結合深度學習演算法及自主巡檢載具於智慧蘆筍生長管理系統之優化

Abstract

蘆筍為一高經濟價值的溫帶作物。過往研究發現，採用留母莖栽培法(mother stalk method)於溫室種植能使蘆筍適應臺灣的濕熱氣候。此方法透過保留一定數量母莖進行養分供給，以提高嫩莖生長效率並增加產量。然，此方法需大量人力管理溫室內的母嫩莖數量，增加田間管理成本。本研究旨在整合並優化前期研究之自主導航巡檢載具及影像辨識模型，以自動化技術建立更高效溫室蘆筍生長管理系統。 在自主導航巡檢載具方面，本研究藉由分層密度聚類算法(Hierarchical-DBSCAN)改進現有的DBSCAN LiDAR居中導航策略。透過計算局部距離建立層次化的密度樹狀結構，根據設定最小群集資料點數量自動挑選分群結果，藉此機制使演算法更適應於田間複雜多變點雲分佈情境。本研究亦開發一套車輛軌跡自動評估演算法以衡量居中導控成效。實驗結果證實，無論於靜態或動態試驗中，HDBSCAN均可實現更準確的分群表現與居中穩定性。在超寬頻(Ultra-Wideband, UWB)定位部分，本研究通過理論分析和實驗調整系統錨點(anchor)架設方式，以減少信號傳輸中因擬葉遮擋造成的Non-Line-of-Sight現象。動態UWB實驗在生長高密度階段之40公尺x 10公尺田區內，實現將均方根誤差(root-mean-square error, RMSE)誤差從34.1公分降至22.6公分，顯著提升此定點監測系統之可用性。在生長辨識模型優化方面，本研究採用基於Transformer架構的Mask DINO模型配合Swin-Large特徵提取骨幹，在交集聯集比(intersection over union, IoU)設定為0.5閾值下，以平均準確度(average precision, AP50)為評估指標，Box AP與Mask AP分別達到94.6%與94.1%的水準，相較先前採用的Mask R-CNN模型，無論在預測植株遮罩、數量，和區域密度等方面，均提供更精確的目標識別與生長資訊。 最後，透過整合前期開發的監測網頁介面，將UWB定位資訊與影像辨識模型結合，可提供管理者易於檢視的田間實際密度分布資訊，生長監測成果亦與硬體開發業界合作，將其應用於藥肥自動噴藥車之開發。透過UWB資訊與密度查詢API設計，提供更精確的藥肥變量施灑作業。藉由本研究提出的蘆筍生長管理系統，結合機電整合技術及深度學習演算法，可望為農民提供更高效的田間監測與管理工具，實現減輕田間作業的人力需求與負擔。

Asparagus is a high-value temperate crop. Previous research has shown that using the mother stalk production in greenhouse helps asparagus adapt to Taiwan's humid and hot climate. This method improves growth efficiency and increases yield by retaining a certain number of stalks to supply nutrients. However, this method requires significant manual labor to manage the number of stalks and spears in the greenhouse, thereby increasing field management costs. This study aims to integrate previous research findings and optimize a self-guided patrol robot and an image identification model to reduce labor demand through automation, establishing a more efficient greenhouse asparagus growth management system. In terms of the self-guided patrol robot, this study improves the existing LiDAR-centered navigation strategy using the Hierarchical Density-Based Spatial Clustering of Applications with Noise (Hierarchical-DBSCAN) algorithm. By calculating local distances to establish a hierarchical density tree structure and selecting clustering results based on the minimum number of cluster data points set, the algorithm adapts better to complex and variable point cloud distribution scenarios in the greenhouse. This study also develops an automatic robot trajectory evaluation algorithm to measure centering control effectiveness. Experimental results demonstrate that HDBSCAN achieves more accurate clustering performance and centering stability than DBSCAN in both static and dynamic tests. For the Ultra-Wideband (UWB) positioning system, theoretical analysis and experiments were conducted to adjust the system anchor setup, reducing Non-Line-of-Sight (NLOS) phenomena caused by pseudo-leaf obstructions during signal transmission. Dynamic UWB experiments in a 40 m x10 m field during the high-density growth stage reduced root-mean-square error (RMSE) from 34.1 cm to 22.6 cm, significantly enhancing the usability of this point monitoring system. In optimizing the growth identification model, this study adopts the Mask DINO model based on the Transformer architecture with the Swin-Large backbone. Using an intersection over union (IoU) threshold of 0.5 and evaluating with average precision (AP50), the model achieved Box AP and Mask AP of 94.6% and 94.1%, respectively. Compared to the previously used Mask R-CNN model, it provides more accurate target identification and growth information in terms of plant masking, count, and area density. Finally, by integrating the previously developed monitoring web interface with UWB positioning information and the asparagus identification model, the system provides easy-to-view field density distribution information for managers. The growth monitoring results have also been applied in collaboration with the hardware development industry to develop an automatic variable pesticide spraying robot. Using UWB information and a density query API design, it enables more precise variable pesticide and fertilizer application. The asparagus growth management system proposed in this study combines mechatronics integration technology and deep learning algorithms, offering farmers more efficient field monitoring and management tools, thereby reducing labor demand and burden in field operations.