Obtaining the spectral reflectance of an object is the key to accurately reproducing an objects true color under various lighting conditions, which plays an important role in industries with high color requirements, such as textiles and clothing, publishing and printing, online e-commerce, telemedicine, etc. The purpose of spectral reflectance reconstruction is to use training samples to establish the mapping relationship between RGB trichromatic values and high-dimensional vector of spectral reflectance obtained by general equipment such as digital cameras to avoid the problems of high cost, complex operation and low resolution caused by the use of a spectrophotometer and other professional equipment. Due to the limitation of uneven or inconsistent training sample distribution, the selection of training sample sets greatly impacts the spectral reflectance reconstruction processes. The representative color samples selection for local learning-based spectral reflectance reconstruction are discussed in this paper. From a physical point of view, the spectral reflectance vector is a smooth curve, and the selection of training samples should consider both the spatial distance and the similarity of the shape. A method based on improved weighted Euclidean distance is proposed for sample selection. The weighted Euclidean distance between the testing sample and the training sample is combined with the vector angle distance, and different weights are given as the similarity measure, which aims to ensure the similarity between training samples and target samples. The experimental results show that the proposed method can significantly reduce the chromaticity error while ensuring the minimum root mean square error. Moreover, after adding noise, it maintains the minimum root mean square error and chromaticity error, showing the method has good generalization performance.