Multiscale Opportunistic Fusion of NEON, LANDSAT OLI, MODIS and VIIRS DATA for the Validation of Satellite Based Vegetation Index Time Series

Abstract

Space-borne remote sensing-based Vegetation Index data have historically been used as an indicator of green vegetation health and productivity. Vegetation Indices are robust, empirical and proxy measures of vegetation activity at the land surface. They are designed to enhance the vegetation reflected signal from measured spectral responses by combining two (or more) wavebands, often in the red (0.6 - 0.7 µm) and NIR wavelengths (0.7-1.1 µm) region. Vegetation indices are also used to support questions related to plant responses to climate change. The VI time series from various Earth Observing Systems is one of the longest records to date, providing data covering more than 40 years. It is an invaluable data about ecosystem functioning and status over space and time. A large number of satellite sensors are used to construct vegetation indices aiming at observing and monitoring the environmental. Validation of these VI time series still depends on in-situ approaches to establish the accuracy and errors of satellite-derived VIs, but in-situ validation is limited in spatial footprint and is very costly. Yet, in-situ validation remains a critical endeavor because of issues related to spatial, spectral, temporal resolutions and processing methods differences across these sensors and missions. Here, we have developed an opportunistic validation approach in support of long-term VI time series data from sensors like the Moderate Resolution Imaging Spectroradiometer (MODIS), the Visible Infrared Imaging Radiometer Suite (VIIRS), the Advanced Very High-Resolution Radiometer (AVHRR), and Landsat. The approach is based on the opportunistic use of high-resolution data from the National Ecological Observatory Network (NEON) Airborne Observation Platform (AOP) hyperspectral sensor. NEON is an NSF funded effort for tracking and documenting the ecosystem change over 20 eco-climatic domains across the US. In this work, NEON AOP data is assumed to be ground truth, due to its hyperspectral nature, high resolution, and proximity to the ground which eliminates most of the atmosphere the source of most issues in remote sensing. In this work we used data from eight NEON sites, representing various types of land cover conditions ranging from dense to mixed forest to desert and semi-arid biomes. This validation framework consists of three steps: 1) Spatial and spectral convolution of the NEON data, 2) Statistical analysis to compute the uncertainty, precision, accuracy, (UPA) and errors in the data records, 3) Correlation analysis across the different sensors to elucidate sensor continuity. The general objective is to extrapolate these site-specific results and provide a systematic characterization of the VI data record error and uncertainty budget over time and space and from the various synoptic sensors. Overall, our results point to a general trend of high accuracy and precision in the global VI time series, a strong intrinsic continuity across the different sensors at both ends of the vegetation density spectrum, and slightly more noise and lack of accuracy in heterogeneous and open canopies. This is because the results are mediated by different factor in the spatial and temporal domains. Results indicate high correlation among MODIS and Landsat VI data (coefficient of determination~95% and p value <0.01). For VIIRS we observed some difference compared to the other two sensors. For instances, NDVI global relationships are same for Landsat and MODIS whereas VIIRS shows weaker correlation than others. Unlike NDVI, VIIRS exhibited similar behavior as MODIS for EVI2 global correlation. In addition, concordance plots show similar correlation which helped to verify the results obtained from linear regression. From the concordance correlation analysis, we evaluated 95% confidence interval to understand range of datapoints. For most of the sites sample datapoints fall into three sigma limit, which indicated that the model operates efficiently and produced result of highest quality. For dense forest areas exception is occurred mostly due to cloud. Also, these differences can be occurred due to the distinction between sensors resolutions. For sparsely vegetated and dense forest areas the datasets were highly precise (50%) indicating the strong compactness of data points with small standard deviation. This was in contrast to the semi-arid regions and less dense forests areas, that showed more data spread from the average value, and higher standard deviations, indicating less precision. These results are consistent with previous research works that suggest homogeneity versus heterogeneity play a major role in data accuracy and noise.