以卷積神經網路從Sentinel-2資料測繪外來種植物之入侵分佈

Mapping distribution of invasive alien plant species from Sentinel-2 data using convolutional neural networks

具入侵性的外來種植物擴散能力強，不僅會威脅生物多樣性，也對農業造成巨大的經濟損失，銀合歡(Leucaena leucocephala (Lam.) de Wit)是世界百大外來入侵物種之一，現已嚴重威脅到臺灣恆春半島的生態系。遙測影像可觀測大範圍的地表資訊，具潛力進行銀合歡入侵分佈的測繪工作，本研究以恆春半島為研究區，整合Sentinel 2影像資料與卷積神經網路(convolutional neural network, CNN)，建立可預測銀合歡覆蓋率的深度學習模型，進而測繪銀合歡的入侵程度並估算分佈面積。研究結果顯示，運用CNN所建立的IPNet模型在準確度指標上(R2約為0.78)，比其他經過測試的深度學習模型表現更好。依據覆蓋率高低分為5種入侵程度，推估屬銀合歡入侵嚴重的面積計有4,442 ha，占全區約11%。總結而言，本研究所開發的IPNet模型是一種低成本、高效率的測繪方法，並且為銀合歡的入侵問題，提供了明確的空間分佈資訊。

Invasive alien plant species have a strong spreading ability, which not only threatens biodiversity but also causes substantial economic losses in agriculture. Leucaena leucocephala (Lam.) de Wit is one of the top 100 alien invasive species in the world, and they have been seriously threatening the ecosystem of the Hengchun Peninsula. Remote sensing imagery can provide a wide range of surface information with a potential for mapping the distribution of L. leucocephala invasion. This study focused on the Hengchun Peninsula, integrating Sentinel-2 data with a convolutional neural network (CNN) to develop a deep learning model capable of predicting the cover fraction, mapping the degree of invasion, and estimating the distribution of L. leucocephala. The results indicate that the proposed IPNet model using CNN was significantly better than the other tested deep learning models in terms of the accuracy metrics (R2≒0.78). The invasion was categorized into five degrees based on the cover fractions, with severe invasion estimated to cover an area of 4,442 hectares, accounting for approximately 11% of the entire region. In conclusion, the IPNet model developed in this study is a low-cost with high-efficiency mapping method that provides clear spatial distribution information on L. leucocephala invasion.