A Machine Learning Approach for NDVI Forecasting based on Sentinel-2 Data

The Normalized Difference Vegetation Index (NDVI) is a well-known indicator of the greenness of the biomes. NDVI data are typically derived from satellites (such as Landsat, Sentinel-2, SPOT, Plèiades) that provide images in visible red and near-infrared bands. However, there are two main complications in satellite image acquisition: 1) orbits take several days to be completed, which implies that NDVI data are not daily updated; 2) the usability of satellite images to compute the NDVI value of a given area depends on the local meteorological conditions during satellite transit. Indeed, the discontinuous availability of up to date NDVI data is detrimental to the usability of NDVI as an indicator supporting agricultural decisions, e.g., whether to irrigate crops or not, as well as for alerting purposes. In this work, we propose a multivariate multi-step NDVI forecasting method based on Long Short-Term Memory (LSTM) networks. By careful selection of publicly available but relevant input data, the proposed method has been able to predict with high accuracy NDVI values for the next 1, 2 and 3 days considering regional data of interest.