To screen the key indicators that affect the robustness of tomato plug seedlings during the spring nursery and transplanting period and realize its rapid non-destructive testing, this paper measured 5 seedling indicators, then used vector normalization and the independence weight coefficient method to determine each indicator. According to the weighting results, two indicators containing more comprehensive information and greater influence are selected: chlorophyll and dry quality. The simplified seedling evaluation value composed of the two indicators can approximate the comprehensive evaluation value. The correlation coefficient is 0.92, which greatly reduces the number of indicators required for quality testing, and can well represent the robustness of tomato seedlings during the spring seedling transplanting period. At the same time, the visible-near infrared spectrum data of each plug seedling is extracted and pre-processed by denoising and multi-scattering correction (MSC). This way, it can eliminate the spectral interference information caused by light scattering and make it more usable than the original spectral information. Subsequently, the spectrum-physical and chemical value symbiosis distance (SPXY) algorithm is used to divide the sample set. The distance between the samples is calculated using two variables of the band value and the evaluation value to maximize the characterization of the sample distribution to improve the difference and representativeness of the samples. Secondly, the competitive adaptive weighting algorithm (CARS) and the uninformative variable elimination-successive projections algorithm (UVE-SPA) are used to optimize the spectral feature wave number, reduce the spectral data dimension and obtain simplified spectral information that can better reflect the spectral characteristics and reduce redundancy. Finally, partial least squares-support vector machine (LS-SVM) and convolutional neural network (CNN) based on U-Net model transformation are applied. After extracting the characteristic wavelength, the preprocessed spectral data and the spectral data are respectively used as the input of the model and established a non-linear mapping model of spectral data and comprehensive evaluation value. We can carry out comparison and selection. The results show that the spectral information of the bands filtered by the UVE-SPA preprocessing method is more abundant and effective. The regression effects of the models built for the two preprocessed optimal bands are overall better than the models built for the full bands; the modeling effect of the CNN model is overall. It is better than the LS-SVM model, and the UVE-SPA-CNN model has the best effect on the regression analysis of spectral data and seedling evaluation values. The correlation coefficients of the modeling set and prediction set are 0.988 and 0.946, respectively, and the values of the root mean square error are 0.085 and 0.025, respectively, which provide a theoretical basis for directly using spectral data to obtain the evaluation value of tomato seedlings that incorporate multiple factors, thereby judging the robustness of the seedlings.