透過您的圖書館登入
IP:18.216.94.152
  • 期刊

空間轉換和取樣方法對於森林分類準確度影響之研究

Spatial Transformation and Sampling Strategy on Computer-Assisted Classification in Forest Thematic Mapping

若您是本文的作者,可授權文章由華藝線上圖書館中協助推廣。

摘要


遙測技術應用於森林分類所面臨的共同問題是分類準確度還低於分析者所要求的期望值。爲求提高分類準確度,本文以空載多譜掃描影像爲資料,探討造成遙測影像分類準確度低的原因,並提出空間轉換和不同取樣方法(逢機與系統取樣)以消除或減低造成分類準確度低的原因。同時,檢定所提出的方法是否能改善分類的結果。若能改善,則將之應用於日後遙測資料之電腦分類處理,並提高遙測技術與森林資源資訊系統整合應用的可行性。 試驗結果指出:根據傳統監察方法之區集取樣法,利用最大概似分類式和線性判別分類式所得的結果分別爲80.7%和97.3%。而本文所提出的空間轉換和不同取樣方法所得的結果分別爲: 空間轉換方法:最大概似分類式爲81.8%,而線性判別分類式爲97.6%。 系統取樣方法:最大概似分類式爲90.9%,而線性分別分樣式爲98.0%。 逢機取樣方法:最大概似分類式爲92.2%,而線性判別分樣式爲96.7%。 由上述分類所得結果可知,在應用監察分類方法時,本文所得出的空間轉換和取樣方法確能改善分類的結果,而改善程度的多寡,以逢機取樣法爲最佳,系統取樣法其次,空間空間轉換爲第三,因此,本文建議以後有關森林遙測資料的分類處理時,若使用監察分類法,可採用逢機或系統取樣方法以進行選取具低空間機關的訓練樣本,達到提高分類準確度的目的。

並列摘要


The effects of spatial transformation and sampling strategy on spectral classification using airborne MSS data were investigated. The objective was to investigate whether spatial transformation and sampling strategy can eliminate the effect of spatial correlation and improve the classification accuracy in order to in crease the possibility of integrating remote sensing and geographic information system. The classification accuracies obtained from the conventional block selection and the proposed spatial transformation and sampling strategy were as follows. 1. Block selection: the maximum likelihood classifier is 80.7% and the linear discriminant classifier is 97.3%. 2. Spatial transformation: the maximum likelihood classifier is 81.8% and the linear discriminant classifier is 97.6%. 3. Systematic sampling strategy: the maximum likelihood classifier is 90.9% and the linear discriminant classifier is 98.0%. 4. Random sampling strategy: the maximum likelihood classifier is 92.2% and the linear discriminant classifier is 96.7%. For the above classification results, it is known that the proposed spatial transformation and sampling strategy did improve the supervised classification accuracy. And the priority of classification accuracy is random sampling, systematic sampling, and spatial transformation, respectively. Therefore, using random sampling and systematic sampling strategies in the selection of training classes is suggested to increase the accuracy of supervised classification.

延伸閱讀