A novel blending algorithm for satellite-derived high resolution spatio-temporal normalized difference vegetation index

Abstract

The Landsat and Modis satellites have been providing spectacular imagery of Earth’s surface for over 40 years and 20 years respectively. Joining them is another recent initiative by European Space Agency (ESA) of multispectral Sentinel satellites first launched in 2016. These together provide a variety of vegetation indices, in particular, Normalized Difference Vegetation Index (NDVI) which is widely adopted to analyze vegetation patterns and crop phenological characteristics to understand and predict expected yield, pest attacks, water deficiency etc. However, a big challenge in using the data from multiple satellites to derive common integrated insight is the appropriate bias correction especially as the spatial and temporal resolutions do not match across various satellites. For instance, Landsat has 30mx30m spatial resolution (SR) and 16 days temporal resolution (TR), Modis has 250mx250m spatial resolution but 1 day temporal resolution whereas Sentinel has 10-60mx10-60m spatial resolution with close to 5 days temporal resolution. Along with varying resolutions, presence of significant cloud cover for long time periods is also a factor which affects the performance of vegetation health models based on raw data from these satellites especially in countries like India with long monsoon seasons. We propose a novel blending algorithm to bring together remote sensed data from Landsat, Modis and Sentinel satellites and derive a new enhanced High Definition Normalized Difference Vegetation Index (HDNDVI) for daily vegetation health monitoring at the fine spatial resolution of Sentinel data. The proposed algorithm analyses past observed NDVI data from the above mentioned three satellites, corrects for anomalous sensor data due to cloud cover and sensor saturation, determines the bias between different combinations of satellite data and applies bias correction to derive the HD-NDVI. As it determines the bias from the recent past for a given region, the proposed algorithm can be scaled globally and applied across different geographic regions and seasons. In this paper we present the derived HD-NDVI from our algorithm for Burdwan region of West Bengal in India across the full crop cycle of approximately 98 farmers. We show HD-NDVI images and the latest images of all the three satellites to highlight the differences and significant gain in spatial and temporal resolution achieved by HD-NDVI. The correlation between HD-NDVI and Landsat, Sentinel and Modis satellite data for the entire crop cycle is also analyzed across the full crop cycle. HD-NDVI has the highest correlation with Sentinel and the average, maximum and minimum Pearson Correlation coefficients are 0.9611, 0.992 and 0.59 respectively. Second in terms of correlation is Landsat with average, maximum and minimum correlation coefficients of 0.898, 0.982 and 0.048 respectively. The lowest correlation is with Modis with average, maximum and minimum correlation coefficients of 0.69, 0.87 and -0.00113. The above correlation numbers indicate that HD-NDVI provides daily NDVI at the spatial resolution of Sentinel. Summarizing, HD-NDVI acts as a virtual satellite that has the best of temporal and spatial resolutions compared to the widely adopted Landsat, Sentinel and Modis satellites.