Comparison of different machine learning algorithms for predicting maize grain yield using UAV-based hyperspectral images

Timely and accurately predicting maize grain yields will contribute to making adaptive measures to improve management practice and to adjust consumption patterns for ensuring food security. Unmanned aerial vehicles (UAV) are widely used to obtain high-temporal and high-spatial resolution remote sensing images of crops, enabling a possible sensor performance comparison. To date, few studies have compared the potential abilities of multispectral-based and hyperspectral-based images, only sensitive spectral wavelength and full hyperspectral spectra, and various machine learning approaches in estimating physiological characteristics such as chlorophyll meter values, leaf area index (LAI), and agricultural grain yields in high vegetation coverage. In this study, the multispectral and hyperspectral images with the ground measurement of crop traits were collected on 13 and 22 September 2021 in Nanpi experimental station, CangZhou, China. The potential ability of multispectral and hyperspectral images for estimating chlorophyll meter values, retrieving LAI, and predicting maize grain yields were explored and compared using the formed two-band (2D) vegetation indices (VIs) and 2D textural indices (TIs). The sensitive spectral wavelengths were confirmed using correlation analyses, then the sensitive spectral wavelength formed VIs and the full hyperspectral spectra were also compared for predicting maize grain yield using five commonly applied machine learning approaches and five deep learning approaches of convolutional neural network (CNN). The results indicated the narrow bands of hyperspectral remained high sensitive with chlorophyll meter values, leaf area index (LAI), and agricultural grain yields than multispectral images in high vegetation coverage. The adoption of full hyperspectral spectra significantly improved the accuracy of maize grain yield predictions compared with adopting VIs built only using sensitivity spectral wavelength. Based on selected VIs, random forest regression (RF) and LightGBM achieved the highest accuracy, R2 (RMSE) were 0.90 (0.55 t/ha), and 0.85 (0.59 t/ha), respectively. While based on full hyperspectral spectra, RF and CNN150 performed the best, with R2 (RMSE) being 0.92 (0.53 t/ha), and 0.91 (0.59 t/ha), respectively. This research concluded the integration of full hyperspectral spectra in combination with RF were highly recommended for predicting maize grain yields, especially for crops in high vegetation coverage.