Abstract:

Tobacco SPAD value was estimated based on hyperspectral fractional differential, and the aim was to improve the accuracy of hyperspectral data estimating tobacco SPAD value. Firstly, the optimal transformation method was determined to estimate the SPAD value of tobacco leaves, and fractional differential processing was carried out. Then, the characteristic wavelengths were selected based on correlation analysis, out-of-bag(OOB) data importance, and random forest (RF). Finally, the SPAD value estimation model of tobacco leaves was constructed. The results showed that: (1) The characteristic wavelengths used to estimate the SPAD value of tobacco leaves mainly included the green band (499 nm, 500 nm), the red edge band (634 nm, 636 nm, 702 nm, 703 nm, 732 nm), the near-infrared band (972 nm, 1286 nm, 1289 nm, 1295 nm, 1298 nm, 1316 nm), short infrared bands (1450 nm, 1453 nm, 1456 nm, 1806 nm). (2) The RF-SPAD model constructed with the characteristic wavelength of order 1.9 had the highest accuracy, R ²= 0.690, which was 22.1%, 42.6%, 87% higher than the 0, 1 and 2 orders respectively; RMSE= 2.799, which was reduced by 13.5%, 20.2% and 27.8% compared with 0, 1, and 2 orders, respectively. Compared with the integer order model, the RF-SPAD model constructed by using the characteristic wavelength of order 1.9 effectively improved the estimation accuracy of tobacco SPAD value, and provided a new idea for the hyperspectral fractional order differential technology to estimate SPAD value.

Key words: hyperspectral fractional differential, SPAD value, tobacco leaves, importance of out-of-bag data, random forest