非破壞性感測技術應用於不同空間規模之農業光譜資訊分析

Abstract

基於影像和光譜學的非性破壞性測量技術在今日已成為獲取農業和作物相關資訊的重要工具。然而由於農業非性破壞性測量應用的情境多元，不同研究問題間的空間尺度差異甚大，從基於遙測技術的橫跨多縣市的大範圍監測至單一田區，甚至近距離非破壞性的單一植株監測，皆運用非性破壞性測量相關技術，而不同空間尺度會對資料的需求以及分析方法上帶來差異。本研究旨在探討不同空間尺度下，非性破壞性測量技術在應用於農業資訊的分析上時分析工具上的探討，以及研究上的差異和限制。本研究提供了一套適用於不同空間尺度下，影像或光譜資料的分析框架，將農業遙測的流程區分為光譜資料獲取、數據校正和前處理、特徵轉換、模型建立等步驟，並回顧與展示了每個步驟中可供選擇的工具或方法，以統整不同空間尺度下的所使用的方法論。為了展示非性破壞性測量技術在不同空間尺度下的應用，本研究透過三個分別在不同空間尺度下的案例，顯示非性破壞性測量技術應用於農業資訊分析的成果。在大空間尺度的案例一中，本案例使用基於無人機收集的可見光影像，使用了色彩特徵，以及本案例所提出的基於灰度共生矩陣的紋理特徵 (Gray-Level Co-occurrence matrix vector, GLCMv)，並結合機器學習的 XGBoost 演算法，在嘉義與台南兩縣市進行田區地物種類的辨識。結果顯示，最佳特徵組合下的準確度達到82%，同時相較於傳統紋理特徵，本案例所提出的 GLCMv 能顯著的提升模型準確度。在中等空間尺度的案例二中，本案例使用了基於一款商用自走式割草機器人為載具，加裝可見光相機後，在多塊公園和開放式空間的草地上，於執行割草任務的同時進行影像收集，並利用植生指數 (Vegetation Index, VI) 以及 k-means 分群演算法來計算草坪上草和綠草的覆蓋率，以及結合機器人的座標位置，利用覆蓋率指標繪製地圖和尋找空間中低覆蓋率的熱點區域。結果顯示基於影像所得到的草覆蓋率和綠草覆蓋率兩項指標皆和專家的標註結果存在有相同的趨勢，案例中也建立起一套影像指標和專家標註間的轉換表。同時在地圖繪製的部分，不論是使用單次影像或是時序性割草任務，都能有效觀察到熱點區域的位置，且所辨識的熱點區域和人工在草地旁觀察到的結果一致，而在將影像指標轉換為專家標註的系統後，能更有效的提供管理上的建議。同時該系統在用於高爾夫球場等高度管理的場域時，也能在影像指標的層級上辨識出覆蓋率較低的區域。在小空間尺度的案例三中，展示了使用手持式高光譜儀，在單株芋植株的葉面上測量反射率，以預測一個月後芋地下部的軟腐病 (Bacterial soft rot) 罹病情形。本案例將芋地下部因軟腐病造成的重量折損率，透過 0% 和 5% 兩個閾值轉換為折損與否以定義罹病情形，並在光譜特徵上使用了主成分分析 (Principal Component Analysis, PCA) 、稀疏化以及基於平均和變異係數等統計量的特徵篩選三種不同的特徵篩選與處理技術來處理高維度的高光譜儀資料，利用基於深度學習的自編碼器 (Autoencoder) 建立異常檢測器，辨識罹病的異常植株，並利用特徵篩選的方式挑選出對分類結果有高影響力的關鍵波段。結果顯示，在 5% 的折損閾值，以及使用 forward selection 進行特徵篩選時，本案例能在測試中達到最高 0.81 的 F1 score，有效的及早預測出罹病植株。而同時本案例也篩選出 17 個重要的波段，分別位於紫外光、綠光、紅光、紅光邊緣、近紅外光、短波長紅外光等範圍中，並且有 12 個波段已在前人研究中被提及和葉綠素、黃酮類化合物、蛋白質與氮含量和水分等植物體內的化合物含量有關。在不同案例間，基於不同的空間尺度以及研究目標上而存在有載具、波段和資料處理與模型建立方式上的差異。但如本研究所提出的框架所示，在每個步驟中，不同尺度下的案例在欲達成的目標上是一致的，方法選擇上的差異則是由於資料的收集或形態上的不同而導致。

Non-destructive sensing techniques based on imaging and spectroscopy have become an important tool for obtaining agricultural and crop-related information today. However, due to the diverse contexts of agricultural non-destructive measurement applications, there existed huge differences in spatial scales across different research questions. These range from large-scale monitoring by remote sensing across counties to single fields, and even close-range sensing of individual plants, all of which belong to non-destructive measurement-related technologies. Different spatial scales result in variations in data requirements and analytical methods. This study aims to explore the analytical tools and research differences and limitations associated with the application of non-destructive measurement technologies in agricultural information analysis across different spatial scales. This study provides a framework for analyzing image or spectral data across different spatial scales, dividing the agricultural remote sensing process into steps such as spectral data acquisition, data correction and preprocessing, feature transformation, and model establishment. It reviews and demonstrates the tools or methods available at each step to integrate the methodologies used across different spatial scales. To demonstrate the application of non-destructive measurement techniques at different spatial scales, this study presents three cases under different spatial scales to shows the application of non-destructive measurement techniques in agricultural information analysis. In Case Study 1, which focuses on a large spatial scale, this case uses visible light images collected by drones, employing color features and the texture features based on the gray-level co-occurrence matrix vector (GLCMv) proposed in this case, combined with the XGBoost algorithm of machine learning, to identify the types of land cover in farmland areas in Chiayi and Tainan counties. The results showed that the accuracy reached 82% under the optimal feature combination. Compared to traditional texture features, the GLCMv proposed in this case significantly improved the model accuracy. In Case 2, which focuses on the medium spatial scale, this study utilized a commercial autonomous lawn mower as the carrier, equipped with a visible light camera, to collect images while performing mowing tasks on multiple parks and open-space grasslands. Vegetation Index (VI) and the k-means clustering algorithm to calculate the coverage rates of grass and green grass on the lawn. By combining the robot's coordinate position with the coverage rate indicators, maps were created and low-coverage hotspot areas in the space were identified. The results showed that the two indicators—grass coverage and green grass coverage—derived from the images exhibited the same trends as the expert annotations. A conversion table between the image indicators and expert annotations was also established in this case. In terms of map generation, whether using a single image or sequential mowing tasks, hotspot locations can be effectively identified, and the identified hotspots align with those observed manually alongside the grass. After converting image metrics into expert annotations, the system can provide more effective management recommendations. Furthermore, when applied to highly managed environments such as golf courses, the system can identify areas with lower coverage at the image indicator level. In Case Study 3, which was conducted on a small spatial scale, a handheld hyperspectral imager was used to measure the reflectance of the leaves of individual taro plants to predict the incidence of bacterial soft rot in the underground parts of the taro plants one month later. The weight loss rate caused by soft rot in the underground parts of taro plants was used to determine whether damage had occurred, using two thresholds of 0% and 5% to define the abnormal or not. Spectral features were processed using three different feature selection and processing techniques: principal component analysis (PCA), sparsification, and feature selection based on statistical measures such as mean and coefficient of variation to process high-dimensional hyperspectral data. An anomaly detector was established by a deep learning-based autoencoder to identify abnormal plants with disease, and feature selection was used to select key bands with a high impact on the classification results. The results showed that at a 5% loss threshold and using forward selection for feature selection, this case achieved a maximum F1 score of 0.81 in the test, effectively predicting diseased plants at an early stage. Additionally, this case identified 17 important bands spanning ultraviolet, green, red, red edge, near-infrared, and short-wave infrared regions, with 12 of these bands previously mentioned in prior studies as being associated with plant compounds such as chlorophyll, flavonoids, proteins, nitrogen content, and moisture levels. Across different cases, differences in spatial scales and research objectives lead to variations in platforms, bands, data processing, and model development methods. The objectives sought in each step are consistent across cases at different scales, while differences in method selection arise from variations in data collection or form.