Bayesian optimized deep learning and ensemble classification approach for multiclass plant disease identification
Abstract Early detection of plant diseases is crucial in smart agriculture to prevent significant crop losses and reduce reliance on chemical pesticides. While many existing methods leverage customized neural network architectures or transfer learning, they often suffer from limited accuracy, model...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-07-01
|
| Series: | Discover Sustainability |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s43621-025-01648-1 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Early detection of plant diseases is crucial in smart agriculture to prevent significant crop losses and reduce reliance on chemical pesticides. While many existing methods leverage customized neural network architectures or transfer learning, they often suffer from limited accuracy, model interpretability, and dataset imbalance. This study proposes a novel hybrid model that integrates a pre-trained Inception v3 architecture for feature extraction with a Bayesian-optimized activation strategy and a fine-tuned Random Forest classifier for final decision-making. The architecture involves freezing specific layers within Inception v3 to retain essential low-level features while adapting high-level features for the target domain. Bayesian optimization is used to identify and combine optimal activation functions, enhancing the network's capacity to learn complex disease patterns from tomato leaf images. The extracted features are then classified using a Random Forest model that is tuned for performance on imbalanced data using class weights and key hyperparameters. The proposed approach was evaluated on a multi-class dataset comprising 10 categories, including 9 diseased and 1 healthy class of tomato leaves from the PlantVillage dataset. Results show that the model achieves a classification accuracy of 99.6%, representing a + 5.4% improvement over standard transfer learning techniques. The integration of Bayesian-optimized activations and a tuned Random Forest enhances both precision and robustness across classes, including minority ones. This design not only improves accuracy but also supports a more balanced and interpretable classification process. The model demonstrates significant potential for real-world agricultural applications, particularly where early and accurate disease detection is essential for timely intervention and improved crop management. |
|---|---|
| ISSN: | 2662-9984 |