Progressive loss-aware fine-tuning stepwise learning with GAN augmentation for rice plant disease detection

Modern technology like Artificial Intelligence (AI) must be used in the agricultural sector if sustainable agricultural output is to be achieved. One of the most convenient strategies for resolving current and future issues is data-driven agriculture. For this, disease prediction is a major task for precise farming. For predictive analysis and precise agriculture monitoring systems, with the application of AI, Machine Learning (ML) and Deep Learning (DL) play vital roles in building a more robust system. In this work, we will design a DL-integrated rice disease prediction system to be implemented for precise farming. Improvisation of the developed model to detect rice plant diseases & pest attacks with a high level of precision. In this work, the Progressive Loss-Aware Fine-Tuning Stepwise Learning (PLAFTSL) model is proposed for disease detection. For step-wise learning fine-tuned ResNet50 model is used with the introduction of freezing and unfreezing layers. This reduces the training parameters and thus computational complexity. The introduction of the step-wise and progressive loss-aware layer will result in fast convergence and improved training efficiency during information exchange among layers respectively. Our proposed work uses a dataset from two sources. The result analysis is presented with an ablation study. Additionally, the baseline model, ResNet50, is used to display the outcomes of the ablation. The results demonstrate that the fine-tuned model results in better performance as compared to the transfer learning model. The Conditional Generative Adversarial Network (cGAN) augmentation is also added to the designed model which will improve detection effectiveness and can also manage the imbalance in input data. The model has achieved approx. 98% accuracy and outperforms better with comparative state-of-art models. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.