Title: 類神經網路於多光譜影像分類之應用
The Application of Neural Network for the Image Classification of Multispectral Data
Authors: 邵泰璋
Tai-Chang Shao
Tian-Yuan Shih
Keywords: 遙感探測;類神經網路;影像分類;Remote Sensing;Neural Network;Image Classification
Issue Date: 1998
Abstract: 水稻是台灣地區重要的經濟作物,相關的問題一直是政府單位關切的重點。台灣省糧食處每年採用航照遙測技術調查稻作,並以人工辨識航照方法,估算水稻田面積與產量;若能發展適當的自動分類辨識方法,將可減少時間、人力與物力之投入,且避免人為判釋上的主觀差異。 本研究採用類神經網路,其模仿人類神經元記憶思考的處理模式與容錯性的特點,適合分類工具的發展。研究中,選用監督式理論較具代表的倒傳遞類神經網路,與混合監督式與非監督式概念的學習向量量化類神經網路,並採用兩種不同資料編碼輸入網路模式,分別針對彰化地區多時段SPOT衛星影像與多時段正規化差分植生指數影像作水稻田分類工作。分類成果與傳統高斯最大似然分類法相比較,最後並加入紋理影像輔助分類,探討其對分類工作上的助益。 研究結果顯示,在使用多時段光譜影像時,粗編碼之倒傳遞類神經網路展現出最佳的分類效果;而使用多時段正規化差分植生指數影像時,則以一般正規化編碼之倒傳遞類神經網路表現較佳。就整體而言,類神經網路確實比傳統高斯最大似然分類法為佳,尤其以倒傳遞類神經網路最為有效,學習向量量化類神經網路次之。另外選擇三種紋理影像輔助分類,但從成果顯示,僅二階角動量紋理輔助多時光譜影像分類時,有略微提高分類精度外,其餘紋理影像的加入並未能提升分類精度。
Rice is the most important economical crop in Taiwan, and related problems are always the major concern of the government. Taiwan Provincial Food Department utilizes aerial photo interpretation for rice crop inventory each year to calculate the rice areas and crop field. If an automated classification method can be developed, the amount of time, manpower, and resources needed in current work can be reduced. Meanwhile, the errors caused by human subjective interpretation can be avoided. This research uses artificial neural network (ANN) which simulates human neuron and fault-tolerance for classification. In this study, error back-propagation (BP) and learning vector quantization (LVQ) neural network algorithms are selected. Meanwhile, two data coding techniques are applied for data represatation to input network model. The data used in the experiment are multi-temporal SPOT images and multi-temporal NDVI images of Changhua area. All the classification results are compared with those produced by gaussian maximum likelihood algorithm. Finally, the contribution of texture images for classification are studied. The experiments reveal that BP with coarse coding provides the best classification accuracy for multi-temporal spectral images, and BP with normal normalized coding provides the best outcome for multi-temporal NDVI images. In general, neural network approaches are better than maximum likelihood classification. Especially BP, and LVQ is the second-best. Regarding texture images, besides ASM image, the other texture images did not improve the classification accuracy.
Appears in Collections:Thesis