應用深度學習演算法於溫室網紋洋香瓜花朵辨識與開花期預測

Abstract

溫室栽植網紋洋香瓜(Cucumis melo, L. var reticulatus Naud.)需要精準的田間管理與授粉以確保果實的高品質和高產量。然而，人工授粉工作費力耗時，有限的授粉時間更進一步提高栽植管理的難度，增加人力需求。隨著智慧農業的引進，開發自動化授粉系統成為一個有潛力的解決方案。本研究旨在整合機器學習方法，解決自動化授粉任務中雄雌花識別與預測最佳授粉期間的需求，輔助溫室網紋洋香瓜自動化授粉系統之開發。 針對上述需求，本研究開發三階段流程，包括網紋洋香瓜花朵識別、花朵發育監測和開花期間預測。在第一階段，基於YOLOv9架構開發網紋洋香瓜雄雌花辨識模型和花瓣分割模型，用於辨識花朵性別並提取花瓣輪廓。辨識模型的F1-score達到0.975，而分割模型的遮罩準確度均值(mask average precision, mask AP)達到0.928。第二階段對花朵發育程度提供更精確的量化方式，提出自定義之「開花角度」與其計算演算法，其後利用計算出之開花角度建立開花曲線，用於監測並分析授粉的關鍵起始與結束時間。第三階段，進一步加入蒐集到環境感測器之數據，包括：溫度、濕度與光照度，搭配季節性自回歸整合移動平均含有外生變量(seasonal autoregressive integrated moving average with exogenous factors, SARIMAX)、長短期記憶(long short-term memory, LSTM)和自編碼器(autoencoder)等三階段時間序列模型，用以建議開花曲線的預測模式。最終模型效能比較，以自編碼器模型預測之曲線可達均方根誤差(mean squared error, MSE) 0.009為最佳，提供較為準確的開花曲線。此外，自編碼器模型在檢測開花結束時間和預測開花期間方面也顯示出卓越的性能。此模型於兩組冬季開花曲線上預測的開花期間僅有兩小時的誤差；並且能在三組測試集數據上都提供完整的開花期間預測。 本研究應用機器視覺於網紋洋香瓜花性辨識，並結合提出之開花角度演算法，監測花朵發育。進一步整合開花曲線與環境數據，利用時間序列模型預測開花期間，旨在優化最佳授粉期間的判讀。透過本研究提出之三階段流程，推動自動化授粉系統開發的進程，在未來有望減輕溫室網紋洋香瓜栽植管理的人力負擔。

Muskmelon (Cucumis melo, L. var reticulatus Naud.) cultivation in greenhouse environments requires precise pollination and field management to ensure high fruit quality and yield. However, manual pollination is time-consuming and labor-intensive. The limited pollination window further complicates the cultivation management, increasing the labor demands. With the introduction of smart agriculture, developing an automated pollination system has become a promising solution. This study aims to integrate machine learning methods to address the needs of identifying muskmelon flower genders and predicting the optimal pollination period, supporting the development of an automated pollination system for greenhouse cultivated muskmelon. To meet these requirements, this study developed a three-stage system comprising muskmelon flower identification, flower development recording, and flowering period prediction. In the first stage, muskmelon flower gender identification (MFGI) model and petal segmentation (PS) model based on YOLOv9 architecture were developed to identify flower genders and extract petal contours. MFGI model achieved 0.975 F1-score, and the PS model attained a mask AP of 0.928. In the second stage, to determine flower development more accurately, a self-defined measure “flowering angle” and its calculation algorithm were proposed. The calculated flowering angle was used to establish flowering curves, analyzing the flowering start and end point for pollination. In the third stage, environmental data including temperature, humidity, and solar irradiance were integrated with flowering curves. Subsequently, time series models of seasonal autoregressive integrated moving average with exogenous factors (SARIMAX), stacked long short-term memory (LSTM), and the autoencoder model were developed and compared to suggest the most suitable method for flowering period forecasting. The model comparison revealed that the autoencoder model achieved the best performance, with a curve mean squared error (MSE) of 0.009, providing more accurate flowering curves. Additionally, the autoencoder model demonstrated excellent performance in detecting the end of the flowering period and predicting the flowering duration. For two sets of winter flowering curves, the prediction error for the flowering period was only two hours. The autoencoder was capable of providing complete predictions for the flowering period across all three datasets. This study applied computer vision methods for muskmelon flower gender identification and flower development recording. Further integrated flowering angel with environmental data, utilized time series models to predict flowering periods, aiming to optimize the interpretation of the optimal pollination window. The proposed three-stage system advanced the development of automated pollination systems, potentially reducing the labor burden of greenhouse muskmelon cultivation management in the future.