MSIT Theses and Dissertations (2021)

Permanent URI for this collection

http://hdl.handle.net/11071/12748

Browse

Browsing MSIT Theses and Dissertations (2021) by Subject "Convolutional Neural Networks (CNNs)"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    A Computer vision-based model for crop yield prediction using remote sensing data
    (Strathmore University, 2021) Kiragu, Daniel Mburu
    Arguably, crop yield data forms the most important measure of crop productivity in agriculture. With adequate crop yield data, local and international bodies can develop effective agricultural policy leading up to sustainable food supplies and elevated food security. However, timely acquisition of crop yield data can be a cumbersome task as existing crop yield prediction approaches face numerous challenges. In this study, these challenges are identified as high cost and high dimensionality of data required for the prediction activities as well as limited scaling of the resultant prediction models. In efforts of overcoming these challenges, this study leveraged an alternative source of data to design and develop a cheap, accurate and scalable deep learning model using convolutional neural networks. Satellite imagery datasets were used as the primary and only source of data for training the model. This benefited the study in two major ways. Firstly, off, the approach automatically took care of the high dimensionality problem as demonstrated in the GEMS data. Second, satellite imagery data is readily available globally, a factor that greatly reduced the costs needed to collect real-time data for the study. Validation of the developed model was done using 10% of the overall dataset acquired. Reliability of the model in performing crop yield predictions was captured using an MSE loss function for each epoch trained. Cumulatively, the model achieved an MSE loss score of 3.6.
  • Thumbnail Image
    Item
    A Deep normalized neural network model for strawberry fungal leaf disease detection
    (Strathmore University, 2021) Kerre, Deperias Webula
    Strawberry is one of the cash crops that are being grown in Kenya for both export and local consumption. However, strawberry fungal leaf diseases are threatening the existence of this crop which is an important input in the agricultural production sector. The types of strawberry fungal leaf diseases resulting to greater losses in production include Strawberry Leaf scorch, Strawberry Leaf Spot and Strawberry Leaf Blight. The biggest challenge the farmers face is that of correctly classifying these diseases based on observable leaf features. Famers have incurred losses due to poor/incorrect control measures which results from the misdiagnosis of these diseases. This scenario is more pronounced in rural settings where the farmers have a limited access to expertise in modern agricultural production. As a result of this, automated classification of strawberry plant fungal leaf diseases is highly desired. The literature review found several computer vision techniques that have been leveraged in Strawberry fungal leaf disease detection. Among these solutions are the convolutional neural network-based models. Despite the high detection accuracy, the models do not cover another of strawberry fungal leaf diseases such as leaf spot and fail to generalize well on unseen data. The models also do not consider cases where more than one disease occur on the same part of the plant, in this case the leaf. In this study, a deep learning model was proposed for classifying fungal leaf diseases in strawberry based on an experimental research design. The model generalized well on previously unseen data and considered a scenario where multiple diseases occur on the same leaf (Leaf Scorch and Leaf Blight). The model also covered strawberry Leaf Spot that was not covered by any of the existing deep learning models. Data samples containing a total of 1,134 leaf images, categorized into five classes including healthy leaf images were split into 80% training and 20% validation. The disease classes include strawberry leaf spot, leaf scorch, leaf blight and a class where two diseases (Leaf Blight and leaf Scorch) occur together. The model was trained on 30 epochs from scratch with batch normalization implemented within the convolutions in Keras framework and validated using a confusion matrix. The model achieved an outstanding classification accuracy of 98%, precision of 97% , recall of 95.7% and an F1-score of 96.3%.