輪廓線估計法之探索: 以葉面積預測為例

Abstract

隨著科技的進步，影像分析技術在農業領域受到了廣泛的關注。許多農業專家和研究人員認識到，利用影像可以更有效率的監測田間作物的生長情況，觀察植株外觀、或是反射回來的光波都可以有效的獲取田間資訊、即時去進行適當的措施從而提高作物產量。然而，有些影像是從上向下拍攝的，如果作物密度大或葉片較大，下層葉片很容易被遮蔽，這使得專家難以準確評估下層葉片的生長狀況。為了解決這個問題，本研究開發了一個工作流程來估算葉片重疊面積，使用OpenCV影像處理工具進行影像前處理、影像分割和物體輪廓偵測。通過計算葉片輪廓的梯度值，瞭解葉片輪廓變化的特性，例如曲率和邊界形狀的變化。並使用DTW（Dynamic Time Warping）動態時間校正演算法計算不同葉片梯度之間的相似距離，利用影像中的輪廓梯度訊息，將部分葉片對應回最相似的單一葉片位置，進而推估被遮蔽葉片的面積。隨後，應用統計分析方法來驗證推估出的葉片面積是否與真實值相似。根據統計分析結果，將得到的數據利用對數轉，換得到該方法利用平均數所得到的整體MAPE平均值為6.39%，平均歸一化均方根誤差（NRMSE）為0.10%。在刪除極端值後，平均MAPE降低到4.60%，平均NRMSE降低到0.06%。這些結果表明，刪除極端值提高了葉面積估算的準確性。此方法不僅可以讓農業專家和農民更好地了解田間植物的生長狀況，從而做出更精確的管理決策，未來可將這一方法應用於各種物體和場景，例如環境監測和自然資源管理。通過影像分析技術，我們可以更準確地瞭解不同場景中物體的分佈和特徵，從而制定更有效的管理策略和決策。

As technology advances, image analysis techniques have gained widespread attention in agriculture. Many agricultural experts and researchers have recognized that using images can more accurately and efficiently monitor the growth of field crops, thereby increasing crop yields. However, because these images are taken from a top-down perspective, if crop density is high or the leaves are large, the lower leaves are easily overlapped, making it difficult for experts to assess the growth status of the lower leaves accurately. To address this issue, this study developed a workflow to estimate the overlap area of the leaves using a dataset containing both single and overlapping images. This workflow employs OpenCV for image preprocessing, including object edge detection, image segmentation, and object contour detection. By calculating the gradient values of leaf contours, the workflow captures the changes in leaf contour and boundary shape variations. The similarity between contour profiles is calculated for all images in the dataset using the Dynamic Time Warping (DTW) algorithm. This process identifies a contour fragment that closely matches the overlapping leaves and subsequently estimates the overlapped area. Statistical analysis methods are then applied to verify whether the estimated leaf area is similar to the true value. According to the statistical analysis results, the log-transformed Mean Absolute Percentage Error (MAPE) and Normalized Root Mean Square Error (NRMSE) values are as follows: The log mean MAPE for the original performance is 6.39%, and the log mean NRMSE is 0.10%. After removing extreme values, the log mean MAPE decreases to 4.60%, and the log mean NRMSE decreases to 0.06%. These results indicate that removing extreme values improves the accuracy of leaf area estimation.This method allows agricultural experts and farmers to understand the growth status of field plants better, thereby making more precise management decisions. We hope it can also be applied to various objects and scenarios in the future, such as environmental monitoring and natural resource management. Through image analysis techniques, we can more accurately understand the distribution and characteristics of objects in different scenarios, thereby formulating more effective management strategies and decisions.