Predicting on multi-target regression for the yield of sweet potato by the market class of its roots upon vegetation indices

Abstract

Single-target regression can accurately predict the crop’s performance but fails to generalize problems with more than one true and cross-validatable solution. An alternative to output multiple numeric values upon the input, we think, would be multi-target regression (MTR) with either Random Forest (RF) or k-nearest neighbors (KNN). Therefore, we captured the advantages of high-resolution remote sensing and multi-target machine learning into an immersive single framework then analyzed if it could be possible for accurately predicting for the yield of sweet potato by the market class of its tuberous roots (i.e., Extra < 0.15 kg; 015 ≤ Extra AA ≤ 0.45 kg; and Extra A > 0.45 kg) upon imagery data on summer and winter full-scale fields. The remote sensing captured the spectral changes on both fields and enabled the MTR to accurately predict for the yield of sweet potato in total and by the market class of harvestable roots upon normalized difference vegetation index (NDVI) and its derivative version (GreenNDVI) as well as upon soil-adjusted vegetation index (SAVI). The SAVI-RF framework predicted the summer field to yield marketable roots at the proportions of 2.04 t ha−1 Extra, 3.89 t ha−1 Extra AA and 2.08 t ha−1 Extra A, and the spectral data from the mid-stage of cultivation at 296 growing degree days (GDD) minimized its mean absolute error (MAE) to 2.66 t ha−1. The GNDVI-RF framework predicted the winter field to yield 1.64 t ha−1 Extra, 5.02 t ha−1 Extra AA and 3.65 t ha−1 Extra A, with an error of 3.45 t ha−1 upon spectral data from sampling on the late stage at 966 GDD. Our insights are timely an absolutely will open up the horizons for harvesting high-quality roots to commercialization, industrialization and propagation, and scaling up this essentially provocative yet emerging crop for food safety and energy security.