- 期刊
影像參數加權統計法PN指標於土地利用型高光譜特徵選粹上之研究
A Statistical Based Hybrid PN Criterion for Band Selection of Airborne Hyperspectral Data for Land Covers Classification
摘要
高光譜資料所記錄的各波段光譜波長範圍遠比多光譜資料窄小，資料維度約數十至數百波段，資料量非常龐大，且各波段資料所含光譜資訊同質性偏高，如何慎選具有資訊代表性的少數光譜波段，以達到環境與資源監測目的，是一個非常重要的課題。本文利用1997年農業委員會高光譜遙測計畫所取得DAIS-3715高光譜資料，分析濱海地區的15種土地利用型，藉以探討高光譜資料的波段選擇問題。研究結果顯示，利用高光譜資料各波段影像的均值與變異數等參數，發展出的影像參數加權統計法－PN指標具有很高的效能，可以精確的篩選DAIS-3715高光譜影像資料特性，使用近紅外光區的0.96、0.84、0.82μm以及可見光區的0.59μm等四個波段的光譜資訊，即可使訓練樣區和評估樣區之全區分類準確度及Kappa同意度係數高於95%及90%，分類使用的電腦時間可以降低到1/7.5倍的時間。PN指標選定的四個波段對訓練樣區的分類準確度約比全部24個波段的分類準確度減少3%，但其於評估樣區的準確度卻可高出約3%。
並列摘要
Hyperspectral data have tens to hundreds bands and finer or better spectral, spatial, and radiometric resolution than multispectral data and are hence widely used in land covers researches in recent years. Reducing the dimensionality and keeping the information of hyperspectral data become an important works in applying such great volume and spectral homogeneously information of remote sensing data. This paper uses the DAIS-3715 data, a product of hyperspectral remote sensing program conducted by Committee of Agriculture of ROC in 1997, to classify 15 types of land-cover of seashore area of southern Taiwan. It aims to study the efficiency of proposed statistical based hybrid criterion, PN, in reducing the data dimensionality or selecting heterogeneously spectral bands for seashore land-cover classification. Results showed that PN criterion is excellent in selecting heterogeneous spectral bands of DAIS-3715 data for classifying the seashore land-cover. PN method can choose effectively only four bands, denoted as band no. 24, 18, 17, and 5 which correspondingly records the near infrared signals of 0.96, 0.84, 0.82 micrometer and visible signal of 0.59 micrometer, to achieve a good accuracy and which is almost identical to the one of the classification by using all of 24 bands. PN method saved about 7.5 times of the computer time of land uses classification. The accuracy assessments for all of the classification with different band combinations that decided by each band selection criterion are assessed from the same sets of training and assessing data. Overall accuracy (OA) and kappa_hat accuracy (k) of the classification using only four bands selected by PN criterion for both training and assessing data are 95% and 90% respectively. Although, the accuracy of the PN method based classification is less about 3% than the ones of all 24 bands based classification for the training data while it is more about 3% for the assessing data.