Evaluation of spatiotemporal fusion methods for high resolution daily NDVI prediction

Abstract

The Normalized Difference Vegetation Index was introduced for monitoring vegetation dynamics. This index can be extracted from multispectral sensor data, such as Landsat and MODIS sensors, and therefore the NDVI can be obtained with high spatial resolution but low temporal resolution when using Landsat or with high temporal resolution but low spatial resolution when using MODIS. Spatiotemporal fusion methods were proposed as a solution for this limitation. By using these methods, images with high spatial and high temporal resolution can be obtained. STARFM, ESTARFM and FSDAF are ones of the methods that have been successfully applied for spatiotemporal fusion. The objective of this study is to compare and evaluate these three methods and apply it on actual NDVI Landsat 8 and MODIS data in the region of Tadla in Morocco, to generate daily NDVI at 30m resolution. This evaluation was supervised by experts in CRTS and this through two approaches. The evaluation approach one is applying the three methods to predict Landsat NDVI for 16 days based on predicted images. The evaluation approach two is based on predicting Landsat NDVI for 4 months and evaluating the results with available real Landsat images with statistic parameters. The Results show that only the ESTARFM method can handle the propagation of error for evaluation approach one and it is less sensible to the quality of inputs. For evaluation approach two, the ESTARFM method gives more accurate results than the STARFM and FSDAF method if input two pairs Landsat and MODIS NDVI are used from previous days with a RMSE attending 0,06.