Phenology based NDVI time-series compensation for yield estimation analysis

Normalized difference vegetation index (NDVI) has been correlated with various vegetation parameters using different preprocessing methods, corrections and sampling time based on the aim of the study. In yield estimation studies, maximum NDVI value of the season and the same day of the year NDVI value, which are based on chronological sampling time, are used within different techniques from statistical analysis to machine learning. However, analysis of biological systems based on their chronological timing results in an error increase at data extraction phase due to the non-linearity among phenological stages, representing plant development and its variability. In this study, a phenology based optimum NDVI sampling time is determined and proposed as a replacement of chronologically sampled NDVI time for yield estimation analysis. It may not be possible to have or acquire satellite images for the desired NDVI date due to the temporal resolution of existing remote sensing satellites and meteorological limitations. Therefore, a compensation process based on Adaptive Savitzky-Golay filter and using the existing images is proposed to constitute a new NDVI value for the desired day of the season. The study area is situated in the Southeastern Anatolia region of Turkey within the Fertile Crescent where wheat was first cultivated 10000 years ago. The region has the highest durum wheat production, supplying %46 of the whole production in Turkey. 8-day interval, Landsat-7 and Landsat-8 NDVI time-series are analyzed for seasonal vegetation development with TIMESAT software for the 2014-2016 period. Ground-based ancillary data was obtained within the Turkish Agricultural and Environmental Informatics Research and Application Center (TARBIL) project. Trend analysis of NDVI time-series was performed using Adaptive Savitzky-Golay filter, form of a moving average, adapting to the upper envelope of the data points. Two different sampling methods representing chronological and phenological approaches in addition to the max NDVI value are used to determine the optimum NDVI day. Phenological sampling is carried out as 10-day intervals starting from the emergence phase indicating the start of the season whereas 15 April, representing the long-term annual mean peak NDVI date of the study area was used for chronological sampling. Adaptive Savitzky-Golay filtering and different sampling combinations were used to perform correlation analysis with annual yield data. Best sampling method along with the optimum NDVI sampling day of the season was determined based on the correlation analysis. It is observed that the combinations with phenological sampling corresponding to the first node stage according to Food and Agriculture Organization (FAO) guidelines have the highest correlation. Regression analysis between agrometeorological data with and without compensated NDVI and yield variables showed that the usage of compensated NDVI had higher correlation for wheat yield estimation. The results showed that, in comparison with the conventional approaches, the usage of phenology based compensated NDVI, enhanced the yield estimation percentage. Along with the possibility of producing ancillary data from remote sensing images, this approach will minimize the need for ground-based observations that are time and money consuming.