Using Machine Learning Methods Combined with Vegetation Indices and Growth Indicators to Predict Seed Yield of Bromus inermis

Smooth bromegrass (Bromus inermis) is a perennial, high-quality forage grass. However, its seed yield is influenced by agronomic practices, climatic conditions, and the growing year. The rapid and effective prediction of seed yield can assist growers in making informed production decisions and reducing agricultural risks. Our field trial design followed a completely randomized block design with four blocks and three nitrogen levels (0, 100, and 200 kg·N·ha-1) during 2022 and 2023. Data on the remote vegetation index (RVI), the normalized difference vegetation index (NDVI), the leaf nitrogen content (LNC), and the leaf area index (LAI) were collected at heading, anthesis, and milk stages. Multiple linear regression (MLR), support vector machine (SVM), and random forest (RF) regression models were utilized to predict seed yield. In 2022, the results indicated that nitrogen application provided a sufficiently large range of variation of seed yield (ranging from 45.79 to 379.45 kg ha⁻¹). Correlation analysis showed that the indices of the RVI, the NDVI, the LNC, and the LAI in 2022 presented significant positive correlation with seed yield, and the highest correlation coefficient was observed at the heading stage. The data from 2022 were utilized to formulate a predictive model for seed yield. The results suggested that utilizing data from the heading stage produced the best prediction performance. SVM and RF outperformed MLR in prediction, with RF demonstrating the highest performance (R2 = 0.75, RMSE = 51.93 kg ha-1, MAE = 29.43 kg ha-1, and MAPE = 0.17). Notably, the accuracy of predicting seed yield for the year 2023 using this model had decreased. Feature importance analysis of the RF model revealed that LNC was a crucial indicator for predicting smooth bromegrass seed yield. Further studies with an expanded dataset and integration of weather data are needed to improve the accuracy and generalizability of the model and adaptability for the growing year.