Abstract: As a cornerstone of global food security, maize (Zea mays L.) demands the development of advanced breeding strategies to enhance yield, stress resistance, and adaptability. High-throughput, non-destructive, and accurate acquisition of plant phenotypic parameters is therefore crucial for accelerating the intelligent breeding process. While three-dimensional (3D) point clouds acquired by LiDAR provide unprecedented detail of plant architecture compared to 2D imaging, their widespread application in real-world field conditions faces a significant bottleneck. The primary challenge lies in the accurate instance segmentation of individual plants within dense populations. Complex plant morphologies, extensive canopy adhesion—where leaves of adjacent plants are tightly interwoven—and mutual occlusion among plants create a convoluted spatial arrangement that traditional clustering or geometry-based algorithms fail to resolve. This frequently leads to fragmented or incorrectly merged plant instances, severely constraining the efficiency and reliability of subsequent phenotypic data extraction.To address these critical limitations, this study proposes a novel instance segmentation framework, 3D-MaizeNet. This framework integrates LiDAR data with deep learning to accurately extract individual maize plants and achieve high-throughput measurement of key agronomic traits, such as plant height and stem height. The proposed methodology comprises three innovative stages. First, to preserve the structural integrity of individual plants, which is often compromised by simplistic preprocessing methods, we introduce an adaptive block segmentation technique based on crop row detection. This method leverages the inherent row-planting pattern of agricultural fields to intelligently partition the large-scale point cloud into meaningful, plant-centric blocks. This approach effectively minimizes interference from overlapping canopies in adjacent rows, facilitating the construction of a high-quality, field-derived point cloud dataset that is critical for robust model training. Second, to enhance the network's ability to learn fine-grained geometric details from complex canopy structures (e.g., leaf angles and stem orientations), we design a local spatial encoding module. Concurrently, we integrate an attention-based aggregation down-sampling module to mitigate the loss of key spatial features during feature extraction, selectively preserving salient information that is vital for distinguishing between closely packed plants.Finally, leveraging the high-fidelity instance segmentation results, we established an automated pipeline for the high-throughput quantification of plant height and stem height—two pivotal phenotypic parameters closely related to yield potential and lodging resistance. Experimental results from field-scanned data rigorously validated the efficacy of our proposed framework. 3D-MaizeNet achieved a mean Average Precision (mAP) of 95.9% and an overall accuracy of 96.4% in instance segmentation, demonstrating its superior performance in both identifying and delineating individual plants. Furthermore, the agronomic validity of the extracted traits was confirmed through strong correlations with manual ground-truth measurements, with coefficients of determination (R²) of 0.91 and 0.89 for plant height and stem height, respectively. The high-throughput and precise phenotyping platform developed in this study provides critical technical support for advancing maize genomics research, facilitating Genome-Wide Association Studies (GWAS), and ultimately accelerating the pace of molecular breeding for next-generation crop improvement.