Lokupitiya, E, Lefsky, M, Paustian, K (2010). Use of AVHRR NDVI time series and ground-based surveys for estimating county-level crop biomass. INTERNATIONAL JOURNAL OF REMOTE SENSING, 31(1), 141-158.
Crop biomass and residue production are major components of cropland carbon dynamics that can be estimated using yield data from ground-based surveys. In the USA, surveyed yield data are available at county level and have been widely used for various research, economic and policy purposes, in addition to biomass estimation. However, survey data may be unavailable for certain times and/or locations and thus biomass estimates using remotely sensed data might be used to fill in any missing biomass data for estimating residue production and carbon dynamics in croplands. Compared to ground-based surveys, remotely sensed data are collected on a regular schedule and may also provide more spatially resolved data. We analysed composite biweekly Normalized Difference Vegetation Index (NDVI) data obtained using the Advanced Very High Resolution Radiometer (AVHRR) sensor and crop aboveground biomass (AGBM) estimated from available county-level yield data reported by the National Agricultural Statistics Service (NASS) for three crops (corn, soybean and oats) during 1992, 1997 and 2002. The aim of the study was to explore the relationships between NDVI and crop biomass to complete the missing biomass data in counties where no NASS-reported yields are available for biomass estimation. AGBM was estimated from Pathfinder biweekly NDVI, using canonical correlation analysis (CCA) and best subset multiple regressions incorporating canonical variates from NDVI time series. Cross-validation of model estimates was performed by randomly splitting the dataset into training and application subsets, simulating a 10-40% range of missing values. NDVI and crop biomass in Iowa during a given year were well correlated, with coefficient of determination (R-2) values > 0.8 in most cases. Using the available (training) data from a single year or a combination of years to derive models for filling the missing (validation) data within the same time period yielded a mean estimated biomass with < 1% relative error and bias. However, models applied to out-of-sample years had lower (< 0.4) R-2 values for the relationships between biomass and NDVI, although the mean residuals were low.