On the Selection of Features for the Prediction Model of Cultivation of Sweet Potatoes at Early Growth Stage

With the changes in the agricultural environment, the cultivation of sweet potatoes today cannot be managed extensively as in the early days. The increase in production costs makes it necessary for enterprises that manage agricultural lands to estimate the premium grade yield more accurately, reduce the risk and cost of supply shortages, and even further improve premium grade yield and maximize the value of agricultural properties. In the past, the forecast of premium grade yield used manual inspection and field sampling to judge the growth status of sweet potato and infer the final output of premium-grade sweet potatoes, but such an estimation method requires long-term tracking and could not do early supply planning. Therefore, this study aims to develop a prediction model for sweet potato production, using cultivation information provided by a sweet potato depot manufacturer in Taiwan, agro-meteorological information provided by an agricultural improvement farm in Taiwan, and the wholesale price of sweet potato provided by Taiwan Council of Agriculture. Combining the information from the planting day to 4 days after planting, we determine the features of the model that predicts excellent potato yield during the early planting stage of sweet potatoes. We investigate hierarchical feature clustering and simulated annealing feature selection, removing features that are unrelated to yield prediction, and redundant features that are highly correlated with other features. Our results show that the 3-day-after-planting information may provide predictable information in the feature selection stage, and the wholesale price of sweet potato was found to be related to the premium-grade yield in the experiment.