深度學習應用於航照影像果園坵塊判釋

Orchard Classification and Parcel Segmentation in Aerial Imagery Using Deep Neural Networks

陳彥碩(Yen-Shuo Chen)

指導教授 ： 周呈霙

國立臺灣大學/生物資源暨農學院/生物機電工程學系/碩士(2022年)

https://doi.org/10.6342/NTU202202109

摘要

農作物的產量過剩或產量不足會影響至供需不平衡，進而產生價格波動，因此，使農民或農業相關單位能夠監控農地進而推估產量或控制品質是個重要的課題。然而，目前監控農地的方法為人估監控，這是一種成本高、耗時、費力、且易出現錯誤的方法，故本篇提出自動化的監控系統，此系統能自動分割每塊農地，此系統分成三個部分：轉換地理資訊成訓練模型的資料集、使用深度捲積神經網路模型、和將模型嵌入至常用的地理資訊系統ArcGIS Pro。此篇使用三種神經網路模型： Mask R-CNN、PointRend、和Yolact，期平均分割準確度分別為0.585、0.592、和 0.528，且使用我的模型來判釋一張航照的時間低於1分鐘。此篇的方法能有效率的分割航照圖中的坵塊面積，且可在ArcGIS Pro中使用本篇訓練好的模型，使農業相關產業或使用地理資訊系統的研究者有更有效率的監控面積的方法。

關鍵字

深度學習 ； 果園判釋 ； 航照圖 ； 遙測 ； 坵塊分割

並列摘要

A surplus or shortage of agricultural fruits often leads to a severe imbalance between supply and demand. Therefore, farmers or agricultural entities monitoring orchards in different geographic areas to predict yield and quality was significant. However, manual monitoring is costly, time-consuming, and unstable. My study proposed a method to automatically identify orchards within a geographic area. The method consisted of three parts. Initially, the geographical information data was changed to COCO data format. The rectified aerial images were fed into a deep convolutional neural network (DCNN). The best segmentation mean precision (mAP) of Mask R-CNN, PointRend and Yolact were 0.585, 0.606, and 0.528, respectively. Then I implemented my well-trained models to the ArcGIS Pro. The processing time for an aerial image with my models was lower than 1 minute. As a result, the model predicted the acreage. Open regions are efficiently classified and segmented. The method could help the experts and reduce their loading of the works.

並列關鍵字

Deep learning ； Orchard classification ； Aerial imagery ； Remote sensing ； Parcel segmentation

參考文獻

Badrinarayanan, V., Kendall, A., Cipolla, R. (2017). Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 39(12), 2481-2495.

Bianco, S., Cusano, C., Napoletano, P., Schettini, R. (2017). Improving CNN-based texture classification by color balancing. Journal of Imaging, 3(3), 33.

Bolya, D., Zhou, C., Xiao, F., Lee, Y. J. (2019). Yolact: Real-time instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 9157-9166).

Carlson, T. N., Ripley, D. A. (1997). On the relation between NDVI, fractional vegetation cover, and leaf area index. Remote sensing of Environment, 62(3), 241-252.

Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., ... Lin, D. (2019). MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155.