogonal matrix, a regularization term is added to the SoftMax training loss. The reshaping loss is defined as follows:

$$L_{Res}={‖I-A{A}^T‖}_F^2$$

(6)

where $A$ represents the feature alignment matrix predicted by the network. The incorporation of orthogonal transformations helps preserve input information, ensuring the integrity of the feature transformation process. The regularization term enhances optimization stability while improving the overall performance of the model.

This strategy enables task-specific loss functions to guide network learning at different training stages, effectively avoiding gradient conflicts while improving convergence speed and training stability. The experimental results demonstrate that the proposed network achieves outstanding performance in both point cloud completion and reshaping tasks for cotton leaves, further validating the effectiveness and applicability of the loss function freezing strategy in multitask deep learning networks.

2.3. Evaluation Measures

2.3.1. Chamfer Distance

This study utilizes the chamfer distance (CD) as the primary evaluation metric for the cotton leaf dataset. CD is widely used in point cloud reconstruction and generation tasks and serves as an effective measure of similarity between two points. A higher CD value between the denoised point cloud and the reference point cloud indicates poor completion performance, while a lower CD value signifies better results. The metric evaluates the difference between the completed and ground-truth point clouds, with values approaching zero indicating greater shape similarity and superior completion quality.

In addition to CD, we report per-point $L_1$ and $L_2$ CDs to provide complementary error profiles. The $L_1$ CD is defined as $L_1={\displaystyle \sum {}_i\left|Q_i-{\widehat{Q}}_i\right|}$, where $Q_i$ and ${\widehat{Q}}_i$ denote corresponding points in the reference and predicted clouds, respectively. $L_1$ aggregates absolute coordinate differences, treating all deviations uniformly. The $L_2$ CD is formulated as $L_2=\sqrt{{\displaystyle \sum {}_i(Q_i-{\widehat{Q}}_i)^2}}$. Because the $L_2$ norm squares deviations before aggregation, it places greater emphasis on large errors than the $L_1$ norm does, making it particularly sensitive to substantial discrepancies between point clouds.

2.3.2. Earth Mover’s Distance

The Earth Mover’s Distance (EMD) quantifies the minimum effort required to transport a distribution (conceptualized as “soil”) from one location to another. This effort is calculated as the total amount of “soil” moved multiplied by the distance it travels. As shown in Equation (7), unlike CD, EMD requires that the two point clouds, $Q$ and $Q^{\prime }$, have the same size. A lower EMD value indicates greater similarity in the resolution and density of the point clouds. This metric captures the spatial alignment between points in the complementary network’s output and the corresponding plant sample point cloud, thereby validating the accuracy and reliability of the point cloud reconstruction.

$$EMD(Q,\widehat{Q})=\underset{\varphi :Q\to \widehat{Q}}{\min }{\displaystyle \frac{1}{\left|Q\right|}}{\displaystyle \sum _{q\in Q}^K‖{q-\varphi (q)‖}_2}$$

(7)

where $\pi$ is a bijection. Because EMD is defined only for point sets with the same cardinality $\left|Q\right|=\left|\widehat{Q}\right|=K$, we fix $K=2048$ as described above and compute the assignment using an approximate matching solver. The term $‖q-\varphi (q)‖$ denotes the Euclidean distance between a ground-truth point $q$ and its matched prediction $\varphi (q)$. This formulation enables EMD to capture fine spatial discrepancies between the generated and ground-truth point clouds, often providing a more sensitive similarity assessment than CD.

2.3.3. F-Score

The F-score evaluates model performance by combining precision and recall through the harmonic mean, providing a balanced measure of both metrics. Unlike CD, the F-score mitigates potential inaccuracies arising from point-to-point distance aggregation and serves as a more reliable and comprehensive evaluation metric. By integrating performance across multiple dimensions, the F-score offers a robust assessment of the model’s overall effectiveness.

Let $G={\left\{g_i\right\}}_{i=1}^K$ be the ground-truth point set and $Y={\left\{y_i\right\}}_{i=1}^K$ the prediction after cardinality normalization. We fix $K=2048$ for all evaluations; if a method outputs a different size, both $G$ and $Y$ are FPS-downsampled to 2048 beforehand. Distances are Euclidean and reported in the same units as CD/EMD. Given a global matching radius $d$, precision and recall are computed on the two point sets via nearest-neighbor distances:

$$\begin{array}{l}P(d)={\displaystyle \frac{1}{K}}{\displaystyle \sum _{y\in Y}1}\left[\underset{g\in G}{\min }{‖y-g‖}_2<d\right],\\ R(d)={\displaystyle \frac{1}{K}}{\displaystyle \sum _{g\in G}1}\left[\underset{y\in Y}{\min }{‖g-y‖}_2<d\right].\end{array}$$

(8)

The F-score is the harmonic mean:

$$F-Score(d)={\displaystyle \frac{2P(d)R(d)}{P(d)+R(d)}}$$

(9)

where $K=2048$ (cardinality normalization), and $d$ is a single global matching radius chosen on the validation set and kept constant across all methods.

3. Results

3.1. Training Visualization

At a 50% missing data rate (Figure 4a,b), the training loss curves of state-of-the-art methods, including GRNet, PF-Net, PMP-Net, PMP-Net++, SnowflakeNet, and PointTR v2 [40], exhibit markedly different convergence behaviors. GRNet is highly unstable, with large oscillations and no clear convergence by 160 epochs. PF-Net and PMP-Net drop rapidly at the start but plateau early, suggesting overfitting or convergence to suboptimal minima. PMP-Net++ and SnowflakeNet are smoother yet stagnate after ~30 epochs. PointTR v2 descends quickly in the first 10–15 epochs and is steadier than GRNet, but it shows an early transient spike and then saturates at a small, non-zero floor above our method. In contrast, our model (Figure 4b) exhibits a fast, near-monotonic decrease within the first ~20 epochs and stabilizes at a substantially lower plateau with minimal oscillation, indicating more reliable optimization and stronger learning at the 50% sparsity level; the visual comparisons at this setting (Figure 5 and Figure 6) further show superior preservation of leaf contours and vein continuity with a more uniform point distribution.

When the missing data rate increases to 75% (Figure 4c,d), these trends become more pronounced. GRNet remains severely unstable and fails to converge. PF-Net, PMP-Net, PMP-Net++, and SnowflakeNet still plateau early and show little additional loss reduction beyond the initial epochs, highlighting the difficulty these methods face under extreme sparsity. PointTR v2 has clear strengths: it descends rapidly at the start and is noticeably steadier than GRNet, reaching a lower loss than PF-Net and PMP-Net at this setting. Nevertheless, it soon saturates at a non-zero floor that remains above our method. By contrast, our model maintains robust convergence: the loss drops to a low level within about 20 epochs and then stays stable with minimal fluctuation (Figure 4d). Consistent with the curves, the visual comparisons at 75% (Figure 5 and Figure 6) show that although PointTR v2 produces reasonable global shapes, it tends to display uneven point density, boundary discontinuities, and vein fragmentation, whereas our reconstructions preserve contours and vein continuity with fewer artifacts and more uniform sampling. These results underline the effectiveness of our architecture, particularly the combination of adaptive density-aware processing and multiscale feature fusion, mitigating the adverse effects of severe data scarcity and sustaining low, stable loss values.

To further investigate the cause of this training stability and efficiency, Figure 4b reveals that the proposed model incorporates a novel parameter-freezing strategy. This technique significantly accelerates early convergence, as evidenced by the rapid decline in loss within the first 15 epochs. During subsequent epochs (15–80), the loss exhibits minor oscillations before stabilizing, a behavior directly attributable to the selective freezing of network parameters that guides optimization toward global minima. As the missing data rate increases from 50% to 75%, the model consistently demonstrates adaptive convergence characteristics, confirming its robustness and flexibility across varying levels of point cloud incompleteness. Taken together, both the comparative and causal analyses establish that the proposed model consistently outperforms state-of-the-art methods in terms of convergence efficiency and reconstruction accuracy, effectively addressing the challenges posed by moderate to severe point cloud data loss.

The completion network was implemented in PyTorch (version 1.10.2, CUDA 11.2) and optimized with Adam. We trained all components for 301 epochs with a global batch size of 16 and an initial learning rate of 0.001. Training was performed on 4 × NVIDIA GTX 2080 Ti with CUDA 11.6, Ubuntu 22.04.1 LTS, and Python 3.8. Evaluation protocol: before scoring, GT and predictions are FPS-normalized to K = 2048; F-score uses a fixed global matching radius d = 0.01.

3.2. Cotton Leaf Completion for Missing Shapes

3.2.1. Completion Results for 75% Missing Data

To rigorously assess point cloud completion performance under high missing rate scenarios, a comparative analysis was conducted using the cotton leaf dataset with a 75% missing-data rate. As summarized in

Table 1. Quantitative evaluation of point-cloud completion methods on the cotton leaf dataset with a 75% missing-data rate.

Means	${\mathit{L}}_1$	${\mathit{L}}_2$	EMD	F-Score
GRNet [20]	99.9812	60.6202	0.0629	0.1158
PF-Net [21]	49.1054	7.7551	0.0885	0.0765
PMP-Net [22]	47.2924	6.5268	0.0731	0.0285
PMP-Net++ [23]	57.2051	9.3005	0.0935	0.01
SnowflakeNet [24]	56.1553	9.8552	0.0536	0.0131
PointTR v2 [40]	45.1987	6.6914	0.1022	0.0519
Ours	24.5246	1.5396	0.0404	0.1192

, the proposed model demonstrates statistically significant improvements over existing advanced methods across all major metrics.

Specifically, in terms of the two mean distance metrics ($L_1$ and $L_2$), the proposed model reduces the error by 75.5% and 97.5%, respectively, compared with GRNet, which represents the baseline with the highest error values. Relative to PointTR v2, our model further reduces these two mean distance metrics by 45.7% and 77.0%, respectively. For EMD, our model achieves a 24.6% lower error than the best-performing reference model, SnowflakeNet, and a 60.5% lower error than PointTR v2. Furthermore, the proposed model attains the highest F-score among all compared methods, improving by 2.9% relative to GRNet, by 809.9% relative to SnowflakeNet, and by 129.7% relative to PointTR v2, thereby demonstrating a more effective balance between completion accuracy and structural integrity. Taken together, these statistical results highlight the comprehensive performance advantage of the proposed model for point cloud completion under severe data loss conditions, establishing its effectiveness for reconstruction in complex and highly sparse scenarios.

Following the comparative analysis of various models using quantitative metrics such as EMD, F-score, and CD to evaluate the performance of point cloud reconstruction on the cotton dataset with a 75% missing data rate, a visual assessment was conducted to further substantiate and elucidate these numerical findings. This step is essential, as quantitative metrics alone may not fully capture critical aspects such as the uniformity of the point cloud distribution and the preservation of detailed morphological features. Therefore, to comprehensively evaluate the reconstruction fidelity, Figure 5 presents complementary multiangle renderings of six Gossypium hirsutum leaf samples under the condition of 75% missing data, effectively illustrating the reconstruction capabilities of the various models and providing deeper insight into their practical advantages and shortcomings.

Figure 5 presents a visual comparison of reconstructed cotton-leaf point clouds under 75% missing data. GRNet has difficulty capturing complex structural details, particularly for Leaves (4) and (5), where the results are fragmented, noisy, and far from the ground-truth morphology. PF-Net shows significant distortions and uneven density distributions, most evident in Leaves (3) and (4), with elliptical point patterns that radiate from central regions and cause shape deformation and loss of detail. PMP-Net and PMP-Net++ consistently produce sparse and over-smoothed reconstructions, for example, Leaves (2) and (6), yielding rounded outlines that lack critical geometric nuances. SnowflakeNet provides relatively complete shapes, yet clear layering artifacts and nonuniform densities remain; in Leaves (3) and (4), the detailed contours are not fully recovered. PoinTR v2 improves global shape fidelity compared with GRNet, PF-Net, and the PMP family, but it still underperforms SnowflakeNet: fewer serrations and marginal veins are restored and edges are blurrier with stronger local density fluctuations, especially in Leaves (3) and (4). By contrast, our method reconstructs sharper leaf margins and more continuous venation with fewer artifacts, achieving the closest resemblance to the ground truth.

To further elucidate these microstructural differences, Figure 6 provides a fine-grained visual comparison of leaf edge reconstruction between SnowflakeNet and the proposed method under conditions of 75% missing data. SnowflakeNet generally succeeds in generating globally complete leaf shapes, yet its outputs frequently display uneven point distributions and noticeable layering artifacts. These issues are especially evident in samples (1), (2) and (4), where high levels of input noise result in poorly defined contours and fragmented edge structures. In these cases, the reconstructed point clouds exhibit discontinuities along the boundaries, with local geometric details frequently blurred or absent. Layering effects are visually apparent, manifesting as stratified, nonuniform patches that detract from the morphological realism of the reconstructed leaves.

In contrast, our proposed model clearly outperforms the other methods in terms of reconstruction quality across all six leaf samples, effectively maintaining uniformity in the point cloud distribution and preserving intricate morphological features. Leaf (4) and Leaf (5) highlight the strength of our approach, where reconstructed point clouds closely replicate the original leaf shape, displaying enhanced topological accuracy and detailed structural fidelity. Overall, these visual comparisons highlight the robustness of the proposed model in reconstructing cotton leaf point clouds under conditions of severe data loss. The model consistently maintains structural fidelity and preserves fine morphological features, thereby demonstrating its suitability and effectiveness for high-precision point cloud completion in challenging phenotyping applications.

When the missing data rate is reduced to 50%, our model demonstrates even greater reconstruction fidelity, underscoring its strong robustness and adaptability to varying data conditions. In scenarios with lower missing rates, the reshape module refines the global leaf shape, enhancing geometric accuracy, whereas the upsampling and freezing modules work together to restore local features and improve point cloud uniformity. This progressive enhancement in performance, observed consistently as the amount of missing data decreases, confirms the model’s generalizability and structural integrity, validating its practical applicability across diverse data-loss scenarios.

3.2.2. Completion Results for 50% Missing Data

On the basis of the results presented in

Table 2. Quantitative evaluation of point-cloud completion methods on the cotton leaf dataset with a 50% missing-data rate.

Means	${\mathit{L}}_1$	${\mathit{L}}_2$	EMD	F-Score
GRNet [20]	57.9596	26.4345	0.0466	0.1794
PF-Net [21]	53.9436	9.4655	0.0802	0.0857
PMP-Net [22]	41.7531	5.2057	0.0808	0.0764
PMP-Net++ [23]	47.9835	6.6025	0.0820	0.0230
SnowflakeNet [24]	32.2416	3.7345	0.0553	0.1858
PointTR v2 [40]	42.2060	5.8453	0. 1407	0. 0119
Ours	37.0253	4.6137	0.0533	0.1329

, a statistical analysis of the cotton leaf dataset at a 50% missing data rate reveals notable performance differences among the evaluated architectures. Compared with SnowflakeNet, the proposed model significantly improves the point distribution uniformity, even though SnowflakeNet results in slightly higher local geometric accuracy. This enhanced uniformity contributes to improved morphological integrity and practical applicability in phenotypic analysis, highlighting the overall superiority of the proposed approach for robust point cloud reconstruction under incomplete data conditions. Specifically, the proposed model reduces the EMD by 3.6% relative to SnowflakeNet, highlighting its superior ability to maintain a consistent point cloud distribution. Visual inspection of the reconstructed point clouds further corroborates these findings, as our model produces more uniformly distributed points and preserves greater morphological detail, resulting in reconstructions that closely align with the ground truth.

Further analysis indicates that architectural refinements, such as the strategic reduction in convolutional operations in local feature encoding and the enhancement of global attention weighting, resulted in a predictable trade-off. Although there was a slight increase in localized geometric error, with $L_1$ increasing by 14.8% and $L_2$ increasing by 23.5%, the shape completeness as measured by the F-score remained statistically equivalent to SnowflakeNet. This outcome demonstrates that our approach effectively reallocates computational resources from low-level feature extraction to high-level contextual modelling, thereby prioritizing topological fidelity over local metric optimization. Although GRNet achieved the lowest EMD value among all methods, its visual reconstructions presented distribution and completeness deficiencies. PointTR v2, a recent transformer baseline, shows mid-range $L_1/L_2$ but the largest EMD and a very low F-score, indicating difficulty in maintaining structural completeness and uniform point density under 50% sparsity. By comparison, the proposed model delivered a strong balance, with competitive EMD and a consistently superior F-score, outperforming PF-Net, PMP-Net, and PMP-Net++ across all the key metrics. These results establish the model’s effectiveness in achieving both uniform point distribution and detailed morphological preservation in challenging, incomplete data scenarios.

To rigorously assess the performance of various models in the task of cotton leaf point cloud reconstruction, this study designed a controlled experiment involving 50% data omission. As illustrated in Figure 7, a comparative analysis reveals substantial disparities among the models, particularly in terms of reconstruction quality and fidelity to fine details.

The experimental results indicate that the conventional GRNet, despite demonstrating fundamental reconstruction capabilities, has significant limitations in preserving complex local details and maintaining uniform point cloud distributions. These shortcomings are particularly evident in Leaf (4), where GRNet cannot reconstruct critical high-curvature features, including apical indentations, and produces substantial noise artifacts. PF-Net, while exhibiting an improvement over GRNet in terms of global structural accuracy, still suffers from uneven point density distributions. Notably, Leaf (4) exhibited an elliptical scattering pattern originating from the center, resulting in pronounced distortions in overall leaf shape and a considerable loss of local detail. Similarly, PMP-Net and PMP-Net++ yield sparse point clouds, with reconstructed contours tending toward simplistic arc-shaped forms that inadequately represent intricate leaf structures. PointTR v2, though capable of recovering a plausible global outline, frequently shows density clustering and over-smoothed details; in Leaf (2) and Leaf (3) points concentrate along the midrib leaving lamina voids, and in Leaf (5) and Leaf (6) the margins remain jagged with broken veins, indicating incomplete structural recovery under sparsity. As illustrated in Figure 8, although SnowflakeNet produces relatively complete reconstructions, it frequently results in uneven point distributions and pronounced layering artifacts. These shortcomings become especially apparent in challenging cases such as Leaf (5) and Leaf (6), where substantial input noise leads to significant contour ambiguity and a degraded edge definition in the reconstructed results.

In contrast, the proposed model demonstrates consistently superior reconstruction performance across all six leaf samples, maintaining both global morphological integrity and high-fidelity local detail. The reconstructed point clouds exhibit substantially improved uniformity and edge continuity, closely match the ground truth and effectively preserve fine structural features. Qualitative analysis further reveals that the model not only restores the overall leaf shape with high accuracy but also excels in capturing intricate microstructural details. Notably, the following advantages are observed: (1) the uniformity and continuity of the point cloud distribution surpass those achieved by the control models; (2) the accuracy of leaf edge contour reconstruction is significantly enhanced, showing excellent agreement with the morphological features of the ground truth; and (3) the preservation of surface detail fidelity is superior to that of all other methods, highlighting the model’s strong capability in feature extraction and reconstruction. These experimental results unequivocally demonstrate the model’s enhanced ability to address both large-scale completeness and fine-grained accuracy in complex cotton leaf point cloud reconstruction, thereby underscoring its potential as a robust solution for high-precision phenotypic analysis and advancing the state-of-the-art in point cloud reconstruction research.

3.3. Cotton Leaf Completion for Sparse Point Cloud

To evaluate the performance of cotton leaf point clouds with 75% sparsity systematically, quantitative and qualitative comparisons were conducted against state-of-the-art completion methods and the proposed algorithm. Performance was assessed via four widely adopted metrics: $L_1$ distance, $L_2$ distance, EMD, and F-score.

On the basis of the results in

Table 3. Quantitative evaluation of point-cloud completion methods on sparse cotton leaf data.

Means	${\mathit{L}}_1$	${\mathit{L}}_2$	EMD	F-Score
GRNet [20]	108.1170	83.9074	0.0975	0.0291
PF-Net [21]	34.6940	3.4182	0.0513	0.1004
PMP-Net [22]	39.5188	4.4992	0.0764	0.0139
PMP-Net++ [23]	44.6338	5.5138	0.0795	0.0048
SnowflakeNet [24]	26.6929	1.7999	0.0206	0.0900
PointTR v2 [40]	25.2929	2.3216	0.3493	0.0787
Ours	28.7565	2.1111	0.1780	0.0869

, a systematic evaluation was conducted to compare state-of-the-art completion methods on sparse cotton-leaf point clouds. Compared with the next-best method, PMP-Net++, the proposed model achieves markedly better reconstruction quality: $L_2$ (CD) decreases from 5.51 to 2.11 (a 61.7% reduction), $L_1$ decreases from 44.63 to 28.76 (a 35.6% reduction), and F-score increases from 0.48% to 17.8%; EMD remains comparable. Notably, compared with SnowflakeNet, which reports the lowest $L_1$/$L_2$/EMD, our method attains a substantially higher F-score (an 8.8 percentage-point improvement), indicating more complete and uniform reconstructions under severe sparsity. These results highlight the model’s robust balance between reconstruction fidelity and structural completeness in the presence of large missing regions and challenging occlusions.

Figure 9 provides a qualitative comparison of point cloud completion results for sparse cotton leaves using different methods. As illustrated, the input point clouds are highly sparse, presenting significant challenges for accurate reconstruction. GRNet produced outputs with substantial noise and incomplete structural recovery, while PF-Net captured only the coarse outline of leaves, resulting in interior voids and irregular point distributions. PMP-Net and PMP-Net++ maintained coherent global morphology, but persistent internal gaps and surface discontinuities were still evident due to sparsity, compromising local topological integrity. PointTR v2 recovers a reasonable outline and cleaner interiors than GRNet and PF-Net, but under high sparsity, it shows density clustering and over-smoothed lamina; margins remain jagged and veins partly broken, indicating incomplete fine-detail recovery.

In contrast, both SnowflakeNet and the proposed method generated more complete and structurally coherent point clouds. However, the proposed algorithm more effectively preserved fine morphological features, such as leaf serrations and vein patterns, especially in Leaves 3 and 4, resulting in reconstructions closely aligned with the ground truth. This qualitative evidence is consistent with the quantitative results presented in Table 3 and further confirms the enhanced capability of the proposed approach for detailed phenotypic analysis in agricultural applications, particularly for accurate cotton leaf structural recovery and subsequent phenotypic parameter extraction.

3.4. Ablation Study

To comprehensively evaluate the effectiveness of each architectural component, an ablation study was conducted on the point cloud completion model using a self-constructed cotton dataset with a 75% missing data rate. Three key modules (upsample, reshape, and frozen) were systematically enabled or disabled, resulting in five model configurations, as summarized in

Table 4. Ablation study of module contributions to point-cloud completion performance under a 75% missing-data rate on the cotton leaf dataset. Model variants include upsampling, reshaping, and freezing modules.

ID 1	Upsample	Reshape	Frozen	${\mathit{L}}_1$	${\mathit{L}}_2$	F1	EMD
1	√	×	×	45.6528	6.7480	0.0838	0.0669
2	×	√	×	31.583	2.6188	0.0763	0.0458
3	×	√	√	30.7775	2.5197	0.0848	0.0488
4	√	√	×	27.0097	1.8167	0.0893	0.0410
5	√	√	√	26.3892	1.8116	0.1711	0.0404

¹ Each combination is assigned a unique ID ranging from 1 to 5. Bold values indicate the best performance across the compared settings.

. Each configuration was assigned a unique identifier (ID 1–5) for systematic comparison. Through these targeted comparisons, the influence of each individual network component and the combined contributions to the final reconstruction quality were clearly identified.

ID 1 included only the upsampling module while disabling both reshaping and freezing.

ID 2 enables the reshaping module without upsampling and freezing.

ID 3 activated both upsampling and reshaping but excluded freezing.

ID 4 integrates the upsampling and reshaping modules, disabling the freezing module.

ID 5 represents the complete model, which simultaneously incorporates upsampling, reshaping, and freezing modules.

On the basis of the results in Table 4, the full model (ID 5), which integrates upsampling, reshaping, and freezing, achieves the best overall performance. Relative to the baseline configuration (ID 1), ID 5 reduces EMD by 39.6% and increases the F1 score by 104.2%, indicating substantial gains in geometric fidelity and structural completeness. Models incorporating the reshaping module (IDs 2–5) consistently outperform those without it, yielding on average 67.5% lower $L_2$ error and 25.7% higher F1 than the non-reshaping baseline. Adding the upsampling module further enhances performance: within the reshaping family, configurations that include both upsampling and reshaping (IDs 4 and 5) achieve up to 16.4% lower $L_1$ error and 17.2% lower EMD compared with their reshape-only counterparts (IDs 2 and 3).

Interestingly, the freezing parameters had varying impacts on performance. While it enhanced the results when it was combined with both upsampling and reshaping (ID 5), it slightly diminished performance when it was used with reshaping alone (as seen when comparing IDs 2 and 3). These findings validate the effectiveness of our architectural design, demonstrating that the full integration of all three components yields the most robust and accurate point cloud completion results.

3.5. Robustness Test

We conduct all robustness tests on the cotton-leaf dataset. In the first test, we vary the extent of incompletion to 50% and 75% while keeping the output cardinality fixed (K = 2048) and evaluating under the same protocol (normalization, FPS, CD/L1, CD/L2, EMD, F-score). The missing pattern is Axial-Cut, where a half-space cut removes a large sheet-like portion of the lamina. As shown in Figure 10, our network reconstructs coherent contours, restores serrations and vein continuity, and maintains uniform sampling even when three quarters of the points are removed. In the second test, we switch to Sphere-Removal missing, where spherical windows carve local voids across the leaf surface at the same 50%/75% ratios. Figure 11 illustrates that the model reliably infills these punctate holes, suppresses artifacts, and preserves global shape and boundary sharpness with balanced point density. Together, the two tests demonstrate that the proposed method is robust to both sheet-like and localized occlusions under severe sparsity. Across Axial-Cut and Sphere-Removal tests at 50%/75% missing ratios, the model sustains low, stable losses and produces visually faithful, uniformly sampled completions, verifying that the framework generalizes well to distinct occlusion geometries and severe sparsity and is applicable to diverse data-loss conditions.

4. Discussion

4.1. Quantitative and Quality Analysis

The quantitative results across multiple benchmarks consistently showed that the proposed method either outperformed or closely matched state-of-the-art models, particularly under challenging conditions with 75% missing data. While PMP-Net attains the lowest $L_1$ and $L_2$ errors and SnowflakeNet the lowest EMD among the reference methods, the proposed model achieves the highest F-score, exceeding SnowflakeNet by 809.9%, reflecting superior structural completeness and point-distribution uniformity. In addition, the proposed model reduces the $L_2$ distance by 76.4% relative to PMP-Net and the EMD by 24.6% relative to SnowflakeNet.

Importantly, quantitative metrics alone may not provide a comprehensive assessment of reconstruction quality. Previous studies have demonstrated that models that achieve high scores on global metrics such as EMD or CD can still produce suboptimal outcomes with respect to spatial uniformity and perceptual realism [34,41]. Although SnowflakeNet achieves superior results in certain numerical indicators, qualitative analysis reveals significant limitations in its practical application. In particular, qualitative evaluations reveal that SnowflakeNet often generates nonuniform point distributions and pronounced layering artifacts, corroborating the challenges previously identified by Sun et al. [24] and Weng et al. [42]. These deficiencies are especially evident in the reconstruction of fine-scale structures such as leaf edges and venation, thereby compromising both the morphological integrity and the perceptual fidelity of the reconstructed point clouds. This limitation has also been highlighted in prior evaluations of generative point cloud models [43]. In contrast, the proposed model consistently achieves visually superior reconstructions, preserving intricate structural details and maintaining a uniform distribution, both of which are critical for practical applications and downstream phenotypic analysis. This observation underscores a key limitation of relying exclusively on global distance-based metrics, because these metrics may fail to capture local structural fidelity and visual realism. The practical utility of point cloud completion methods is therefore more accurately reflected through qualitative visualization and a comprehensive, multimetric evaluation strategy rather than through single-metric rankings alone.

Building upon this insight, further analysis was conducted through ablation studies to elucidate the specific contributions of individual architectural modules. Both the reshape and upsampling modules were shown to be indispensable for achieving optimal geometric fidelity and improving the F-score, while the parameter-freezing module played a crucial role in stabilizing the training process and accelerating convergence. The integration of these modules enables the network to robustly reconstruct complex geometric patterns and maintain a uniform point distribution, even in the presence of substantial data loss. Together, these findings underscore the necessity of combining strong quantitative performance with qualitative robustness to ensure reliable application in real-world phenotypic analysis tasks.

4.2. Loss Function Freezing Strategy

The loss function freezing strategy demonstrates strong applicability in multitask deep learning networks. By selectively freezing the loss function parameters of specific tasks at designated training stages, this approach reduces task interference and enables each task to learn independently and efficiently. In multitask learning scenarios, the optimization of multiple objectives often introduces gradient conflicts, which can impede training stability and hinder convergence. Numerous studies have shown that loss freezing and adaptive loss scheduling are effective in mitigating these conflicts, thereby facilitating stable and efficient optimization [44,45,46]. This is especially important in complex networks where shared parameters are influenced by competing gradients from different tasks. Loss function freezing has been successfully applied in computer vision and natural language processing to improve training stability, accelerate convergence, and enhance the robustness of multitask architectures [47]. Techniques such as GradNorm [48], adaptive loss balancing [49], and curriculum-based or phased training [50] have improved in both convergence performance and generalization performance by dynamically adjusting the contributions of different losses during training. In the context of point cloud completion, the loss function freezing strategy implemented in this study effectively stabilized the training process, mitigated gradient conflicts, and improved convergence rates. Consequently, the model exhibited greater robustness when processing complex datasets, particularly cotton leaf point clouds with extensive missing regions. These results are consistent with broader findings in multitask learning literature, where loss freezing strategies are recognized for enhancing optimization efficiency and overall predictive quality.

Despite these strengths, several limitations should be acknowledged. The proposed approach was validated exclusively on a self-constructed cotton leaf dataset, and its generalizability to other plant species and diverse point cloud modalities remains to be established. The computational complexity introduced by multistage feature extraction and upsampling could also limit its feasibility for real-time deployment in large-scale or high-throughput scenarios. The current framework is designed primarily for geometric completion and does not yet incorporate color or spectral information, which could provide additional improvements in the precision and utility of phenotypic analysis. Future research will focus on extending the model’s applicability to a wider variety of plant phenotypes and real-world field datasets. The integrating of multimodal data, such as hyperspectral or RGB point clouds, has the potential to further increase the effectiveness of plant health monitoring. Continued algorithmic advancements aimed at lowering computational cost, improving scalability, and incorporating uncertainty estimation for more reliable completion in highly ambiguous cases will also be pursued.

5. Conclusions

To address the challenges posed by large-scale missing data (missing ratio ≥ 50%) and severe sparsity (points ≤ 128) in cotton leaf point clouds, an end-to-end super-resolution completion network is proposed. This network is designed to reconstruct uniform, high-resolution, and fully complete 3D point cloud models. It integrates an MRE to extract hierarchical features from locally incomplete point clouds, thereby enhancing the representation of both semantic and geometric attributes. A PPD is employed to iteratively generate missing point clouds layer by layer, enabling precise reconstruction of both missing regions and intricate details. To ensure consistency and uniformity, PointNet++ is leveraged to fuse the incomplete input point cloud with the generated completed point cloud, producing a fully reconstructed and evenly distributed point cloud model. Additionally, a loss function freezing strategy is implemented, allowing the network to efficiently execute task-specific objectives across different training stages, ultimately enhancing the overall reconstruction accuracy and quality.

Extensive experiments demonstrated that the proposed model delivers robust and accurate completion even under extreme scenarios characterized by large-scale data loss and severe sparsity. The reconstructed point sets are dense, uniform, and boundary-preserving, reducing fragmentation and sampling bias and supporting consistent geometric analysis across growth stages and occlusion patterns. This strong performance under the most challenging conditions implies that the model is also well-suited for small-scale or mild sparsity cases, where completion quality can be expected to be even higher. These results indicate that the proposed framework provides a reliable and generalizable solution for 3D point cloud completion in diverse agricultural phenotyping applications.