Timely, comprehensive and accurate estimation of banana yield can provide growers with decisions on variable fertilization, irrigation, harvest planning, marketing and forward sales. To improve the accuracy of banana remote sensing yield estimation, totally 71 banana fields in Fusui County, Guangxi were used as the study area, and a remote sensing prediction model for banana yield in 2019—2020 was conducted by using time-series Sentinel-2 remote sensing image data, combined with field measured yield data. The method firstly obtained Sentinel-2 images during the key banana phenological period of 2019—2020, then the threshold segmentation and morphological open operation methods were used to remove cloud and cloud shadow coverage areas, the average normalized difference vegetation index (NDVI) values of each plot were extracted, and finally the BSO-SVR model was used to predict and evaluate the banana yield in combination with the actual measured data of banana yield. The results showed that compared with the grid search (GS) and grey wolf optimizer (GWO) algorithms to optimize the penalty factor and kernel function parameters of the SVR model, the brain storming optimization algorithm proposed had higher prediction accuracy and faster prediction speed. The running times of the BSO-SVR model in 2019 and 2020 were 0.320s and 0.331s, respectively, and for the validation set, the R 2 of the BSO-SVR model was 0.777 and 0.793 in 2019 and 2020, respectively; for the test set, the R 2 of the BSO-SVR model was 0.765 and 0.636 in 2019 and 2020, respectively, except that the R 2 of the BSO-SVR model in 2019 is slightly lower than that of the GS-SVR model (R 2=0.797) in 2019, except that the R 2 of the BSO-SVR model was higher than that of the GWO-SVR model and the GS-SVR model, and in addition, the overall performance of the RMSE and MAE of the BSO-SVR model was optimal in 2019—2020 compared with that of the GWO-SVR model and the GS-SVR model, indicating that the prediction results of the BSO-SVR model were closer to the actual values and with higher forecasting accuracy. Compared with the traditional ridge regression (RR) and partial least squares regression (PLSR) models, in 2019, the BSO-SVR model had the highest R 2, followed by the RR model, and the PLSR model was the worst, where the BSO-SVR model had R 2 above 0.75 for both the validation and test sets, which was 0.113 and 0.174 higher than that of the RR model, and 0.192 and 0.184 higher than thta of the PLSR model, respectively. Meanwhile, the BSO-SVR model had the lowest RMSE and MAE compared with the RR model and PLSR model, indicating that the BSO-SVR model had good results in forecasting banana yield in 2019. In 2020, the BSO-SVR model had the best overall performance, with the average R 2 of 0.715 for the validation and test sets, and the R 2 of the validation and test sets were higher than that of the RR model by 0.035 and 0.014, respectively, and better than that of the PLSR model by 0.040 and 0.035, while the RMSE and MAE of the BSO-SVR model also had the best overall performance. The banana time-series yield estimation model proposed achieved accurate yield prediction of banana field plots, which can provide an effective way for field-scale crop yield estimation.