Fusion of Three Optical Sensors for Nondestructive Detection of Water Content in Lettuce Canopies

• Published in: Journal of applied spectroscopy Vol. 88; no. 1; pp. 194 - 202
• Main Authors: Gao, H. Y., Mao, H. P., Zhang, X. D., Ullah, I., Wei, X. H.
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.03.2021; Springer; Springer Nature B.V
• Subjects: Algorithms; Analytical Chemistry; Aquatic resources; Artificial neural networks; Atomic/Molecular Structure and Spectra; Canopies; Color imagery; Correlation coefficients; Empirical analysis; Feature extraction; Forecasting; Genetic algorithms; Lettuce; Machine learning; Moisture content; Optical measuring instruments; Physics; Physics and Astronomy; Principal components analysis; Sensors; Spectra; Spectrum analysis; Substrates; canopy temperature; RGB images; spectrum; nondestructive detection of water content; multisensors data fusion
• ISSN: 0021-9037; 1573-8647
• DOI: 10.1007/s10812-021-01158-8

Experiments were conducted to develop and assess a method by which the water content of a lettuce canopy can be nondestructively detected and estimated using a combination of spectra, RGB images, and canopy temperature. To this end, 130 lettuce samples grown in four different substrate water content levels were collected for data acquisition. In the spectroscopy procedure, five spectral intervals (380 variables) were selected by backward interval partial least squares and were further reduced to 48 wavelength variables, chosen using a genetic algorithm based on Savitzky–Golay smoothing and log (1/R) transformation. Then, 967, 1170, 1221, 1406, 1484, 1942, and 1985 nm optimum spectral variables were selected by the successive projection algorithm. Thirteen plant features were extracted from top- and front-view RGB images. These features comprised morphological, color, and textural features. An empirical crop water stress index was established based on dry and wet reference surfaces via thermal imagery. Subsequently, a principal component analysis was applied to the spectral variables and the image features, and an extreme learning machine was used to construct the multisensor and single-sensor models. The results show that the multisensor model had a correlation coefficient of prediction of 0.9018, which was found to be approximately 9.4 and 15.7% better than that of the spectral and image models. This work demonstrates that integrating spectra, RGB images, and canopy temperature with suitable algorithms offers a high potential for use in the nondestructive measurement of water content in lettuce, considerably improving accuracy over that using a single-sensor modality.