CAJ | 학술논문

针对遥感卫星图像的云检测，提出了基于最小化支持向量数分类器的云检测方案，解决传统分类器训练样本多、易陷入局部最优的问题。使用该分类器对QuickBird高分辨率遥感图像进行云检测，检测正确率达99%以上。实验表明：在确定分类器内部结构参数过程中，与传统的交叉验证法相比，基于支持向量数的方法不仅能够准确预测分类器推广性能的变化趋势，从而确立最优化的参数组合，并且实现简单，大大减少了计算的复杂度。与传统的BP神经网络相比，该方法所需训练样本少，分类性能好。
침대요감위성도상적운검측，제출료기우최소화지지향량수분류기적운검측방안，해결전통분류기훈련양본다、역함입국부최우적문제。사용해분류기대QuickBird고분변솔요감도상진행운검측，검측정학솔체99%이상。실험표명：재학정분류기내부결구삼수과정중，여전통적교차험증법상비，기우지지향량수적방법불부능구준학예측분류기추엄성능적변화추세，종이학립최우화적삼수조합，병차실현간단，대대감소료계산적복잡도。여전통적BP신경망락상비，해방법소수훈련양본소，분류성능호。
The classifier plays an important role for cloud detection in remote sensing image. Traditional Classifiers demand excessive training samples and have risks to fall into local optimum. To solve these deficiencies, SVM was presented as the classifier to achieve cloud detection based on SVD as feature vectors. Meanwhile, the method of minimizing the support vector count was introduced to substitute cross- validation method for optimal parameters selection. Experiment over high resolution remote sensing images QuickBird showed, with this method, the correction rate of cloud detection could be higher than 99%. It also suggested support vector count could reflect the classifier’s estimation accuracy and was more easy to compute. The SVM classifier established in this way, compared with BP neural network, needed fewer training samples but achieved higher accuracy, it showed better performance in cloud detection field.