GreenhouseSplat: Gaussian Splatting for Greenhouses

• GreenhouseSplat is a diverse set of greenhouse frameworks that use Gaussian splatting to enable detailed 3D reconstruction, semantic mapping, and simulation.
• The active semantic mapping system fuses a low-resolution Octomap with dense 3D Gaussian splats to achieve fruit-level reconstruction improvements of up to 28.6% under segmentation noise.
• The simulation and multispectral package converts RGB and multispectral imagery into 2D/3D splats using wavelet regularization for efficient few-shot reconstruction and benchmark evaluation.

GreenhouseSplat denotes a small but expanding cluster of Gaussian-splatting systems for greenhouse and horticultural scenes rather than a single canonical method. In current arXiv usage, the name refers most directly to an active semantic mapping architecture that couples a low-resolution semantic Octomap with dense semantic 3D Gaussian Splatting for mobile-manipulator phenotyping (Cuaran et al., 17 Jan 2026), a photorealistic greenhouse simulation framework and dataset based on 2D Gaussian Splatting and MAST3R-SfM (Tabaa et al., 2 Oct 2025), and, in an informal package-level sense, a sparse-view multispectral reconstruction suite centered on LGDWT-GS, a greenhouse dataset, and a standardized benchmark (Salehi et al., 23 Jan 2026). Across these usages, the common denominator is the use of Gaussian splats as explicit scene primitives for greenhouse reconstruction, rendering, planning, and evaluation.

1. Terminological scope and principal usages

Across the cited literature, GreenhouseSplat is not a standardized term. It names at least three distinct greenhouse-oriented systems, and one paper explicitly notes that the label is not used verbatim in the main text even though it is a natural shorthand for the released package (Salehi et al., 23 Jan 2026). A further source of ambiguity is that the label has also been glossed in a non-agricultural direction as “green, communication-efficient mixed reality using Gaussian Splatting,” underscoring that the term remains unsettled as a research designation (Liu et al., 18 Apr 2025).

Usage	Core representation	Primary function
Active semantic mapping	Low-res semantic Octomap + semantic 3DGS	Target-aware reconstruction and planning
Photorealistic simulator and dataset	MAST3R-SfM + 2DGS	ROS simulation, RGB/LiDAR rendering, localization
Sparse-view greenhouse package	LGDWT-GS + multispectral dataset + benchmark	Few-shot reconstruction and evaluation

The first usage is operational and robotic. It treats GreenhouseSplat as an active semantic mapping system for horticultural environments, designed for fruit counting, fruit or plant volume estimation, and related phenotyping tasks (Cuaran et al., 17 Jan 2026). The second usage is synthetic and infrastructural. It turns ordinary RGB photos of real cucumber plants into 2DGS assets that can be assembled into ROS-based greenhouse simulations with camera and LiDAR rendering (Tabaa et al., 2 Oct 2025). The third usage is methodological and benchmark-oriented. It bundles a wavelet-regularized 3DGS method, a controlled multispectral greenhouse dataset, and an open benchmarking package for sparse-view reconstruction (Salehi et al., 23 Jan 2026).

This multiplicity matters technically. GreenhouseSplat may refer to 2D Gaussian Splatting or 3D Gaussian Splatting; to mobile manipulation, simulation, or few-shot novel-view synthesis; and to semantic mapping, photorealistic rendering, or multispectral reconstruction. Any precise use of the term therefore depends on the cited paper rather than on a universally accepted architecture.

2. Hybrid active semantic mapping for horticultural environments

In "Active Semantic Mapping of Horticultural Environments Using Gaussian Splatting" (Cuaran et al., 17 Jan 2026), GreenhouseSplat is an active 3D reconstruction framework for horticultural environments using a mobile manipulator. Its central design is representational complementarity. A low-resolution semantic Octomap with voxel size $\delta S = 0.05$ m is maintained solely for occupancy reasoning, semantic information gain, and collision checking, while a dense semantic 3DGS model is optimized for accurate geometry, photometry, and semantics. The low-resolution occupancy map preserves explicit known, free, and unknown space; the Gaussian model preserves continuous geometry and small-scale plant detail.

The 3DGS component represents the scene as isotropic 3D Gaussians $\{g_i\}_{i=1}^N$, each with center $\boldsymbol{\mu}_i \in \mathbb{R}^3$, radius $r_i \in \mathbb{R}$, color $\mathbf{c}_i \in \mathbb{R}^3$, opacity $o_i \in [0,1]$, and semantic label $$\mathbf{s}_i \in \mathcal{S}=\{\text{fruits},\text{leaves},\text{background}\}$$. The raw density is

$$f_i(\mathbf{x}) = o_i \exp\left(-\frac{\lVert \mathbf{x} - \boldsymbol{\mu}_i \rVert^2}{2r_i^2}\right),$$

and rendering uses front-to-back alpha compositing,

$$C(\mathbf{p}) = \sum_{i=1}^N \mathbf{c}_i\, f_{2D_i}(\mathbf{p}) \prod_{j=1}^{i-1}\left(1-f_{2D_j}(\mathbf{p})\right).$$

Depth, semantic, and silhouette images are rendered analogously. A key implementation choice is to treat camera poses as known from manipulator kinematics and use 3DGS purely for mapping rather than for SLAM.

Mapping is optimized jointly over depth, color, and semantics. The paper introduces a confidence-weighted semantic loss,

$$L_s(\mathbf{p}) = \bigl| conf^2(\mathbf{p})\,( S(\mathbf{p}) - S_{gt}(\mathbf{p}) ) \bigr|,$$

which down-weights uncertain segmentation labels. This is paired with semantic-aware densification: fruits are sampled densely, roughly one Gaussian per pixel, while non-target semantics are downsampled by 90%. The resulting representation is deliberately non-uniform. It preserves phenotyping targets at high fidelity while constraining memory and compute on leaves and background.

Viewpoint planning alternates between exploitation and exploration. Exploitation begins by clustering fruit-labeled Gaussians with DBSCAN using semantic cluster radius $\{g_i\}_{i=1}^N$0 m and min-samples $\{g_i\}_{i=1}^N$1. Candidate viewpoints are then sampled on a sphere of radius $\{g_i\}_{i=1}^N$2 m with azimuth discretization $\{g_i\}_{i=1}^N$3 and elevation discretization $\{g_i\}_{i=1}^N$4. Exploration viewpoints are sampled on planes parallel to crop rows, avoiding explicit frontier extraction in the Octomap. Informative-view scoring is split accordingly: Unknown Voxel Count is used for exploration, while OSAMCEP is used for exploitation as an occlusion- and semantic-aware expected information gain. The best $\{g_i\}_{i=1}^N$5 viewpoints become graph nodes, each node is connected to $\{g_i\}_{i=1}^N$6 nearest neighbors in joint space, and only the first $\{g_i\}_{i=1}^N$7 waypoints of the best path are executed before replanning.

The reported gains are both geometric and task-level. In simulation, GreenhouseSplat achieves fruit-level reconstruction metrics of CD $\{g_i\}_{i=1}^N$8 m, $\{g_i\}_{i=1}^N$9, $\boldsymbol{\mu}_i \in \mathbb{R}^3$0, and $\boldsymbol{\mu}_i \in \mathbb{R}^3$1 without segmentation noise, and CD $\boldsymbol{\mu}_i \in \mathbb{R}^3$2 m, $\boldsymbol{\mu}_i \in \mathbb{R}^3$3, $\boldsymbol{\mu}_i \in \mathbb{R}^3$4, and $\boldsymbol{\mu}_i \in \mathbb{R}^3$5 with segmentation noise. The corresponding 0.01 m Octomap-only baseline reports $\boldsymbol{\mu}_i \in \mathbb{R}^3$6 without noise and $\boldsymbol{\mu}_i \in \mathbb{R}^3$7 with noise. The abstract summarizes this as a 6.6% improvement in fruit-level $\boldsymbol{\mu}_i \in \mathbb{R}^3$8 score under noise-free conditions and up to 28.6% under segmentation noise, together with a 50% reduction in runtime. For phenotyping, GreenhouseSplat reports fruit volume accuracy of approximately $\boldsymbol{\mu}_i \in \mathbb{R}^3$9 and fruit count accuracy of approximately $r_i \in \mathbb{R}$0 without noise; under noise, the corresponding values are approximately $r_i \in \mathbb{R}$1 and $r_i \in \mathbb{R}$2. By comparison, the 0.01 m Octomap baseline shows strong overestimation in volume and severe variance in count under noise.

Robustness and memory usage are treated as first-class design constraints. An ablation shows that removing the confidence term in the semantic loss degrades performance under segmentation noise from $r_i \in \mathbb{R}$3 to $r_i \in \mathbb{R}$4. Memory experiments show that without downsampling, mapping a single row peaks at 12.4 GB of GPU memory and produces approximately 140 MB Gaussian parameter files, whereas 0.9 downsampling for non-targets reduces usage to 4.6 GB and file size to approximately 20 MB. This suggests that the system is not merely a higher-fidelity reconstructor than Octomap, but a domain-specialized semantic mapper optimized around fruit-level phenotyping.

3. Photorealistic greenhouse simulation and the 2DGS dataset

In "GreenhouseSplat: A Dataset of Photorealistic Greenhouse Simulations for Mobile Robotics" (Tabaa et al., 2 Oct 2025), the term denotes both a reconstruction pipeline and a released dataset. The pipeline converts regular RGB imagery of real cucumber plants into 2D Gaussian Splatting assets and integrates them into a ROS-based simulator with camera and LiDAR rendering. The dataset comprises 8 reconstructed row segments, specifically the start and end segments of 4 cucumber rows, covering 82 unique cucumber plants.

The acquisition and reconstruction pipeline is tailored to long, self-similar greenhouse rows. A human operator captures RGB images with a handheld DSLR camera at 12.5 fps while walking around the rows, with extra overlap at the row ends. Sparse reconstruction is then performed with MAST3R-SfM: MAST3R generates dense correspondences between all image pairs, and a standard SfM pipeline is run on top of those matches. Before Gaussian training, each reconstruction is manually aligned so that the $r_i \in \mathbb{R}$5-axis points upward, the $r_i \in \mathbb{R}$6-axis runs along the plant row, and row ends point in the positive $r_i \in \mathbb{R}$7-direction. This orientation step is important because later post-training rotation could corrupt view-dependent appearance.

The radiance-field representation is 2DGS rather than 3DGS. Each primitive has a 3D center $r_i \in \mathbb{R}$8, a local tangent basis $r_i \in \mathbb{R}$9, scales $\mathbf{c}_i \in \mathbb{R}^3$0 stored logarithmically as $\mathbf{c}_i \in \mathbb{R}^3$1, a color $\mathbf{c}_i \in \mathbb{R}^3$2, and an opacity $\mathbf{c}_i \in \mathbb{R}^3$3. For a given view, the projected ellipses are composited front-to-back:

$\mathbf{c}_i \in \mathbb{R}^3$4

Training uses the published 2DGS method of Huang et al., default hyperparameters, and 30,000 iterations per row end. After training, each model is manually cleaned by removing irrelevant background and unreliable open-end regions. Since SfM recovers geometry only up to scale, a global scale factor $\mathbf{c}_i \in \mathbb{R}^3$5 is then applied to positions and logarithmic Gaussian scales:

$\mathbf{c}_i \in \mathbb{R}^3$6

The released dataset preserves both reconstruction quality and simulator utility. Per-segment metrics include L1 error, PSNR, and Gaussian count. Reported PSNR values range from 27.9 dB to 32.1 dB, with Gaussian counts between 273,297 and 478,812; the paper notes that most segments achieve PSNR of approximately 30–32 dB. Each segment stores camera poses, intrinsics, scale normalization, and globally aligned coordinates, and can be used either directly in a splatting renderer or as a ROS simulation asset.

The simulator architecture is explicitly robotics-oriented. Row segments are assembled in Blender into a coherent greenhouse, a Clearpath Jackal UGV is used as the mobile platform, and ROS 2, tf2, and RViz provide the middleware layer. Camera rendering is organized as a thin ROS camera client plus a separate 2DGS rendering backend. LiDAR is not ray-traced directly; instead, four 90° depth images are rendered to cover 360°, and the point cloud is generated by back-projecting depth through the camera intrinsics,

$\mathbf{c}_i \in \mathbb{R}^3$7

The resulting point clouds are merged and published on a /lidar topic. The paper also demonstrates AprilTag localization by constructing fiducials as Gaussian splats and rendering them directly in the scene. The reported evaluation is qualitative rather than numerical: estimated TF frames from the AprilTag detector coincide visually with the true tag positions in RViz, with no visible offset.

This version of GreenhouseSplat is therefore a greenhouse-scale simulation substrate rather than a semantic phenotyping system. It is designed to close a realism gap in agricultural robotics by replacing procedural plants, coarse point clouds, or hand-modeled meshes with radiance-field assets reconstructed from real greenhouse imagery.

4. The sparse-view multispectral package centered on LGDWT-GS

In "LGDWT-GS: Local and Global Discrete Wavelet-Regularized 3D Gaussian Splatting for Sparse-View Scene Reconstruction" (Salehi et al., 23 Jan 2026), GreenhouseSplat is best understood as an informal label for a combined package: the LGDWT-GS method, a controlled multispectral greenhouse dataset, and an open-source few-shot 3DGS benchmarking toolkit. The paper itself consistently names the method LGDWT-GS and the data a multispectral greenhouse dataset, but presents the three components as a tightly coupled release.

Algorithmically, LGDWT-GS keeps the standard 3DGS renderer and modifies the loss. The composite objective is

$\mathbf{c}_i \in \mathbb{R}^3$8

Global DWT supervision decomposes the rendered and ground-truth images into Haar subbands $\mathbf{c}_i \in \mathbb{R}^3$9, $o_i \in [0,1]$0, $o_i \in [0,1]$1, and $o_i \in [0,1]$2, and applies a weighted loss that sets $o_i \in [0,1]$3 near zero in order to de-emphasize unstable diagonal high-frequency content under sparse-view conditions. Local DWT supervision computes a low-frequency energy ratio

$o_i \in [0,1]$4

selects patches in the lowest 20% percentile of the $o_i \in [0,1]$5 distribution, and applies patchwise DWT loss over the $o_i \in [0,1]$6 and $o_i \in [0,1]$7 bands. The intended effect is to stabilize global geometry while selectively sharpening local detail.

The greenhouse dataset is a controlled multispectral acquisition rather than an in-row mobile-robot dataset. It uses an MSIS-AGRI-1-A snapshot multispectral camera with four spectral bands—Green at 580 nm, Red at 660 nm, Red Edge at 735 nm, and NIR at 820 nm—together with synchronous 4-channel LED illumination. The plant species are Sorghum, Tomato, Alocasia, Cotton, and Grape. Imaging occurs at a controlled greenhouse station with a motorized turntable, a uniform black background, and two identical multispectral cameras placed on opposite sides of the turntable and translated vertically across four heights. Each plant is imaged at ten rotational steps of 36° and yields approximately 80–100 multispectral frames; across all species, the release contains nearly 500 spatially aligned multispectral images.

The package also standardizes few-shot evaluation. It supports vanilla 3DGS, FSGS, DNGaussian, and LGDWT-GS, and defines protocols for LLFF with 3 views, MipNeRF360 with 24 views, and the greenhouse dataset with 10 views per plant. Training is performed on a single NVIDIA A100 and is reported to converge in under 3 minutes per scene. In multispectral mode, new Gaussians are spawned where either RGB or NIR residuals are high, ensuring that both modalities participate in densification.

The greenhouse results separate the effect of multispectral supervision from that of DWT regularization. Averaged across Cotton, Grape, Sorghum, Tomato, and Houseplant, the single-channel setting reports 28.55 PSNR, 0.813 SSIM, and 0.422 LPIPS. Multispectral training without DWT improves this to 30.06 PSNR, 0.890 SSIM, and 0.258 LPIPS. The full multispectral plus DWT configuration reaches 30.51 PSNR, 0.892 SSIM, and 0.258 LPIPS. The paper therefore attributes the largest gain to multispectral shared-geometry training, with DWT providing an additional but more modest refinement.

5. Related greenhouse and plant splatting lineages

Two adjacent developments clarify how the GreenhouseSplat family connects to broader plant-oriented Gaussian-splatting research. The first is temporal phenotyping. "GrowSplat: Constructing Temporal Digital Twins of Plants with Gaussian Splats" reconstructs per-time-step 3DGS models from 15 calibrated RGB cameras in the NPEC Maxi-Marvin system, then aligns those snapshots into a 4D digital twin through feature-based coarse registration and ICP-based refinement (Adebola et al., 16 May 2025). Per-time-step reconstruction uses Splatfacto-MCMC, plant masks, an initial point cloud prior, an opacity-minimization loss, and a lighting factor. Temporal alignment filters Gaussians by scale and quaternion validity, computes normals and FPFH descriptors, performs Fast Global Registration after RANSAC, and refines with Colored ICP. The reported datasets are Sequoia with 40 time steps and average $o_i \in [0,1]$8 of 2.4 days, and Quinoa with 55 time steps and average $o_i \in [0,1]$9 of 1.4 days. No explicit quantitative metrics are reported, but the work shows that Gaussian splats can support temporally consistent greenhouse-style plant reconstruction beyond static scene capture.

The second is multi-spectral rendering with shared per-splat features. "Multi-Spectral Gaussian Splatting with Neural Color Representation" introduces a 3DGS model in which all spectra share geometry while per-splat feature embeddings are decoded by a shallow MLP into band-specific colors (Meyer et al., 3 Jun 2025). The method does not require cross-modal camera calibration, trains jointly over images from multiple independent cameras, and uses a multi-spectral-aware densification rule that considers the maximum residual-driven gradient across bands. On seven agricultural and vegetation scenes, the reported all-band average reaches PSNR 25.65, SSIM 0.763, and LPIPS 0.266, compared with 24.18, 0.718, and 0.344 for ThermalGaussian and 23.62, 0.721, and 0.334 for vanilla 3DGS. The paper also emphasizes vegetation-index rendering, particularly

$$\mathbf{s}_i \in \mathcal{S}=\{\text{fruits},\text{leaves},\text{background}\}$$0

from novel viewpoints. A plausible implication is that future greenhouse systems may unify the semantic mapping objectives of the active GreenhouseSplat architecture with the cross-spectral consistency of neural multi-spectral 3DGS.

Taken together, these neighboring works show that GreenhouseSplat is part of a broader methodological trajectory: from static greenhouse assets, to active semantic maps, to few-shot multispectral benchmarks, to temporal plant digital twins. The shared substrate is explicit splat-based geometry that is differentiable, photometric, and increasingly semantic or spectral.

6. Limitations, misconceptions, and likely directions

A recurrent misconception is that GreenhouseSplat denotes a single greenhouse-scale 3DGS method. The literature does not support that interpretation. One variant is a hybrid Octomap-plus-3DGS active mapper, one is a 2DGS simulation dataset, and one is a package name for wavelet-regularized sparse-view 3DGS with a greenhouse benchmark (Cuaran et al., 17 Jan 2026, Tabaa et al., 2 Oct 2025, Salehi et al., 23 Jan 2026). Another misconception is that all GreenhouseSplat systems are SLAM systems. In the active semantic mapping formulation, camera poses are assumed known from manipulator kinematics, and 3DGS is used purely for mapping rather than for SLAM (Cuaran et al., 17 Jan 2026).

Current limitations are substantial and paper-specific. The active semantic mapping system is evaluated entirely in simulation and explicitly notes likely difficulties from camera pose errors, noisy depth, foliage motion, and non-Lambertian surfaces in real greenhouses (Cuaran et al., 17 Jan 2026). The simulator-and-dataset version remains modest in scale, with only 8 row ends and 82 cucumber plants, and lacks crop, greenhouse, and lighting diversity (Tabaa et al., 2 Oct 2025). The LGDWT-GS greenhouse data are captured in a controlled setting with uniform background and static plants, so generalization to field or in-the-wild agricultural scenes remains open (Salehi et al., 23 Jan 2026). GrowSplat does not yet model non-rigid deformations explicitly and provides qualitative rather than numerical evaluation for temporal registration (Adebola et al., 16 May 2025). MS-Splatting remains a static-scene formulation and notes issues such as color dullness, sequential-capture misalignment, and reflective surfaces that may matter in greenhouse deployment (Meyer et al., 3 Jun 2025).

The published future directions are correspondingly heterogeneous. The active mapping paper points to real-world greenhouse trials, online relocalization, dynamic-scene handling, larger-scale 3DGS tiling and streaming, and planning strategies that may combine reinforcement learning with graph-based search (Cuaran et al., 17 Jan 2026). The simulator paper emphasizes larger datasets, more crops, richer layouts, and sim-to-real validation (Tabaa et al., 2 Oct 2025). The LGDWT-GS package suggests frequency-guided densification and pruning, broader multispectral or hyperspectral extensions, and application to less controlled agricultural environments (Salehi et al., 23 Jan 2026). GrowSplat proposes biological priors, interpolation across unobserved time steps, and quantitative biomass-oriented evaluation (Adebola et al., 16 May 2025). MS-Splatting suggests extensions toward hyperspectral rendering, dynamic 3DGS, and tighter robotics integration (Meyer et al., 3 Jun 2025).

These trajectories suggest that GreenhouseSplat is best understood not as a closed method but as a research direction at the intersection of greenhouse robotics, plant phenotyping, radiance-field simulation, and explicit splat-based scene representation. The unifying technical claim across the literature is narrower and clearer: Gaussian splats provide a flexible intermediate representation that can simultaneously serve high-fidelity rendering, geometric reconstruction, semantic or spectral reasoning, and, in some formulations, active viewpoint planning for greenhouse environments.