Correlation maps to assess soybean yield from EVI data in Paraná State, Brazil

ABSTRACT Vegetation indices are widely used to monitor crop development and generally used as input data in models to forecast yield. The first step of this study consisted of using monthly Maximum Value Composites to create correlation maps using Enhanced Vegetation Index (EVI) from Moderate Resolution Imaging Spectroradiometer (MODIS) sensor mounted on Terra satellite and historical yield during the soybean crop cycle in Paraná State, Brazil, from 2000/2001 to 2010/2011. We compared the ability of forecasting crop yield based on correlation maps and crop specific masks. We ran a preliminary regression model to test its ability on yield estimation for four municipalities during the soybean growing season. A regression model was developed for both methodologies to forecast soybean crop yield using leave-one-out cross validation. The Root Mean Squared Error (RMSE) values in the implementation of the model ranged from 0.037 t ha−1 to 0.19 t ha−1 using correlation maps, while for crop specific masks, it varied from 0.21 t ha−1 to 0.35 t ha−1. The model was able to explain 96 % to 98 % of the variance in estimated yield from correlation maps, while it was able to explain only 2 % to 67 % for crop specific mask approach. The results showed that the correlation maps could be used to predict crop yield more effectively than crop specific masks. In addition, this method can provide an indication of soybean yield prior to harvesting.

摘要：植被指数被广泛应用于监测作物生长发育，且通常作为模型输入数据用于作物产量预测。本研究的第一步为采用月度最大值合成法，基于搭载于Terra卫星的中分辨率成像光谱仪（Moderate Resolution Imaging Spectroradiometer, MODIS）获取的增强型植被指数（Enhanced Vegetation Index, EVI），结合巴西巴拉那州2000/2001至2010/2011年度大豆生育期内的历史产量数据构建相关系数图。本研究对比了基于相关系数图与作物专属掩膜开展作物产量预测的性能，并构建初步回归模型，测试其在大豆生育期内对4个县域作物产量的估算能力。针对这两种方法，本研究均采用留一交叉验证（leave-one-out cross validation）法构建了大豆产量预测回归模型。采用相关系数图的模型实现过程中，均方根误差（Root Mean Squared Error, RMSE）取值范围为0.037吨/公顷至0.19吨/公顷；而采用作物专属掩膜的模型，其均方根误差取值范围为0.21吨/公顷至0.35吨/公顷。基于相关系数图的模型可解释估算产量中96%至98%的方差变异，而基于作物专属掩膜的方法仅能解释2%至67%的方差变异。研究结果表明，相较于作物专属掩膜法，相关系数图可更高效地实现作物产量预测。此外，该方法可在大豆收获前提供产量预估参考。