Enhancing Rice Nutrient Deficiency Classification using Capsnet with
Contextual Attention Routing
Abstract
Rice, being a widely grown food crop, often suffers from nutrient
imbalances, hindering its production, which suffers from a lack of
certain nutrients like potassium, nitrogen, and phosphorus. Identifying
the potential nutrient deficiency in the rice crop can pose a
challenging task because the leaves of rice plants affected by these
nutritional shortages may have disorders such as colour and shape
changes. The classification of nutrient deficits can be done visually by
examining the colour and shape of the leaves. However, it is expensive,
time-consuming, and requires greater expertise. To tackle this issue,
this work proposes a computer vision-based deep learning approach called
CAR-Capsnet, an enhanced version of the capsule network (Capsnet) that
utilises contextual attention routing (CAR) for classifying nutrient
deficiencies in rice crops. The dataset required for training and
evaluation is obtained from the freely available Kaggle data repository.
Pre-processing techniques, such as the Wiener filter and adaptive Otsu
segmentation, are applied to the dataset. Unlike traditional routing
approaches, CAR-Capsnet utilises contextual attention routing to improve
the model’s capability to handle complex visual features and patterns
associated with nutrient deficiencies in rice crops. In the evaluation,
the proposed model’s performance is compared with two other existing
models, CNN and the original Capsnet. Experimental results demonstrate
that CAR-Capsnet outperforms both models, achieving a testing accuracy
of 97.1% and exhibiting superior performance compared to the baseline
models. Additionally, comparing the classifier’s results with previous
research findings confirms that the proposed CAR-Capsnet model is
superior to alternative approaches in classifying nutrient deficiencies
in rice crops.