LeafInst - Unified Instance Segmentation Network for Fine-Grained Forestry Leaf Phenotype Analysis: A New UAV based Benchmark Journal: Computers and Electronics in Agriculture
Taige Luo, Junru Xie, Chenyang Fan, Bingrong Liu, Ruisheng Wang, Yang Shao, Sheng Xu, Lin Cao · 2026
Read LeafInst - Unified Instance Segmentation Network for Fine-Grained Forestry Leaf Phenotype Analysis: A New UAV based Benchmark Journal: Computers and Electronics in Agriculture by Taige Luo, Junru Xie, Chenyang Fan, Bingrong Liu, Ruisheng Wang, Yang Shao, Sheng Xu, Lin Cao (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-03616.txt.
Introduction
Taige Luo and seven colleagues teach drones to outline every leaf on a young forest tree. Tree breeding needs fine leaf measurements, yet almost all phenotyping work targets big-leaf farm crops, not saplings in open fields where leaves vary in scale, brightness, and shape. Their 2026 study surveys field-grown poplar saplings by drone and labels over twelve hundred branches, nearly twenty thousand individual leaves, pixel by pixel.
The released Poplar-leaf dataset is the first pixel-level instance segmentation set built for forestry leaves in open-field scenes. Their network, LeafInst, adds a progressive multi-scale feature pyramid, a module tuned to irregular leaf shapes, and a fusion scheme that cuts redundant features. It beats strong rivals, including a leading transformer segmenter, on both the new set and a large public farm benchmark, and holds up in unseen scenes.
Catalog record
| CREATOR | Taige Luo, Junru Xie, Chenyang Fan, Bingrong Liu, Ruisheng Wang, Yang Shao, Sheng Xu, Lin Cao |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 85298 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.03616
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
LeafInst - Unified Instance Segmentation Network for Fine-Grained Forestry Leaf Phenotype Analysis: A New UAV based Benchmark Journal: Computers and Electronics in Agriculture Taige Luo, Junru Xie, Chenyang Fan, Bingrong Liu, Ruisheng Wang, Yang Shao, Sheng Xu, Lin Cao Abstract Intelligent forest tree breeding has spurred significant advancements in plant phenotyping research. However, the existing studies predominantly concentrate on agricultural scenarios involving large-leaf crops, leaving a notable gap in the fine-grained research on the leaves of sapling trees in open-field environments. Moreover, leaves in natural settings present potential challenges due to variations in scale, brightness, and shape. To tackle these critical issues, this study employed unmanned aerial vehicles (UAVs) to survey field-grown saplings and collect 1,202 branches, encompassing a total of 19,876 leaf instances. Leveraging this data, we introduce and publicly release an open-source UAV RGB remote sensing dataset, named Poplar-leaf. Significantly, this dataset represents the first pixel-level annotated instance segmentation dataset specifically tailored for forestry leaves in open-field scenes. Regarding accurate segmentation, a novel network architecture, LeafInst, is proposed. For scale changes, the Asymptotic Feature Pyramid Network (AFPN) is integrated into the Neck component, which effectively enhances progressive multi-scale visual perception. Additionally, a novel Dynamic Asymmetric Spatial Perception (DASP) module is designed to optimize feature extraction for irregularly shaped leaves. Based on it, we propose a dual-residual Dynamic Anomalous Regression Head (DARH) block to enhance the detection feature fusion. The proposed Top-down Concatenation–decoder Feature Fusion (TCFU) fuses the output with the bounding box regression branch, effectively mitigating redundant feature representations. Quantitative experiments demonstrate that, on the Poplar-leaf dataset, LeafInst achieves a 7.1% improvement in segmentation mAP (68.4) compared with YOLOv11, and a 6.5% gain over the transformer-decoder-based MaskDINO. Furthermore, on the public large-scale agricultural benchmark dataset PhenoBench, LeafInst outperforms MaskDINO by 3.4% in box mAP, reaching 52.7 mAP. Qualitative experiments show that LeafInst outperforms the large visual models in zero-shot scenes. Finally, we present two downstream applications to highlight the real-world utility of the proposed model, while also supporting its robustness across diverse scenarios: robust leaf segmentation in complex scenes, and a novel high-throughput Leaf Growth Condition Indicator (LGCI) for quantitative assessment of leaf phenotypic development (from 1850 leaf instances). Keywords: Deep Learning, Computer Vision, UAV remote sensing, Instance segmentation, Leaf phenotype analysis, Intelligent forestry 1 Introduction With the advancement of smart forestry, research combining "intelligent" models with practical applications has been growing rapidly. Current deep learning-based studies have covered numerous domains, including wheat ear phenotyping identification [ 52 , 17 ] , forest tree species recognition [ 22 , 67 ] , fine-grained classification of forest land use [ 50 , 39 , 57 ] , and crop disease/pest detection [ 13 , 3 , 25 ] . These studies have demonstrated significant practical value. As the most critical component of forests, trees play a vital role in understanding forest dynamics. Forestry breeding represents a significant silvicultural challenge that relies on precise plant phenotyping analysis to select superior varieties with optimal growth characteristics, luxuriant foliage, and adequate water content [ 72 , 45 , 65 ] . Conventional breeding remains an effective approach for forest tree improvement internationally. Leading forestry nations should place greater emphasis on phenotypic and genetic evaluation of germplasm resources, along with the exploration and utilization of unique genetic materials, to establish comprehensive core collections. Long-term breeding programs have entered a phase of accelerated development, where continuous genetic gains in improved varieties are being achieved through techniques such as hybrid breeding and multiple-trait integration breeding. However, systematic genetic evaluation of germplasm resources and core collection establishment have not been conducted for most tree species. Figure 1 : Challenges in forestry leaf phenotyping. (a) Scale changes, (b) Brightness changes, (c) Shape changes. Key technologies such as parental selection, early-stage selection, and stable high-yield seed orchards remain unbroken. The traditional manual evaluation system relies on field inspections, which are time-consuming, labor-intensive, subjective, and potentially damaging to parent trees. This system fails to establish intelligent standardized evaluation criteria and frameworks, resulting in current breeding practices that cannot meet growing consumer demands. In terms of production capacity, taking China as an example, official statistics from 2016 indicate that while the country possesses 4.68 billion mu (approximately 312 million hectares) of forest land, it faces severe timber shortages. The total timber consumption reached 522 million cubic meters, with projections suggesting an annual supply-demand gap exceeding 450 million cubic meters post-2025. This implies an annual transition of 50 million cubic meters of timber production from natural forests to plantations. Currently, most seedlings transplanted to plantation areas are immature, and traditional manual analysis methods demonstrate poor adaptability. Figure 2 : The workflow and downstream applications of the proposed framework LeafInst. Conventional breeding programs primarily emphasize "fast growth, stress resistance, and wood quality" while neglecting the carbon sequestration capacity of individual tree leaves. A tree’s carbon sequestration potential is directly determined by its biomass accumulation rate, which correlates closely with leaf phenotypic traits (e.g., Leaf Area Index (LAI) [ 14 ] , Specific Leaf Area (SLA) [ 16 ] , trunk morphology (DBH) [ 30 ] , height increment [ 38 ] , and root architecture [ 18 , 61 ] ). "Early-stage selection" has emerged as a core challenge in forest tree improvement. As a fundamental indicator of vegetation growth, leaf-derived parameters can effectively evaluate developmental potential in trees [ 24 ] . Individual Tree Segmentation (ITS) has emerged as a research focus in recent years. However, forest tree breeding requires finer-grained phenotypic parameters, driving the trend toward more "granular" research subjects. To assist agricultural/forestry practitioners and ecologists in extracting fine-grained tree structures, analyzing plant phenotypes, and improving superior variety selection, researchers have conducted experiments on tree branches [ 5 , 60 , 51 ] , crown width [ 40 , 34 , 66 ] , and DBH [ 58 , 37 ] . These studies significantly contribute to understanding plant health status and predicting growth trends, enabling the selection of elite varieties at early seedling stages for cultivating high-yield specimens. Nevertheless, few forestry studies address the most critical phenotypic parameter - leaves. Current leaf-level research primarily focuses on visibly diseased/pest-infected leaves [ 26 , 19 ] and agricultural crop leaves [ 2 , 54 , 23 ] . This study represents the first instance-based investigation of forest tree leaves, establishing a fundamental milestone for this field. Building upon agricultural "crop phenomics platform" experiences, this framework can be extended to deploy fixed cameras in plantation test areas to analyze leaf growth rates through video streams, develop dynamic genetic gain evaluation models, and open-source the dataset on global platforms. These efforts will provide data support for worldwide forest ecological breeding initiatives. Compared to agricultural crops, tree leaves present distinct challenges: dense distribution, broad leaf morphology, and close-range sampling. In addition, agricultural leaves often present orthorectified images, which reduce the angles transformation. However, the height and density of trees expose them to environmental variables, leading to uneven illumination, high occlusion density, and morphological deformations. Overall, forestry leaf UAV images present three major challenges: 1. Scale changes (as the drone lens expands and contracts). Scale variations arise from changes in camera configurations and task requirements, leading UAV imagery to focus on different spatial levels, including