Applications of Image Analyses in Agriculture: Challenges and Opportunities

Abstract

作物生育狀況或性狀調查一向是作物栽培管理或培育新品種很關鍵的工作，隨著科技的進步，透過蒐集大量的影像將作物調查步驟自動化、高通量化，已經成為作物科學相關研究的趨勢。然而，收集而來的大量影像資料，除了面臨的是資料儲存空間以及資料分析需要GPU等高階設備支援的挑戰，如何將田間收集而來包含大量雜訊的影像資料，在短時間內轉換為可用的資訊，以也是統計學家必須面對的課題。在本次的演講中，我會分享影像辨識在農業應用上遇到的困難與挑戰，並舉例說明我們目前結合統計方法與農業領域知識，如何加速影像辨識與提升分析速率。影像辨識在農業上的應用也與其他工業或醫學應用上有很大的需求差異，精確度並非主要追求的目標，配合農業產值，如何快速得到辨識結果、並將影像辨識結果大範圍的應用，符合作物生長季節做到精簡管理人力才是農業應用的需求。因此，影像辨識在農業的應用需要不一樣的思維，歡迎統計專家投入，與農業專家合作開發兼具穩定性與實用性的影像分析流程。 ? The identification of crop phenology and traits is essential for crop management or the breeding. With the advancement of science and technology, it becomes feasible to automatically collect crop images with high throughput. However, in addition to the inquiries of data storage and high-end device such as GPUs for data analyses, the large amount of noisy image data makes it challenging to deliver meaningful and usable information through data analyses in limited time. In this talk, I will share the difficulties and challenges encountered in the application of image recognition in agriculture, and the examples of how we currently combine statistical methods and domain knowledge in the agricultural field to accelerate image recognition. The goals of the image applications in agriculture is also very different from other industrial or medical fields. With the relative low value of agricultural products, how to rapidly obtain the results and apply them in a large scale may be the ultimate goal due to the short cropping season and enormous working load for field management. Therefore, the application of image recognition in agriculture requires different thinking. Statistical experts are welcome to cooperate with agricultural experts to develop image analysis processes that achieve both stability and practicality.