Thumbnail Image

Thesis

Browse

20231
Reset filters
Subject: Soil and Water Engineering × Author: Goyal, Pooja × Start date: 2023 ×
Reset filters

Search Results

Now showing 1 - 1 of 1
  • Thumbnail Image
    ThesisItemRestricted
    Developing Artificial Intelligence Algorithm for Identification and Classification of Abiotic Stresses in Maize
    (Punjab Agricultural University, 2023) Goyal, Pooja; Sharda, Rakesh
    Cereal crops like maize are seriously affected by the abiotic plant stresses, which cause substantial reduction in crop yields. New age Artificial Intelligence (AI) and Deep Learning (DL) techniques offer a non-invasive and a practical way for precise detection of these stresses in real time. The present study was planned to develop an AI-based algorithm for identification and classification of abiotic stresses in maize. Three consecutive experimental trials i.e. spring maize (2021), kharif maize (2021) and spring maize (2022) were conducted at the Research Farm of Department of Soil and Water Engineering, PAU, Ludhiana in a split-plot design with varieties in main plot and fertigation treatments in sub-plot. The subplots covered three types of nutritional deficiencies of nitrogen (N), phosphorus (P) and potassium (K), one water deficit plot and one plot with 100% irrigation and fertigation. Data regarding the plant growth and yield attributes were observed at different growth stages and a total of 6035 RGB images were captured throughout the crop growing seasons. These images were classified into five classes according to stress levels to build a DL-based Convolutional Neural Network (CNN) model. Two classification models were built, the first was a binary classification model for drought stress detection, and the second was a multi-class classification model to classify all the abiotic stresses. Further, 2 modelling approaches i.e. custom-CNN model and a transfer learning (TL) approach using 5 state-of-the-art architectures were used. Both ResNet50 and EfficientNet121 models performed the best, achieving a test accuracy and F1-score of 99.26% and 99.22% respectively in case of binary classification model. The ResNet50 model also achieved the highest average weighted accuracy and F1-score of 98.36% and 98.35% respectively for overall abiotic stress detection in maize. The custom-CNN model, with much lesser number of parameters compared to the transfer-learned networks, achieved an acceptable F1-score of 98.44% and 97.04% for drought and abiotic stress classification respectively. Therefore, the custom-CNN model was recommended for real-time abiotic stress assessment in maize crop especially on resource constrained devices.