Joviah Tuhaise

Computer Science

Masters thesis was Titled,"Pixel Classification Methods for Automatic Symptom Measurement of Cassava Brown Streak Disease" Abstract The rapid geographical expansion of the Cassava Brown Streak Disease (CBSD) pandemic has devastated cassava crops in East and Central Africa. This has necessitated an upsurge of surveillance efforts for the disease. To monitor CBSD, surveyors deal with single fields of plants at a time. Diagnosis is performed by examining a cross-section cutting of the cassava root tuber. A score of severity of disease is visually assigned to the plant based on the percentage of necrotised root. This method tends to be sub-optimal since scores are highly subjective due to operator variability. This study investigates various computer vision techniques that could be employed to standardise the scoring. Our investigation follows five stages. In stage one images were acquired using mobile devices. In stage two, different techniques are employed to obtain an annotated data set that can be used to train a classifier. Stage three, several classifiers are employed to classify each pixel of the crop images as healthy or necrotised. In stage four, the performance of these classifiers is evaluated based on the Area Under the Curve (AUC), Mean Absolute Error (MAE) and R2 score. Nearest Neighbour classifier performs best with a R2 score of 0.789. To assess operator variability, we compare two sets of predictions from two different surveyors; a confusion matrix is used express the variability in scores assigned.