研究土地利用有助於規劃和管理土地資源，傳統方式進行國土利用現況調查需要花費大量人力與時間成本來維護。本篇論文以區塊為單位，對台灣本島地區之 SPOT-7 衛星影像進行人工標註，收集台灣本島地區之建物道路、樹林、草地草原、農作物、水體、一般裸露地、農地裸露地七種類別的標註資料集。並應用地理物件影像分析 (GEOBIA) 的方法，將區塊化資料之影像直方圖(histogram image)作為資料輸入修改過之 ViT-B/16 模型，訓練一個基於自注意力機制 (self-attentioin) 的分類模型。本論文使用 2013 與 2021 兩年台灣本島地區的 SPOT-7 衛星影像，分別訓練兩個模型來預測各自的土地使用分類，並針對兩年的土地使用變遷進行研究與分析。;Studying land use contributes to the planning and management of land resources. Traditional methods of conducting land use surveys require significant human and time resources for maintenance. In this paper, using the block as the unit, we manually labeled SPOT-7 satellite imagery of Taiwan, collecting labeled datasets for seven categories: buildings/roads, forests, grasslands, crops, water bodies, general bare land and agricultural bare land. By applying the method of Geographic Object-Based Image Analysis (GEOBIA), we used the histogram distribution of the block-level data as input to the modified ViT-B/16 model, which is based on self-attention mechanism, to train a classification model. Two models were trained using SPOT-7 satellite imagery of Taiwan for the years 2013 and 2021, respectively, to predict land use classifications for each year. The land use changes between the two years were studied and analyzed.
