抽樣方式及抽樣方法對於土地利用變遷模式之影響—以桃園地區為例

Abstract

土地利用變遷的動態過程中具有多重複雜性與不確定性。在眾多領域中皆為重要的研究議題，如:地理學、生態學及社會科學領域。隨著資料科學的進步，土地利用研究目前可處理的網格資料精度日趨精細，亦使得演算時間大幅增加。然而由於土地利用資料相較於一般二維資料具有空間自相關性，容易使得分類模型產生偏誤以至於無法準確代表該研究區的土地利用分布情形。因此如何有效控制資料間的空間自相關性以降低模擬及預測模型中的偏誤(bias)是建立模型中的首要課題。 本研究以桃園地區為研究區比較系統抽樣(systematic sampling)、隨機抽樣(random sampling)與分層隨機抽樣(stratified random sampling,)在不同抽樣比例下應用邏輯斯迴歸(logistic regression)、廣義可加性模型(general Additive Model)、隨機森林(random forest)三種演算法對於CLUE-s模式的土地利用驗證結果之影響。推論結合各土地利用類別之最佳演算法的混合演算模式對於CLUE-s模式是否有過度擬合(overfitted)的問題。此外，本研究亦針對僅使用兩時期土地利用地圖之情形提出可完整校正及驗證模型之方法。 研究結果顯示，抽樣方式差異對於CLUE-s模式之模擬結果並無顯著影響，然而抽樣比例差異對於模型驗證則有顯著影響，且在3種演算法中皆以100%抽樣模型為最佳；若採取90%樣本於邏輯斯迴歸，60%於廣義可加性模型或是90%於隨機森林之CLUE-s模式則為最具效率之抽樣比例。結合各土地利用AUC值最佳演算法之混合模型並無法提升CLUE-s模式土地利用配置精確度，反而影響其他土地利用配置產生誤差。此外，本研究之結果可適用僅有1~2時期之土地利用圖資情形，先以網格抽樣方式切割時期1資料集建立土地利用變遷模型，再以時期2圖資驗證，以確保模型預測結果與實際利用情形相符。

The process of land use change contains multiple complexities and uncertainties. Therefore, it is viewed as an important issue for global environmental change. As data science has advanced, the resolution of raster data has become higher, but this also makes the workload heavier. Additionally, land use data have spatial autocorrelation, which can cause bias in classification algorithms and prevent the model from representing the real land use. Thus, it is crucial to efficiently control the autocorrelation in land use and land change research. This study used Taoyuan as the research area to compare the differences between three sampling methods (systematic sampling, random sampling, and stratified random sampling) in different sampling ratios (100%, 90%, 80%, 70%, 60%, and 50%) with three classification algorithms (logistic regression, general additive model, and random forest) in CLUE-s model performance. Furthermore, it explored the overfitting problem in the mixed model which combined the best suitability of each land use type. Additionally, the research proposed a method that can completely calibrate and validate the model for the situation where only two periods of land use maps are used. First, the results show that there is no significant difference between the sampling methods. It shows a significant difference in sampling ratios, with the 100% sampling model having the highest accuracy in all three algorithms. However, the 90% sample rate in logistic regression, the 60% sample rate in GAM, and the 90% sample rate in random forest are the most efficient modeling methods in the Taoyuan area. Second, the mixed model could not improve the accuracy of the CLUE-s model; instead, the mixed model reduces the allocation of other land use types. Lastly, this research shows a stable and solid method of validating a LULC model by separating the period 1 land use data into a training dataset and a testing dataset with spatial sampling methods and validating the LULC model with the period 2 land use data. The procedure can ensure that the LULC model conforms to the real land use situation. The results of this research could provide a viewpoint on the effect of grid sampling in land-use modeling and land-use land-change research.