A comparison of empirical and neural network approaches for estimating corn and soybean leaf area index from Landsat ETM+ imagery

Abstract

Plant foliage density expressed as leaf area index (LAI) is used in many ecological, meteorological, and agronomic models, and as a means of quantifying crop spatial variability for precision farming. LAI retrieval using spectral vegetation indices (SVI) from optical remotely sensed data usually requires site-specific calibration values from the surface or the use of within-scene image information without surface calibrations to invert radiative transfer models. An evaluation of LAI retrieval methods was conducted using (1) empirical methods employing the normalized difference vegetation index (NDVI) and a new SVI that uses green wavelength reflectance, (2) a scaled NDVI approach that uses no calibration measurements, and (3) a hybrid approach that uses a neural network (NN) and a radiative transfer model without site-specific calibration measurements. While research has shown that under a variety of conditions NDVI is not optimal for LAI retrieval, its continued use for remote sensing applications and in analysis seeking to develop improved parameter retrieval algorithms based on NDVI suggests its value as a “benchmark” or standard against which other methods can be compared. Landsat-7 ETM+ data for July 1 and July 8 from the Soil Moisture EXperiment 2002 (SMEX02) field campaign in the Walnut Creek watershed south of Ames, IA, were used for the analysis. Sun photometer data collected from a site within the watershed were used to atmospherically correct the imagery to surface reflectance. LAI validation measurements of corn and soybeans were collected close to the dates of the Landsat-7 overpasses. Comparable results were obtained with the empirical SVI methods and the scaled SVI method within each date. The hybrid method, although promising, did not account for as much of the variability as the SVI methods. Higher atmospheric optical depths for July 8 leading to surface reflectance errors are believed to have resulted in overall poorer performance for this date. Use of SVIs employing green wavelengths, improved method for the definition of image minimum and maximum clusters used by the scaled NDVI method, and further development of a soil reflectance index used by the hybrid NN approach are warranted. More importantly, the results demonstrate that reasonable LAI estimates are possible using optical remote sensing methods without in situ, site-specific calibration measurements.