比較紋理特徵及多時段光譜特徵對分類正確率之改善

A Comparative Study of Texture and multi-temporal Spectrum Features in Classification Accuracy

江介倫(Jie-Lun Chiang) ； 張世駿(Shih-Chun Chang) ； 蘇元風(Ke-Sheng Cheng) ； 鄭克聲(Jen-Chen Fan) ； 范正成

《中華水土保持學報》 41卷1期 (2010/03) Pp. 1-25

https://doi.org/10.29417/JCSWC.201003_41(1).0001

摘要

衛星影像已廣泛應用於環境資源與自然災害之調查與監測，衛星影像分類為其主要應用之一。當欲區分之類別較詳細複雜時，分類結果之正確率不易提升，因此分類正確率的提升一直是重要研究方向。本研究利用多時段之衛星影像提供更多的資訊，增加土地利用分類之正確率。以石門水庫上游集水區為例，因為不同植生對於土壤的保護程度不同，故將土地利用細分為針葉林、竹林、闊葉林、稻田、農田、建物、裸露地、水體等八類，研究結果發現以SPOT5影像之空間解析度，二階角動差(Angular Second Moment, ASM)、對比(Contrast, CON)、以及熵(Entropy, ENT)等三紋理特徵(feature)並不能有效提升分類正確率；而多時段光譜特徵則可使總體正確率提升。以最大概似法分類時，可從使用單時期光譜特徵的81.97%以下，提升到用三時期光譜特徵時的92.14%；而此時Kappa值亦由77.02%以下提升到90.29%；再以指標克利金法對三時期光譜特徵分類時更可以達到95.76%總體正確率，及94.56%的Kappa值，故知多時期的光譜資訊可有效提升分類之正確率。

關鍵字

地表覆蓋 ； 分類 ； 遙感探測

並列摘要

The satellite images have been widely applied on the investigation and monitor of natural environment and hazard. Satellite images classification is one of the major applications. However, the accuracy of classification still requires improvement. Especially when the category is more detailed, the accuracy of classification is more difficult to improve. In this study, multi-temporal satellite images were used to afford more information in order to improve the classification accuracy. From the result, in the upstream of Shihmen Reservoir, we classified the land cover into eight categories: vegetation, bare soil/built-up, farm, water, paddy field, conifer, broadleaf tree, and bamboo. We found using texture features in this case did not improve classification accuracy but using multi-temporal spectrum features did improve classification accuracy. The classification accuracy is improved from 81.97% (using the single period images) to 92.41% (using the multi-temporal period images). When the classification features are increased from single period images to multi-temporal satellite images, Kappa is improve from 77.02% to 90.29%. When using IK classifier classification accuracy is 95.76% and Kappa is 95.56%. As a result, the multi-temporal images effectively improve the classification accuracy.

並列關鍵字

landcover ； classification ； remote sensing