With the delay of rice maturity in rice-wheat rotation areas in the middle and lower reaches of the Yangtze River, the delay of the sowing date of winter wheat has become the main obstacle affecting the yield, so it is necessary to screen better resistant varieties in late sowing wheat. This study was designed to monitor the relative chlorophyll content of canopy leaves during the early winter wheat growth for late-sowing winter wheat variety screening. In order to explore the feasibility of monitoring chlorophyll content in winter wheat, this study used five single-band spectral reflectance and 15 vegetation indices obtained by UAV as the independent variables. Through recursive feature elimination (RFE) feature variables screening, redundant variables were removed. A remote sensing inversion model of winter wheat’s relative chlorophyll content (SPAD) was established using the BP neural network regression algorithm. Based on the measured leaf SPAD values of winter wheat in the experimental site of Guangling District, Yangzhou city, Jiangsu Province, during 2020—2021, the correlation between remote sensing variables and SPAD values in the two growth stages was analyzed combined with multi-spectral UAV images obtained simultaneously. In addition, feature variables were screened based on the ranking of feature importance among remote sensing variables, and the selected variables were used as the input of the model to construct and screen out the best inversion model for each growth period. Using Ridge regression (Ridge) and Gradient Boosting Decision Tree (GBD) algorithms as a comparison, and R2 and RMSE as model evaluation indexes, the three models’ self-learning ability and generalization ability were analyzed on the validation set. The results showed that the BP neural network model based on optimal spectral information screening showed the strongest regression prediction ability in the two growth periods. R2 and RMSE were 0.806 and 1.861 in the overwintering stage and 0.827 and 0.507 in the jointing stage, respectively. In this paper, the variable selection of UAV multi-spectral data was carried out, and the BP neural network of optimization model constructed had high estimation accuracy. It showed that the effect of early monitoring of winter wheat was better in the elongation stage than in the overwintering stage. It is valuable to use UAV multi-spectrum to estimate the SPAD value of late-sowing winter wheat for variety resistance screening.