A Unified Approach with Multi-Scale EfficientDet based Cotton Pest Detection and PPO-Driven Pesticide Recommendation

• A. Mohan,
• P. Chiranjeevi,
• A. Krishna Mohan,

Vol. 19, No. 8, August 31, 2025
10.3837/tiis.2025.08.006, Download Paper (Free):

• Management of Pests and Diseases (MIPD)
• Artificial Intelligence (AI)
• Full Resolution Convolutional Networks (FrCN)
• Multi-Scale EfficientDet (MS-EfficientDet)

Abstract

Agricultural pests inflict tremendous economic and food security losses. Cotton is plagued by severe insect and pest problems, raising production expenses. Proper field pest identification is critical for management and cost-effectiveness. Existing optimization techniques for cotton pest detection possess significant shortcomings, thus necessitating better solutions. This research proposes a hybrid Optimized FrCN for Cotton pest detection and PPO-Driven Pesticide Recommendation. The system uses two real-world datasets: the Cotton Disease Dataset for Detection and the Cotton-Pest and Disease Dataset for Recommendation. Initially, image augmentation is carried out to expand the size of the dataset using GAN. Next, the augmented images are pre-processed using bilateral filtering, Adaptive Histogram Equalization (AHE), and color space transformation methods. After pre-processing, segmentation is carried out using Hybrid Full Resolution Convolutional Networks (H-FrCN) with Binary cross entropy (BCE)-Dice Loss approaches. Once the segmentation is done, texture, shape, color, and deep features are extracted using Local Phase Quantization (LPQ), Gray-Level Difference Statistics (GLDS), and ResNet101 approaches. Once feature extraction is done, the extracted features are fused, and feature selection is done utilizing a hybrid Black Widow Optimization Algorithm (BWOA) and Artificial Gorilla Troop Optimization (AGTO). Next, Multi-Scale EfficientDet (MobileNet, AlexNet, EfficientDet and RetinaNet) is used for pest and disease detection. Finally, Proximal Policy Optimization (PPO) is used for optimal pesticide recommendation. The entire framework is implemented in Python, achieving an accuracy of 99.6% and a 32% reduction in processing time compared with other existing approaches, such as CNN, DCNN, ResNet, and DenseNet models.

Statistics

Cite this article