Machine learning on multi-spectral imagery to estimate nutrient yield of mixed-species cover crops

Authors: Kharel, Tulsi; Tyler, Heather; Mubvumba, Partson; Huang, Yanbo; Bhandari, Ammar; Fletcher, Reginald; Anapalli, Saseendran; JOSHI, DEEPAK - Kansas State University; Mengistu, Alemu; Birru, Girma; Adhikari, Kabindra; DHAKAL, MADHAV - Mississippi State University; Maskey, Mahesh; Reddy, Krishna; CLAY, DAVID - South Dakota State University

Submitted to: Agricultural & Environmental Letters
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 12/17/2024
Publication Date: 1/16/2025
Citation: Kharel, T.P., Tyler, H.L., Mubvumba, P., Huang, Y., Bhandari, A.B., Fletcher, R.S., Anapalli, S.S., Joshi, D.R., Mengistu, A., Birru, G.A., Adhikari, K., Dhakal, M., Maskey, M.L., Reddy, K.N., Clay, D.E. 2025. Machine learning on multi-spectral imagery to estimate nutrient yield of mixed-species cover crops. Agricultural & Environmental Letters. https://doi.org/10.1002/ael2.70009.
DOI: https://doi.org/10.1002/ael2.70009

Interpretive Summary: Cover crop (CC) mixes should consider not only biomass amount but also economic, ecological service, and functional performance for effective nutrient, moisture, and weed suppression, as well as carbon addition to the cropping system. Quantifying these benefits using ground-based physical measurements is time-consuming and expensive. There is growing interest in the agricultural sector in using remotely sensed imagery to estimate cover crop biomass and its benefits to the cropping system. Scientists from the USDA-ARS and universities selected four ongoing experiments from Crop Production Systems Research Unit, Stoneville, MS to evaluate the feasibility of mixed species biomass and nutrient yield estimation using multispectral imagery. Eleven cover crop treatments with varying grass-legume proportions (GLP) were sampled, and nutrient contents were determined. Biomass sampling occurred during the first and fourth weeks of March and the fourth week of April 2023. A multispectral camera mounted on an unmanned aerial vehicle (UAV) captured imagery of the study sites during each sampling period. Within the range of GLP and dry matter production observed in this study, vegetation indices (VIs) such as the chlorophyll absorption ratio (CARI) and the normalized difference vegetation index (NDVI) were strongly correlated with total NPK yield from biomass. Machine learning algorithms, random forest (RF) and partial least squares (PLS) regression, developed using combined imagery datasets (three time points), were better for biomass (R² = 0.74 with RF) and biomass N% (R² = 0.72 with PLS) prediction compared to the Bio NPK prediction. Biomass N% and K% decreased with increasing GLP when all three time point data were combined. This is attributed to nutrient contents in biomass being higher during the early growth stage, while canopy closure and biomass were lower, resulting in smaller NDVI values compared to the later stage when canopy coverage increased and nutrient contents decreased. By combining satellite-based VIs and cover crop mixture proportions, the relationship between VIs and biomass NPK presented in this study allows for the direct estimation of cover crop nutrient benefits. Growers can apply these NPK credits to their main crop nutrient budgeting.

Technical Abstract: Cover crops (CC) can be used to improve soil health, water quality and crop productivity. However, quantifying these benefits using ground-based physical measurement is time consuming and expensive. Therefore, the objective of this study was to estimate mixed species cover crop biomass and nutrient contents using a remote sensing approach. Eleven cover crop treatments with varying grass-legume proportions (GLP) were sampled and nutrient contents were determined. Biomass sampling occurred during the first and fourth weeks of March and the fourth week of April 2023. A multispectral camera mounted on a unmanned aerial vehicle (UAV) captured imagery of the study sites during each sampling period. Eleven vegetation indices (VIs) were developed using multispectral imagery band combinations. Biomass N (R2 = 0.46-0.60) and K% (R2 = 0.41 -0.71) decreased with increasing GLP which was only weakly correlated to biomass (R2 = 0.04-0.15) and P% (R2 < 0.01). The chlorophyll absorption ratio (CARI) and the normalized difference vegetation index (NDVI) closely followed the biomass NPK yield (Bio_NPK) pattern due to the stronger correlation of VIs (R2 = 0.65 CARI and R2 = 0.74 NDVI) with combined NPK yield (Bio_NPK). Machine learning algorithms random forest (RF) and partial least square (PLS) regression developed using combined imagery dataset (3 time point) were better for biomass (R2 = 0.74 with RF) and biomass N% (R2 = 0.72 with PLS) prediction compared to the Bio_NPK prediction. These results are crucial for scientists to devise appropriate analysis approaches for estimating the benefits of mixed species cover crops.