We present a systematic approach for salinity sensor placement in a polder network, where the objective is to estimate the unmeasured salinity levels in the main polder channels. We formulate this problem as optimization of the estimated salinity levels using root mean square error (RMSE) as the "goodness of fit" measure. Starting from a hydrodynamic and salt transport model of the Lissertocht catchment (a low-lying polder in the Netherlands), we use principal component analysis (PCA) to produce a low-order PCA model of the salinity distribution in the catchment. This model captures most of the relevant salinity dynamics and is capable of reconstructing the spatial and temporal salinity variation of the catchment. Just using three principal components (explaining 93% of the variance of the dataset) for the low-order PCA model, three optimally placed sensors with a greedy algorithm make the placement robust for modeling and measurement errors. The performance of the sensor placement for salinity reconstruction is evaluated against the detailed hydrodynamic and salt transport model and is shown to be close to the global optimum found by an exhaustive search with a RMSE of 82.2 mg/L.