The Latent Semantic Indexing (LSI) method was used for the automatic retrieval of similar images within a database containing 232 water-related landscape images. The principle of LSI is the images dimensionality reduction and the computation of cosine similarity between a query image and other images from the database. The optimal image dimensionality reduction parameter k was determined by means of a scree plot, which is used in principal component analysis. Using k = 8, the photographs displaying similar-looking objects were found. Presented results indicate that low or too high k values can cause a loss of useful information or on the contrary a redundancy of information noise in the analysed images, respectively.