Within-Field Yield Prediction for Sugarcane and Rice Focused on Precision Agriculture Applications

Food and energy security are two main topics when it comes to the on-growing world population. Rice and sugarcane play an important role in this scenario since sugarcane can be used for energy production and rice is one of major staple cereals. In this scenario, Precision Agriculture (PA) management strategies aims to improve productivity, efficiency, profitability, and sustainability, and can help agriculture to fulfill the needs of the growing population in a sustainable way. However, yield maps are essential for PA, but its adoption is still very low. Thus, the main objective of this study was to evaluate the potential of satellite imagery and machine learning to predict yield maps that could support the adoption of precision agriculture practices for rice and sugarcane. Consequently, a framework for the data processing, imagery acquisition and machine learning model generation, was proposed and tested. The results presented a high potential for the usage of those techniques, generating yield maps very similar to the ones obtained from yield monitors (RMSE for rice of 0.9Mg.ha-1 and for sugarcane 3.14Mg.ha-1). Also, in-season yield map prediction was evaluated for rice and sugarcane. Therefore, the prediction was performed for different growth stages by stacking all the images until a specific date. Sugarcane maps were obtained with a satisfactory accuracy early in the season (May-June) (no statistical significance when compared to the predicted maps of the end of the season) whilst for rice the yield maps with the lowest errors were only obtained late in the season. Therefore, sugarcane maps obtained early in the season could be used for in-season management of the crop. On the other hand, the in-season applicability for rice yield maps were limited since accurate maps were obtained at late ripening. However, this information could still be used for harvest planning and nitrogen application on the second harvest of Louisiana's rice. In general, the framework proposed presented a high potential to be used for yield maps prediction. Furthermore, yield maps, an important tool for PA, were obtained with low errors RMSE of 0.83 and 3.14 Mg.ha-1 for rice and sugarcane, respectively.