Explainable deep learning framework for real-time insect pest detection in smart agriculture
Open Access
- 12 October 2026
- journal article
- research article
- Published by Springer Nature in Discover Artificial Intelligence
- Vol. 6 (1) , 1435
Abstract
Global losses of crop production due to insect pest infestations are estimated to be around 20–40% annually, which represents one of the highest threats to agricultural productivity, farmer livelihoods and food security. Current manual scouting is labor intensive, subjective and non-scalable, and many automated scouting methods suffer when they come to deal with different field conditions and offer little insight into decision making. To tackle these challenges, this study proposes Crop Protection in Smart Agriculture—Real-time Insect Pest Detection using Continual Spatio-Temporal Graph Convolutional Network (CPSA-RIPD-CSTGCN), a unified and interpretable deep learning framework that combines the four synergistic components: (i) the Ensemble Transform Schmidt–Kalman Filter (ETSKF) for image preprocessing, including resizing, normalization, and noise suppression; (ii) data augmentation using rotation, flipping, scaling, shearing, and contrast adjustment, which enhances generalization and reduces overfitting; (iii) the Continual Spatio-Temporal Graph Convolutional Network (CSTGCN) that jointly models the spatial dependency across crop regions and temporal evolution of pest dynamics to achieve robust feature extraction and classification; and (iv) the Revolution Optimization Algorithm (ROA) for hyperparameter optimization of CSTGCN weights to adaptively adjust the network. Further, Gradient-weighted Class Activation Mapping (Grad-CAM) visualization is introduced to give post-hoc explainability as it shows regions that are relevant to the biological process and areas that are affected by pests. This framework was tested on IP102 benchmark dataset with 60%-40%-split train-test with 10 economically important classes of pests with 97% accuracy, 98% precision, 98% recall and 98% F1 score, mean ± 0.32% accuracy, across 5-fold cross validation. Among other baseline models, these results consistently outperform VGG16 (82%), InceptionV3 (85%), MobileNetV2 (87%), DenseNet121 (88%), EfficientNetB0 (89%) and task-specific models like EDCP-DIA-RNN(Early Disease and Crop Pest—Diagnosis/Detection using Image Analysis RNN) (94%), EPCY-TPA-ANN (Efficient Prediction of Crop Yield—Through Parameter Analysis ANN) (91%) and ICPM-ESA-CNN (Interactive Crop and Pest Management Enhanced Soil Analysis CNN) (90%). The results of ablation studies showed the solitary effects of each module of the Ensemble Transform Schmidt–Kalman Filter pre-processing and Revolution Optimization Algorithm optimization, and the confusion matrix analysis demonstrated high per-class reliability and only a small amount of misclassification, implying that the proposed method could be applied to real-time, large-scale smart agriculture deployments.Keywords
This publication has 38 references indexed in Scilit:
- High‐Performance Deep Learning for Instant Pest and Disease Detection in Precision AgricultureFood Science & Nutrition, 2025
- An interactive AI-based crop and pest management system leveraging transfer learning for enhanced sustainable agricultureModeling Earth Systems and Environment, 2025
- The ensemble transform Schmidt–Kalman filter: A novel method to compensate for observation uncertainty due to unresolved scalesAtmospheric Science Letters, 2025
- GeoAgriGuard: AI-Driven Pest and Disease Management with Remote Sensing for Global Food SecurityRemote Sensing in Earth Systems Sciences, 2025
- Enhancement of tea leaf diseases identification using modified SOTA modelsNeural Computing and Applications, 2024
- An intellectual model of pest detection and classification using enhanced optimization-assisted single shot detector and graph attention networkEvolutionary Intelligence, 2024
- Optimized recurrent neural network-based early diagnosis of crop pest and diseases in agricultureDiscover Computing, 2024
- Boosting enabled efficient machine learning technique for accurate prediction of crop yield towards precision agricultureDiscover Sustainability, 2024
- Deep Multibranch Fusion Residual Network and IoT-based pest detection system using sound analytics in large agricultural fieldMultimedia Tools and Applications, 2023
- Comparative assessment of Pest damage identification of coconut plant using damage texture and color analysisMultimedia Tools and Applications, 2023