本文選題：高光譜遙感 + 主動(dòng)學(xué)習(xí)　； 參考：《計(jì)算機(jī)應(yīng)用》2017年06期

【摘要】：針對(duì)高光譜遙感影像分類中,傳統(tǒng)的主動(dòng)學(xué)習(xí)算法僅利用已標(biāo)簽數(shù)據(jù)訓(xùn)練樣本,大量未標(biāo)簽數(shù)據(jù)被忽視的問題,提出一種結(jié)合未標(biāo)簽信息的主動(dòng)學(xué)習(xí)算法。首先,通過K近鄰一致性原則、前后預(yù)測(cè)一致性原則和主動(dòng)學(xué)習(xí)算法信息量評(píng)估3重篩選得到預(yù)測(cè)標(biāo)簽可信度高并具備一定信息量的未標(biāo)簽樣本;然后,將其預(yù)測(cè)標(biāo)簽當(dāng)作真實(shí)標(biāo)簽加入到標(biāo)簽樣本集中;最后,訓(xùn)練得到更優(yōu)質(zhì)的分類模型。實(shí)驗(yàn)結(jié)果表明,與被動(dòng)學(xué)習(xí)算法和傳統(tǒng)的主動(dòng)學(xué)習(xí)算法相比,所提算法能夠在同等標(biāo)記的代價(jià)下獲得更高的分類精度,同時(shí)具有更好的參數(shù)敏感性。
[Abstract]:In hyperspectral remote sensing image classification, the traditional active learning algorithm only uses labeled data to train samples and a large number of untagged data are ignored. An active learning algorithm combining untagged information is proposed. First of all, through K-nearest neighbor consistency principle, prediction consistency principle and active learning algorithm information evaluation, the untagged samples with high credibility and certain amount of information are obtained. The prediction tag is added to the tag sample set as a real tag. Finally, a better classification model is obtained by training. The experimental results show that compared with the passive learning algorithm and the traditional active learning algorithm, the proposed algorithm can achieve higher classification accuracy and better parameter sensitivity at the same cost of marking.
【作者單位】： 湖北大學(xué)資源環(huán)境學(xué)院;武漢大學(xué)遙感信息工程學(xué)院;國(guó)網(wǎng)湖北省電力公司檢修公司;
【基金】：國(guó)家自然科學(xué)基金資助項(xiàng)目(41601504)~~
【分類號(hào)】：P237
，

本文編號(hào)：2105941