Machine learning as a tool to predict potassium concentration in soybean leaf using hyperspectral data

Abstract

The soybean grain yield is affected by several factors, among them, the nutritional deficiency caused by low levels of potassium (K⁺) is one of the main responsible for the reduction in grain yield both in Brazil and worldwide. Traditional methods of nutrient determination involve leaf collection and laboratory procedures with toxic reagents, which is a destructive, time-consuming, expensive, and environmentally unfriendly method. In this context, the use of hyperspectral data and machine learning regression models can be a powerful tool in the nutritional diagnosis of plants. However, the comparison among different machine learning algorithms for K⁺ estimation in soybean leaves from hyperspectral reflectance data is yet to be reported. From this, the goal of this research was to obtain K⁺ prediction models in soybean leaves at different stages of development using hyperspectral data and machine learning regression models with wavelength selection algorithms. The experiment was carried out at the National Soybean Research Centre (Embrapa Soja) in the 2017/2018, 2018/2019 and 2019/2020 soybean crop season, at the stages of development V4–V5, R1–R2, R3–R4 and R5.1–R5.3. The experimental plots were managed to obtain different conditions of K⁺ availability for the plants, from severe deficiency level to the appropriate level of nutrient, under the following experimental treatments: severe potassium deficiency, moderate potassium deficiency and adequate supply of potassium. Spectral data were obtained by the ASD Fieldspec 3 Jr. hyperspectral sensor in the visible/near-infrared spectral range (400–1000 nm) and correlated to leaf K⁺ through ten machine learning methods: Partial Least Square Regression (PLSR), interval Partial Least Squares (iPLS), Genetics Algorithm (GA), Competitive Adaptive Reweighted Sampling (CARS), Random Frog (RF, Frog), Variable combination population analysis (VCPA), Principal Component Regression (PCR), Support Vector Machine (SVM), Successive projections algorithm (SPA), and Stepwise. The results showed that K⁺ deficiency significantly reduce grain yield and nutrient content in the leaf, making enabling the clustering separation of all treatments by Tukey’s test. Among the 601 wavelengths obtained by the sensor, the algorithms selected from 1 to 33.28%, largely distributed in the regions of red, green, blue, red-edge and NIR. In all stages of development, it was possible to quantify the nutrient with high accuracy (R² ≅ 0.88). The multivariate regression models from the selection of variables contributed to increase the accuracy (R²) in about 7.65% for the calibration step and 6.45% for the cross-validation step, when compared to the model using the full spectra. The results obtained demonstrate that the monitoring of K⁺ in soybean leaves is possible and has the potential to determine the nutritional content in the early stages of plant development.