Skip to main content
ARS Home » Midwest Area » Columbia, Missouri » Cropping Systems and Water Quality Research » Research » Publications at this Location » Publication #185961
Title: RELATING IMAGE-DERIVED VEGETATION INDICES TO CROP YIELD

Author
item JANG, GAB-SUE - U OF MO
item Sudduth, Kenneth - Ken
item HONG, SUK-YOUNG - NIAST, REP OF KOREA
item Kitchen, Newell
item PALM, HARLAN - U OF MO

Submitted to: Biennial Workshop on Aerial Photography, Videography, and High Resolution Digital Imagery for Resource Assessment Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 3/1/2006
Publication Date: 3/1/2006
Citation: Jang, G., Sudduth, K.A., Hong, S., Kitchen, N.R., Palm, H.L. 2006. Relating image-derived vegetation indices to crop yield. In: Yang, C., Everitt, J.H. (editors), Proceedings 20th Biennial Workshop on Aerial Photography, Videography, and High Resolution Digital Imagery for Resource Assessment. October 4-6, 2005, Weslaco, Texas. 2006 CDROM.

Interpretive Summary: Precision agriculture requires being able to efficiently and economically collect and interpret data describing variability within cropped fields. For some time, remote sensing from satellites or airplanes has been used to describe differences in vegetation vigor and productivity over large areas, such as watersheds, counties, or states. Using this approach to understand within-field grain crop variability could allow pre-harvest estimates of yield, and might enable mapping of yield variations without use of a combine yield monitor, both of which would be advantageous to farmers. Our goal in this study was to determine the feasibility of estimating corn and soybean yield from aerial hyperspectral remote sensing data. In contrast to conventional, or multispectral, remote sensing data where only a few data bands are obtained, hyperspectral data includes tens to hundreds of bands. The specific sensor we used had 20 to 24 bands and provided data for each 6-foot by 6-foot square within our two research fields. We collected images three times during each of two growing seasons, and also obtained grain yield maps using combine yield monitors. We tested a number of band combinations defined by previous researchers for their ability to estimate yield variations. We also used statistical analysis to examine the data for better band combinations. Of the predefined combinations, we found that a ratio of near-infrared to green light reflectance gave us the best yield estimates in this data. Statistically-derived band combinations gave similar results. Although we had good results in a year when crop yields were reduced due to drought, poor results were obtained in another year when rainfall was timely and crop yields were not water-limited. Further work will be required to identify causes of this difference in estimation quality and to improve results. This research will benefit other researchers and practitioners who may be interested in using hyperspectral reflectance data to understand crop yield differences for precision agriculture.

Technical Abstract: Combinations of visible and near-infrared (NIR) bands in an image are widely used for estimating vegetation vigor and productivity. Using this approach to understand within-field grain crop variability could allow pre-harvest estimates of yield, and might enable mapping of yield variations without use of a combine yield monitor. The objective of this study was to estimate within-field variations in crop yield using vegetation indices derived from hyperspectral images. Hyperspectral images were acquired using an AISA aerial sensor on multiple dates during the 2003 and 2004 cropping seasons for corn and soybean fields in central Missouri. Vegetation indices, including intensity normalized red (NR), intensity normalized green (NG), normalized difference vegetation index (NDVI), green NDVI (gNDVI), and soil-adjusted vegetation index (SAVI), were derived from the images using wavelengths from 440 nm to 850 nm, with bands selected using an iterative procedure. Accuracy of yield estimation models based on these vegetation indices was assessed by comparison with combine yield monitor data. In 2003, late-season NG provided the best estimation of both corn (r^2 = 0.632) and soybean (r^2 = 0.467) yields. Stepwise multiple linear regression using multiple hyperspectral bands was also used to estimate yield, and explained similar amounts of yield variation. Remote sensing was better able to estimate yields in the 2003 season when crop growth was limited by water availability, especially on drought-prone portions of the fields. In 2004, when timely rains during the growing season provided adequate moisture across entire fields and yield variability was less, remote sensing estimates of yield were much poorer (r^2 < 0.3).