A machine‐learning based ConvLSTM architecture for NDVI forecasting

Abstract Normalized difference vegetation index (NDVI) is an essential remote measurement for agricultural studies because of its strong correlation with crop growth and yield. Accurate and comprehensive NDVI forecasts thus provide effective future projections of crop yield for precise agricultural planning and budgeting. Previous recurrent neural network (RNN) based forecasting methodologies have only performed single‐pixel or large‐area‐average NDVI predictions. We present an alternative RNN‐based deep‐learning architecture, the convolutional long short‐term memory (ConvLSTM), to supply much more comprehensive and detailed NDVI forecasts. In this paper, a single ConvLSTM is capable of 10,000‐pixel field‐level NDVI predictions, providing a more practical methodology for agricultural producers than single‐pixel studies. We compare our model to the parametric crop growth model (PCGM), another multipixel field‐level NDVI forecasting technique. We test each model over the same set of soybean crop field pixels with the root mean square error (RMSE) metric. The training configuration of each model is defined by the number of seasons of historical data used for weight optimization. When the best training configuration of the model found is used, the ConvLSTM obtains an RMSE of 0.0782, outperforming the PCGM's RMSE of 0.0989 (an improvement of 0.0207 in precision represents a large gain in the accuracy of production volume prediction when projected into large production areas). Finally, by comparing the ConvLSTM predictions with the ground truth data over the entire target region rather than just the soybean crop pixels, we discover that the ConvLSTM can also predict NDVI values over the nonsoybean crop as effectively.