近年來遙測技術進步,地表資訊被高光譜儀器大量的擷取,並透過分析光譜數值特徵獲取地表物質辨識結果;物質分析的方法以圖形識別(Pattern Recognition, PR)分類演算法進行,常見的分類演算法為最鄰近分類(k-Nearest Neighbor, k-NN)演算法,主要是以測試樣本為中心,計算最鄰近的訓練樣本數量,對測試樣本進行特徵分類。然而k-NN演算法在不同類別的訓練樣本交疊情況下容易分類錯誤,所以本論文使用最鄰近特徵向量空間(Nearest Feature Space, NFS)演算法來保留測試樣本類別群聚關係,計算測試樣本至訓練樣本特徵空間的最短距離,以提升辨識效果。 由於高光譜資料量龐大,配合NFS演算法,以一般電腦運算相當耗時,因此,本論文以實現 NFS演算法平行化為主軸,透過統一計算架構(Compute Unified Device Architecture, CUDA)實現訓練樣本為基底的平行化,將不同的訓練樣本特徵空間的運算指定到各個計算核心,並分散測試樣本至對應核心同步執行,配合記憶體資料調配來減少主機與GPU資料傳輸延遲,提升NFS演算法的運算速度。
Recently, the information of ground surface has been recode with Hyperspectral device massively and recognition the ground material by analysis the spectral data. The k-Nearest Neighbor(k-NN) algorithm is widely used in classify, the main idea of k-NN algorithm is that find the k nearest neighbor and voting by their class ID. However, the overlapping of different training sample groups will cause false classification. For overcome this problem, we trying to use Nearest Feature Space(NFS) algorithm to keep the structure of training samples and calculate the nearest distance between test sample and the feature space of training samples. Although, NFS can get a better correctness rate of classification, it will spend a huge time when the training sample is too much. For this reason, we propose a parallelism method of NFS algorithm based on training samples, distributing different feature space to corresponding core of GPUs thought Compute Unified Device Architecture(CUDA). For reduce the transform delay between Host and Device, adapting data between different memories in GPU carefully is needed.