This paper proposes a compressed Convolutional Neural Network (CNN) model for rice yield detection using the weight pruning technique. The initial CNN model achieved an accuracy of 91% on a dataset comprising 3120 images of both yield and unyield rice crops. However, it had a large size of approximately 603MB, posing challenges in terms of deployment and storage. To address this issue, weight pruning was applied to compress the model. The compressed model achieved a significant reduction in size to 186MB, representing a reduction of approximately 69.15%, while maintaining a reasonable accuracy of 86%.

The experiment was conducted in three phases. First, a dataset of 3120 images of yield and unyield rice crops was collected from different farms in Kano metropolis of Nigeria and preprocessed by resizing them to 250x250 pixels. Secondly, a CNN model with 12 layers was designed and trained using the preprocessed dataset. The model achieved an accuracy of 86%. Finally, weight pruning was applied to the trained CNN model to reduce its size. The compressed model exhibited a size of 300MB and an accuracy of 86%.

The results of this study demonstrate the effectiveness of weight pruning as a viable technique for compressing CNN models without significantly compromising their accuracy. The compressed model, with its reduced size, is well-suited for deployment on resource-constrained devices for rice yield detection applications.