A progeny test of half-siblings of Eucalyptus benthamii was analyzed using wood quality traits combined with volumetric information. The test was designed in completely randomized blocks design with a total of 1394 trees planted in the city of Encruzilhada do Sul, Brazil. At six years of age, all the trees were measured and sawing samples were collected from 87 trees, which were processed and read through the NIR. Regression models were calibrated by the partial least squares (PLS) method to correlate the NIR spectra with wet chemical measurements, allowing the assembly of models to estimate values of basic density and pulp yield for all the population. From the values of basic density and pulp yield, obtained by the models, and average annual increment of wood (MAI), the average annual increment of cellulose (MAIpulp) was calculated for each individual in the test. Two types of optimizations were tested: balanced and unbalanced. The balanced optimization consisted of selecting for MAIpulp, on average, 9 individuals in the 40 best families. The unbalanced optimization was performed through the simulation of 72 scenarios. The effect size (Ne), inbreeding rate (F), and accumulated gain were calculated for each scenario. The selection prioritized the best families and the top ranked individuals within each family. Results revealed that the individual heritability in the restricted sense (h²a) for density (0.331) and pulp yield (0.322) were classified as moderate magnitude. For MAIpulp, the h²a was considered high (0.514), which suggested the presence of genetic control and the possibility of obtain gains through selection. The accuracy for all the traits was higher than 77.3%. In addition, the NIR prediction correlation coefficient presented values above 85% in PLS-OPS for basic density and pulp yield, indicating a high predictive potential of technology for selection of E. benthamii genotypes. In the balanced scenario, 369 individuals were selected with a Ne of 119 and a genetic gain of 36%. In the unbalanced scenario, the scenario 53 was selected, which demonstrated the maintenance of 100 individuals and 36 families is necessary to obtain a Ne equal to 54 and a gain of 64.21% relative to the original population. The estimated genetic parameters indicated favorable conditions for selection. The results suggested the necessity to improve the process through the use of NIR technology. For this process, future research may need to adopt models specific to breeding regions and make alterations in the calibration model for each species. The unbalanced optimization was more efficient than the balanced. Using the unbalanced procedure, researchers may be able to accomplish considerable genetic gains with less individuals in the population while maintaining the same rate of genetic variability.