Figure 8.
The four confusion matrices, panels a to d, compare ground truth and prediction for R 1 C P, R 2 L S, R 3 S S, R 4 C F, R 5 R L C A, R 6 S M, R 7 G, R 8 F W, R 8 W, R 9 G P, R 10 P and R 11 I. Panel a has almost all values on the main diagonal, with only a few misclassifications. Panel b retains high diagonal counts, generally between 249 and 256, with several small off-diagonal counts. Panel c also has predominantly high diagonal values, although R 2 L S has 238 correct predictions and 14 predictions as R 3 S S. Panel d contains larger off-diagonal counts, including 162 R 2 L S predictions as R 4 C F and 213 R 11 I predictions as R 4 C F.Confusion matrices obtained by applying the independently trained models to their corresponding test subsets: (a) laser-induced breakdown spectroscopy, (b) Raman, (c) ultraviolet–visual and (d) near infrared (R1-CP, cement paste; R2-LS, limestone sand; R3-SS, siliceous sand; R4-CF, ceramic fraction; R5-RLCA, recycled limestone concrete aggregate; R6-SM, siliceous mortar; R7-G, gypsum; R8-FW, wood fibres; R8-W, wood chips; R9-GP, glass powder; R10-P, polyethylene; R11-I, iron)