Search

This work presents initial results on a multitemporal hyperspectral image analysis method to evaluate the time degradation of pork meat. The proposed method is inexpensive and practically non-destructive. The hyperspectral data is analyzed and the relevant information is reduced to the information in only three wavelengths. The analysis is performed by a binary classifier composed by two stacked autoencoders and a softmax output layer. The use of autoencoders reduces tenfold the dimension of the input space. The proposed classifier has led to 97.2% of correct decisions, which indicates the great potential of the methodology to monitor the safety of meat.

This paper proposes a new kernel Kalman filter formulation for system identification and time series estimation in nonlinear time-varying environments. The unknown nonlinear time-varying function is approximated by a finite-order linear model in a reproducing kernel Hilbert space. The model coefficients define the state of the Kalman filter. Simulation results illustrate the improvement in estimation performance provided by the new algorithm when compared to the classical kernel LMS filter.