Crop Phenology Stage Forecasting and Detection using NDVI Time-Series and LSTM

International audienceForecasting the crop phenology helps in crop production estimation, irrigation scheduling and in crop classification. In this article, we propose a methodology that uses the Normalized Difference Vegetation Index (NDVI), derived from the bands of an optical multispectral satellite (Sentinel-2) data for forecasting crop phenology stages using a Long Short-Term Memory (LSTM) network. Firstly, LSTM is used for forecasting the NDVI values until the end of season and then the stage estimation technique is applied on this forecast data to get the phenology points. This method is experimented on a region in India, where the potato crop was grown. The performance of the LSTM model on test dataset are: MAE≈ 0.19, MSE≈ 0.06 and RMSE≈ 0.23, and the phenology points are detected correctly with a mean deviation of ±10 days