水稻是台灣的主要農作物之一，如果能在其收成前便取得其耕種區塊及面積，將能對水稻之用水量、產量等數值進行預估，有助於政府及早調整相關糧食策略。早期為了取得水稻耕種區域，政府會請專家以人工標註的方式，在航照圖上進行水稻田的數化作業，但此種方式十分仰賴人力資源，且判釋速度及精確度明顯不足。近年來隨著人工智慧相關技術的發展，使用深度學習技術輔助專家進行判釋，不僅能大幅提高判釋效率，也能減少人工判釋造成的誤判情況發生。

然而，適用於水稻判釋的語義分割（Semantic segmentation）方法，諸如FCN或U-net等深度學習模型，其結果皆以像素為單位，各像素皆獨立存在，極易造成椒鹽效應（Salt and pepper effect）的情況出現，導致結果難以被實際應用。通常需要進行一些後處理手段，將結果轉換成shapefile，以利在實務上方便應用。為此，本研究使用傳統區域成長法（Region Growing）輔助語義分割方法，來進行水稻坵塊的區塊化作業，能一定程度上彌補椒鹽效應的問題。但因為傳統區域成長法單一門檻值的方式不夠彈性，無法應對少數例外情況，容易發生區塊過度生長的現象。因此本研究提出使用孿生網路（Siamese network）作為區域成長法在生長時的條件依據，彌補傳統區域成長法在規則上的不足。

Rice is one of Taiwan’s main crops. If we can obtain the farming region and area before its harvest, it will be able to estimate the consumption of water and the yield of rice, which will help the government to decide related strategies as soon as possible. In the early days, in order to obtain the fields of rice, the government would ask experts to mark the rice fields on aerial photographs. However, this method relies on human resources heavily, and the speed and accuracy of interpretation are obviously insufficient. In recent years, with the development of AI related technologies, using deep learning method to assist experts to do interpretation can not only greatly improve the efficiency of interpretation, but also reduce the misjudgments caused by human.

However, Semantic segmentation methods suitable for rice interpretation, such as deep learning models such as FCN or U-net, whose results are based on pixels, and each pixel exists independently, which is very easy to cause the salt and pepper effect, making the results difficult to be applied. Therefore, some post-processing methods are needed to convert the result into a shapefile to facilitate application. For this reason, we use the traditional region growing method assist semantic segmentation method to block rice mounds, which can compensate for the problem of salt and pepper effect. However, because the single threshold method of the traditional regional growth method is not flexible enough, it cannot cope with a few exceptions. Therefore, we propose to use siamese network as a rule in region growing method to improve the traditional regional growth method.

[1] 李哲源，「區塊物件化分類模式於自動化製圖之研究-以水稻田坵塊主題圖為例」，逢甲5927學環境資訊科技研究所碩士論文，2010

[2] Lillesand, T. M. and Kiefer, R. W., 2000, Remote Sensing and Image Interpretation, 4thed., John Wiley & Sons, Inc.

[3] W.Weng andX.Zhu, “UNet: Convolutional Networks for Biomedical Image Segmentation,” IEEE Access, vol. 9, pp. 16591–16603, 2021, doi: 10.1109/ACCESS.2021.3053408

[4] J. Long, E. Shelhamer and T. Darrell, “Fully convolutional networks for semantic segmentation”, Proc. IEEE Conf. Comput. Vis. Pattern Recognition, 2015.

[5] Hojjatoleslami S. A. and Kittler J, “Region growing: A new approach”, IEEE, Image Processing, vol. 7, no. 7, pp. 1079-1084, 1998

[6] R. Adams and L. Bischof, "Seeded Region Growing", IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 16, no. 6, June 1994.

[7] .J. Bromley, I. Guyon, Y. LeCun, E. Sackinger, and R. Shah. “Signature verification using a siamese time delay neural network.” J. Cowan and G. Tesauro (eds) Advances in Neural Information Processing Systems, 1993.

[8] Gonzalez-Betancourt, M., Mayorga-Ruíz, L, “Normalized difference vegetation index for rice management in El Espinal, Colombia.”, Dyna 85 (205), 47–56, 2018

[9] Karen Simonyan, Andrew Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” International Conference on Learning Representations (ICLR), 2015.

[10] Zhang, K., Ge, X., Shen, P., Li, W., Liu, X., Cao, Q., et al. “Predicting Rice Grain Yield Based on Dynamic Changes in Vegetation Indexes during Early to Mid-Growth Stages.” Remote Sens. 11, 387. doi: 10.3390/rs11040387, 2019