Chlorophyll estimation in soybean leaves infield with smartphone digital imaging and machine learning

Abstract

Soybean (Glycine max (L.) Merrill) leaf chlorophyll content is indicative of the plant growth and health issues. However, chlorophyll measurement using the standard chemical procedure is laborious, while the sensor-based electronic options, such as soil plant analysis development (SPAD) meter tend to be highly expensive and made only spot measurements. Therefore, a simpler and less expensive infield method of chlorophyll measurement in soybeans using smartphone camera with image processing and machine learning models was developed. Soybean leaf images (720 images) and SPAD readings were collected from different cultivars (4), with replications (3) and sampling dates (2) from experimental plots. Of the several color vegetation indices (CVIs) tested, the dark green color index (DGCI) had the best correlation with SPAD meter readings (r=0.90), which was further improved by color calibration (r=0.93). The results of the random coefficients model showed that both cultivars and sampling dates had no significant effect (0.06≤P≤0.96), hence data were combined for the analysis. The simpler statistical linear regression (SLR) and polynomial regression (PR), multiple linear regression as well as the advanced machine learning models (support vector machine (SVM), random forest (RF)) tested with color scheme inputs (RGB, DGCI, range pixel count (RPC) of DGCI, and ‘Both’ (RPC + RGB)) produced the best chlorophyll prediction with DGCI, RPC, and ‘Both’ inputs (0.87<R2<0.89; 2.90≤RMSE≤3.41 SPAD units). Overall, these models were not significantly different, but the SVM model found to be the best (R2=0.89 and RMSE=2.90SPAD units). The simpler SLR and PR models with DGCI input (R2≥0.87 and RMSE≤3.1 SPAD units) performed as good as the advanced SVM and RF models. The SVM model had the potential of predicting the chlorophyll directly with the raw RGB input (R2=0.86 and RMSE=3.20SPAD units) without the need of using the standard calibration board. The developed methodology of image processing with machine learning modeling and conversion relationship of measuring infield soybean leaf chlorophyll is efficient, inexpensive, not requiring the standard calibration board, and can be easily extended to other large-scale aerial imaging platforms and field crops.