溫室微型害蟲辨識系統之優化與線上自主學習架構之研究

Abstract

在作物生長期間，蟲害被認為是對農業生產的最大威脅之一。它們危害農作物的生長，降低產量，對農業經濟收益造成嚴重損失。因此，有效管理蟲害的發生，對於農業從業者至關重要。若想要達到有效管理的目標，需要透過即時且準確的害蟲種類和數量數據，擬定正確的管理對策。本研究室先前已開發出一套基於物聯網技術的智慧型蟲害管理系統(intelligent integrated pest and disease management, I2PDM)，使用相機拍攝黏蟲紙，並利用深度學習辨識害蟲種類與數量。本研究目的為優化I2PDM系統中所使用的害蟲辨識模型，使其提供更準確的蟲害資訊，我們使用SRGAN影像增強模型，用於強化害蟲影像的視覺特徵。同時，將害蟲的尺寸納入分類模型中提供額外的資訊，以提高辨識的準確性。以影像增強、尺寸特徵與兩者結合的方法，提出三種新的模型架構，相比於原始的架構，經過優化後的模型分別能提升約2.7%、2.3%與4.4%的F1-score。此外，我們提出一套自動化線上自主學習架構，利用I2PDM系統數據流的優勢，持續收集更多的害蟲影像擴增訓練集，再利用新影像對基礎模型進行優化訓練，用以解決傳統資料收集與訓練模型所需的大量人力與時間。並且透過樣本清理演算法，搭配高斯混合模型對新進樣本進行篩選，確保新收集之樣本正確性，以及與正確樣本之間的特徵相似性，實現自動化樣本收集和模型再訓練的流程。測試結果顯示，在使用三年資料與四種不同基礎模型的情境下，其最終模型都能有效提升約2.6% 到5.8%的水準。後續利用MQTT、ZMQ與TCP等網路傳輸協議，將線上自主學習架構實際部署到I2PDM系統中。經 使用五個月的資料進行測試比較，與基礎模型相比可達2.7%的效能提升。除軟體優化外，本研究亦進行硬體升級，以使用Arducam 64MP替換原有的Raspberry pi camera v2相機模組，最終得到約2倍DPI的影像，取得更細微的害蟲特徵，測試結果顯示，使用新相機所訓練的分類模型相比使用原始相機，約有4.4%的F¬1-score提升。

During crop growth, pests are considered one of the biggest threats to agricultural production. They damage crop growth, reduce yields, and cause significant economic losses to the agricultural industry. Therefore, effective management of pest occurrence is crucial for agricultural practitioners. To achieve effective management goals, accurate and real-time data on pest species and quantities are needed to formulate appropriate management strategies. Our research laboratory has previously developed an Intelligent Integrated Pest and Disease Management (I2PDM) system, which utilizes cameras to capture images of sticky traps and employs deep learning to identify pest species and quantities. The objective of this study is to optimize the pest recognition model used in the I2PDM system to provide more accurate pest information. We employ the SRGAN image enhancement model to enhance the visual features of pest images. Additionally, we incorporate the size of pests into the classification model to provide additional information for improved recognition accuracy. Three new model architectures are proposed using image enhancement, size features, and a combination of both, which achieve F1-score improvements of approximately 2.7%, 2.3%, and 4.4%, respectively, compared to the original architecture. Furthermore, we propose an automated online self-learning framework that leverages the data flow of the I2PDM system to continuously collect more pest images for augmented training sets and optimize the base model through retraining with new images. This approach solves the challenges of traditional data collection and the significant manpower and time required for model training. By utilizing a sample cleaning algorithm with a Gaussian Mixture Model, the newly collected samples are filtered to ensure their correctness and feature similarity with correct samples. This automated process of sample collection and model retraining is achieved. With three years of data and four different base models, the final models show effective improvements ranging from approximately 2.6% to 5.8%. The online self-learning framework is deployed in the I2PDM system using network transmission protocols such as MQTT, ZMQ, and TCP. Through a five-month data test, it shows a 2.7% performance improvement compared to the base models. In addition to software optimization, we also upgraded the hardware by replacing the original Raspberry Pi Camera V2 module with the Arducam 64MP camera module, resulting in images with approximately twice the DPI. This upgrade captures finer pest features. The test results indicate a 4.4% improvement of the F1-score when using the model trained with the new camera compared to the model trained with the original camera.