Combining a cotton 'boll area index' with in-season unmanned aerial multispectral and thermal imagery for yield estimation

Authors: SIEGFRIED, JEFFREY - Kansas State University; Adams, Curtis; RAJAN, NITHYA - Texas A&M University; HAGUE, STEVE - Texas A&M University; SCHNELL, RONNIE - Texas A&M University; HARDIN, ROBERT - Texas A&M University

Submitted to: Field Crops Research
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 11/18/2022
Publication Date: 12/5/2022
Citation: Siegfried, J., Adams, C.B., Rajan, N., Hague, S., Schnell, R., Hardin, R. 2022. Combining a cotton 'boll area index' with in-season unmanned aerial multispectral and thermal imagery for yield estimation. Field Crops Research. 291. Article 108765. https://doi.org/10.1016/j.fcr.2022.108765.
DOI: https://doi.org/10.1016/j.fcr.2022.108765

Interpretive Summary: Accurately estimating yield of cotton crops without physical harvest would be of great benefit for precision agriculture, breeding programs, and other applications. Unmanned aerial systems (UAS) allow collection of crop imagery with unprecedented temporal, spatial, and spectral resolutions and could be better leveraged for this purpose. In this four-year study conducted at College Station, TX, we developed a relatively simple multispectral image classification technique for cotton yield estimation termed “Boll Area Index” or BAI, which is collected after defoliation. We also investigated the potential for additional multispectral and thermal crop indices collected in-season to improve estimate accuracy. BAI was associated with cotton yield each year (R2 = 0.61 - 0.79). Multiple linear regression that included BAI, vegetation indices, and/or canopy temperature from two flight dates produced better yield estimates (R2 = 0.79 - 0.89) than BAI alone. Cross validation of the regression models had R2 values that varied from 0.51 to 0.88 and RMSE varied from 273 to 508 kg ha-1. This is a level of error that may be acceptable for some purposes and not others. For example, the technique may be acceptable for screening cotton lines in the early stages of breeding selection, but may be unacceptable for screening advanced lines when accuracy is crucial. Overall, the results indicate that imagery from just two or three UAS flights provides a detailed dataset for cotton yield prediction, while limiting required labor and computational resources.

Technical Abstract: Crop yield data is critical for precision agriculture, breeding programs, and other activities, but collecting this data at fine scales is labor-intensive. Unmanned aerial systems (UAS) allow collection of imagery with unprecedented temporal, spatial, and spectral resolutions and could be better leveraged to estimate or predict yield while limiting labor requirements. Therefore, the objectives of this study were to develop a relatively simple pixel-based multispectral image classification technique for cotton (Gossypium hirsutum L.) yield estimation, termed “Boll Area Index” or BAI, which is collected after defoliation, and to identify in-season co-predictors derived from multispectral and thermal imagery to improve estimate accuracy. A field study was conducted over four growing seasons (2017-2020) at College Station, TX. The experimental treatments included three irrigation rates (0, 40, and 80% ETc replacement) and eight commercial cotton cultivars each year. Multispectral and thermal infrared imagery were captured biweekly. In addition to BAI, three vegetation indices (Normalized Difference Vegetation Index or NDVI, Normalized Difference Red Edge or NDRE, and Optimized Soil Adjusted Vegetation Index or OSAVI) and canopy temperature were derived from orthomosaics and analyzed. There were positive linear relationships between BAI and seed cotton yield each year (R2 = 0.61 - 0.79). Multiple linear regression including BAI, vegetation indices, and/or canopy temperature from two flight dates produced better yield estimates (R2 = 0.79 - 0.89) than BAI alone. Cameras or payloads with both optical and thermal sensors are ideal for strictly in-season yield estimation endeavors, but thermal was not necessary when BAI was included in the models because canopy temperature provided minimal improvement as a third predictor. Multiple regressions involving NDVI and BAI already had quite strong relationships with yield (R2 = 0.7 to 0.87) without including canopy temperature. Cross validation of multiple linear regression models derived from BAI and NDVI, using data from two years to predict yield in a third, had R2 values that varied from 0.51 to 0.88 and RMSE varied from 273 to 508 kg ha-1. This is a level of error that may be acceptable for some purposes, such as screening lines in early stages of cotton breeding selection, but may be unacceptable for screening of advanced lines when accuracy is crucial. Overall, the results indicate that derivatives from just two or three UAS flights presents a detailed dataset for cotton yield prediction, while limiting labor and required computational resources.