A cross-scale transfer learning framework: prediction of SOD activity from leaf microstructure to macroscopic hyperspectral imaging

Abstract

Superoxide dismutase (SOD) plays an important role to respond in the defence against damage when tomato leaves are under different types of adversity stresses. This work employed microhyperspectral imaging (MHSI) and visible near-infrared (Vis–NIR) hyperspectral imaging (HSI) technologies to predict tomato leaf SOD activity. The macroscopic model of SOD activity in tomato leaves was constructed using the convolutional neural network in conjunction with the long and short-term temporal memory (CNN-LSTM) technique. Using heterogeneous two-dimensional correlation spectra (H2D-COS), the sensitive macroscopic and microscopic absorption peaks connected to tomato leaves' SOD activity were made clear. The combination of CNN-LSTM algorithm and H2D-COS analysis was used to research transfer learning between microscopic and macroscopic models based on sensitive wavelengths. The results demonstrated that the CNN-LSTM model, which was based on the FD preprocessed spectra, had the best performance for the microscopic model, with R_C and R_P reaching 0.9311 and 0.9075, and RMSEC and RMSEP reaching 0.0109 U/mg and 0.0127 U/mg respectively. There were 10 macroscopic and 10 microscopic significant sensitivity peaks found. The transfer learning was carried out using sensitive wavelengths, and the model performed well with an R_P value of 0.7549 and an RMSEP of 0.0725 U/mg. The combined CNN algorithm and H2D-COS analysis demonstrated the viability of transfer learning across microscopic and macroscopic models for quantitative tomato leaf SOD prediction.