Seed is an important input component in rice production. Of the many varieties that had been release, the distinction among varieties is not always clear. Among a large number of varieties error may happen in seed processing, storage and distribution, because of the similarity of their physical shape and size, and the seed appearances are difficult to be distinguished. An alternative to distinguish rice seed varieties is using near infrared (NIR) as sensors and using artificial neural network (ANN) as data processor. This research was aimed to study the accuracy of NIR spectroscopy and ANN for detecting rice seed varieties. NIR reflectances (1000-2500 nm) of seeds of 12 varieties were given pretreatment data such as first derivative, second derivative, normalization and standard normal variates. The pretreatment data were used as input in ANN models. Each variety consisted of 12 samples, each sample was 40 grams. ANN model used backpropagation multilayer perceptron with three layers as input, hiden, and output. Network weights were estimated using gradient descent algorithm. The wave form of NIR spectra was similar among varieties, but had different absorptions in intensities, so they could be used for determining the rice seed varieties. The best model was an ANN with standard normal variate pretreatment as input data. The accuracy of varieties prediction was 100% for training, 99.1% for testing and 98.1% for validation. Results showed that the NIR spectra and ANN model can be used as detection methods in rice varieties.