Multi-sensor data fusion in the nondestructive measurement of kiwifruit texture

Three mechanical-based techniques, including falling impact (FAI), forced impact (FOI), and acoustic impulse-response (AIR), were implemented for the nondestructive prediction of the apparent modulus of elasticity (Ea) and Magness-Taylor firmness (MTf) of kiwifruit, cv. Hayward. Considering the merits and limitations of each method in estimating the texture parameters, this study tried to improve the performance of the predictive models by using the concept of multi-sensor data fusion. Two different fusion strategies, including low-level and mid-level fusions, were accordingly applied using partial least square regression (PLSR) and principal component analysis combined with artificial neural network (PCA-ANN), respectively. Better predictions of Ea were obtained, as compared to those of MTf, in using each method, as well as their combinations in both fusion strategies, thereby demonstrating the better fitness of Ea with the nondestructive data. Moreover, both fusion strategies enhanced the performance of Ea and, in most cases, MTf predictive models, as compared with using each technique individually. Comparing two fused systems showed that the mid-level fusion was more effective than the low-level one, where in the best fused systems (integration of all three sensors), the standard deviation ratio (SDR) values for Ea and MTf were improved by 11.2% and 9.1%, respectively, and the satisfactory results for both Ea (R2p = 0.926, SDR = 3.07) and MTf (R2p = 0.841, SDR = 2.51) were obtained. This study revealed that compared to the implementation of mechanical-based methods individually, and also the low-level fusion of them, their mid-level fusion using PCA-ANN algorithm could be an effective approach for providing more detailed and complementary information about kiwifruit texture.