Rice seed varietal purity inspection using hyperspectral imaging

Hai, Vu and Tachtatzis, Christos and Murray, Paul and Harle, David and Dao, Trung Kien and Le, Thi-Lan and Andonovic, Ivan and Marshall, Stephen (2016) Rice seed varietal purity inspection using hyperspectral imaging. In: Hyperspectral Imaging and Applications Conference, 2016-10-12 - 2016-10-13, Ricoh Arena.

[thumbnail of Vu-etal-HSI2016-Rice-seed-varietal-purity-inspection-using-hyperspectral-imaging]
Preview
Text. Filename: Vu_etal_HSI2016_Rice_seed_varietal_purity_inspection_using_hyperspectral_imaging.pdf
Accepted Author Manuscript

Download (1MB)| Preview

Abstract

When distributing rice seed to farmers, suppliers strive to ensure that all seeds delivered belong to the species that was ordered and that the batch is not contaminated by unhealthy seeds or seeds of a different species. A conventional method to inspect the varietal purity of rice seeds is based on manually selecting random samples of rice seed from a batch and evaluating the physical grain properties through a process of human visual inspection. This is a tedious, laborious, time consuming and extremely inefficient task where only a very small subset of the entire batch of the rice seed can be examined. There is, therefore, a need to automate this process to make it repeatable and more efficient while allowing a larger sample of rice seeds from any batch to be analysed. This paper presents an automatic rice seed inspection method which combines hyperspectral imaging and tools from machine learning to automatically detect seeds which are erroneously contained within a batch when they actually belong to a completely different species. Image data from Near-infrared (NIR) and Visible Light (VIS) hyperspectral cameras are acquired for six common rice seed varieties. Two different classifiers are applied to the data: a Support Vector Machine (SVM) and a Random Forest (RF), where each consists of six one-versus-rest binary classifiers. The results show that combining spectral and shape-based features derived from the rice seeds results in an increase in the precision (PPV) of the multi-label classification to 84% compared with 74% when only visual features are used.