藉由土地利用分析可以了解土地的使用情況及輔助碳權之計算與評估，土地利用現況調查需要付出的時間成本與人力成本不在少數。本篇論文將台灣本島地區之 SPOT-7 輻射校正 (radiometric correction) 衛星影像進行切塊，並對各區塊進行人工標註，以此收集十一種類別標註資料，十一種類別為道路、建築、水體、一般裸露地、農耕裸露地、草原與草地、農作物、針葉林、闊葉林、混淆林、其餘植被。將 ViT-B/16 模型進行修改後，訓練一個分類模型，透過地理物件式影像分析 (GEOBIA) 將區塊化影像資料之直方圖 (histogram image) 作為資料輸入端，以此訓練基於自注意力機制 (self-attentioin) 的分類器。本論文使用 2021 年與 2022 年台灣本島地區之 SPOT-7 輻射校正衛星影像，訓練出兩個模型來進行各自的土地利用分類預測及以 2021 年之模型預測 2022 年之土地利用分類，針對上述土地利用之變化進行研究與分析。;Through land use analysis, we can understand land use and assist in calculating and evaluating carbon rights. Surveying the current land use status requires a lot of time and labor costs. This paper cuts the SPOT-7 satellite image of radiometric correction of Taiwan′s main island into blocks and manually labels each block to collect eleven categories of labeled data. The eleven categories are roads, buildings, water bodies, general bare land, agricultural bare land, grasslands and grasslands, crops, coniferous forests, broadleaf forests, mixed forests, and other vegetation. After modifying the ViT-B/16 model, a classification model is trained, and the histogram image of the block image data is used as the data input end through Geographic Object-Based Image Analysis (GEOBIA) so that the training is based on a self-attention classifier for mechanisms. This paper uses the SPOT-7 satellite images of radiometric correction of Taiwan′s main island in 2021 and 2022 to train two models to predict land use classification and the 2021 model to predict the land use classification in 2022. For the above-mentioned land use the changes to conduct research and analysis.
