Combined Use of Landsat 8 and Sentinel 2A Imagery for Improved Sugarcane Yield Estimation in Wonji-Shoa, Ethiopia

In this study, a support vector regression (SVR) approach based on a radial basis function was used for estimating sugarcane yield in the Wonji-Shoa sugarcane plantation (Ethiopia) combining Landsat 8 (L8) and sentinel 2A (S2A) data. Vegetation Indices(VIs) involving visible, near-infrared, and shortwave infrared bands were calculated from the L8 and S2A sensor observations, and seasonal cumulative values were computed for the period June to October in the 9th month and June to November in the 10th month of the year for 2016/17 to 2018/19 cropping seasons. Sugarcane yield was predicted using the SVR, Multilayer perceptron neural network (MLPNN), and Multiple linear regression (MLR) methods. Then, a tenfold cross-validation approach was implemented for the performance evaluation. The results showed significant correlations between sugarcane yield and cumulative values of VIs computed during the 10th month in the growing season. The results also revealed that the estimation accuracy of sugarcane was better using the combined L8 and S2A (RMSE = 12.95 t/ha, and MAE = 10.14 t/ha) than using the S2A data alone (RMSE = 14.71 t/ha, and MAE = 12.18 t/ha). Comparing SVR results with MLPNN and MLR disclosed that SVR outperforms the other two models in terms of prediction accuracy. Overall, this study demonstrated the successful application of the SVR in developing a model for Sugarcane yield estimation and it may provide a guideline for improving the estimations of sugarcane in the study area.