以深度學習架構建立物種分布模型

Abstract

物種分布模型為一根據生態棲位理論的模型，該模型計算生物出現點位與環境因子相關性，以預測未知時空下棲地適合生物生存程度的模型。過去生態學家基於不同的演算法發展出多種物種分布模型，並基於背景理論及對物種出現資料的限制，物種分布模型持續在設計、評估、解讀上受到廣泛討論。隨著電腦科學的發展，包含物種分布模型等任務皆可以透過深度學習來解決，本研究利用多個深度學習模組建立全新物種分布模型DeepSDM，回應並解決過去所遇到的問題，並藉由真實世界資料與虛擬物種資料比較模型表現之優劣。 本研究以所有物種之兩兩共域次數，套用詞嵌入概念計算物種出現的關聯性作為生物因子；透過注意力模組將其與環境因子結合；以各網格資料可信度加權擬空缺資料所占損失函數的權重；選用U-Net作為模型主要架構建立多物種、多時段、高時間解析度之物種分布模型。對於真實世界鳥類資料，DeepSDM的預測結果圖呈現高度穩定性；MaxEnt的兩個訓練方式MaxEntAll及MaxEntMonths則各顯露傳統物種分布模型單一時段訓練的缺點。DeepSDM與MaxEnt模型對虛擬物種資料預測結果之AUC值在VirtualReal及VirtualIdeal情境下的相異，說明了缺乏正確空缺資料將降低模型表現指標的意義，而在已知正確且完整的出現、缺席資料條件下，較不受物種廣泛度影響的f1值顯示DeepSDM的預測結果較MaxEnt優異。 DeepSDM利用多時段的物種出現－環境因子關聯、所有物種做為模型生物因子、多空間尺度的環境因子，建立涵蓋更多特徵維度的生態棲位超體積，並且以不同過往研究的角度加權擬空缺網格的權重，期以模擬複雜的實際狀況。結果證實DeepSDM比MaxEnt更具備回答物種分布問題的能力，未來該模型將可應用於多種大尺度物種分布任務中，成為最有力的工具之一。

Species distribution models are models based on niche theory that calculates the correlation between the species occurrence and environmental factors to predict the suitability of habitats in unknown time and place. In the past, ecologists have developed various species distribution models based on different algorithms, and these models have been widely discussed in terms of design, evaluation, and interpretation, taking into account background theory and limitations of species occurrence data. With the development of computer science, tasks including species distribution modeling can be solved by deep learning. In this study, we have developed a novel species distribution model called DeepSDM using multiple deep learning modules to address and solve past issues about species distribution models. We have compared the performance of the model using real world data and virtual species data. We used the pairwise co-occurrence frequency of all species and applied the concept of word embeddings to calculate the correlation of species occurrences as the biological factor. We combined it with environmental factors using attention modules and weighted the missing data in each grid based on the credibility of the grid data in the loss function. We used U-Net as the main architecture of the model to establish a multi-species, multi-temporal, and high temporal resolution species distribution model. For real world bird data, DeepSDM predicted results showed high stability, while the two different training methods of MaxEnt, MaxEntAll, and MaxEntMonths each exhibited the limitations of traditional species distribution models trained in a single time period. The AUC of DeepSDM and MaxEnt models for predicting virtual species data differed in the VirtualReal and VirtualIdeal scenarios, indicating that the lack of accurate absence data reduces the significance of model performance indicators. However, under the condition of known and complete presence and absence data, the f1 values, which are less influenced by species prevalence, demonstrated the superior predictive results of DeepSDM over MaxEnt. DeepSDM utilizes the correlations between species occurrences and environmental factors across multiple time periods, uses all species as biological factors, and incorporates environmental factors at multiple spatial scales to establish an ecological niche hypervolume that covers more feature dimensions. Additionally, the weights of the missing grid data are assigned based on different perspectives from previous studies to simulate complex real world situations. The results confirm that DeepSDM has a better ability to answer species distribution questions compared to MaxEnt. In the future, this model can be applied to various large-scale species distribution tasks , making it one of the most powerful tools in the field.