Paddy pest ID now ships with a reason and a tamper-proof log: 97.83% accuracy, 34.2% less pesticide
Synopsis
AgriMind-PestGuard couples three parts for paddy pest management: a multi-scale attention CNN that identifies 14 major pest species, a Grad-CAM-HCA module that grounds each decision in entomological concepts, and a Hyperledger Fabric layer that writes detections, pesticide recommendations and compliance records into a tamper-proof ledger. On a curated 58,400-image field dataset it reaches 97.83% classification accuracy, above the strongest baseline EfficientNetV2-S at 96.38%; explanation faithfulness is 0.91 and three entomologists agree with 93.7% of concept attributions. Across 12 paddy farms, field trials report 34.2% less pesticide use and 8.7% higher yield than conventional practice.
Interpretation
Paddy pest identification now returns a class label together with a checkable justification and a tamper-proof record, not just a prediction. Earlier deep-learning pest systems pursued classification accuracy alone, left decisions opaque, and stored monitoring data in centralized databases open to tampering. On the curated AMPG-58K test set the full MSA-CNN reaches 97.83% accuracy, 1.45 points above the strongest baseline EfficientNetV2-S at 96.38%, with McNemar's test p<0.001.
The recognition network separates discriminative features of differently sized pests through multi-scale fusion and dual channel-spatial attention. A standard feature pyramid aggregates multi-scale features with simple lateral connections and does not weight which scales matter for pest discrimination. Ablation shows the pest-aware feature pyramid adds +1.29% accuracy alone, channel-spatial dual attention adds +1.59%, their combination +1.91%, and the composite loss brings the total gain to +2.21%.
The explanation module maps pixel-level saliency onto entomological concepts, so farmers and entomologists can see whether wing venation, body shape or antennae drove the call. Earlier explainability methods stayed at low-level pixel attribution without grounding in entomological concepts. On 500 randomly sampled test images, Grad-CAM-HCA scores 0.91 faithfulness and 93.7% concept agreement, above Score-CAM at 0.83 and 85.1%.
The blockchain layer writes each detection, pesticide recommendation and compliance record as a signed ledger entry that regulators can verify individually. Earlier agricultural blockchain work stayed at supply-chain traceability or conceptual frameworks, without practical integration and evaluation alongside an AI pest detection system. The Hyperledger Fabric v2.4 network reaches 1,247 TPS throughput and 0.83 s average latency at 200 clients and a 1,500 TPS send rate, with peak throughput of 1,312 TPS at a 2,000 TPS send rate and 3.05 MB storage per 1,000 records.
Perspective
The system is trained and tested on 14 major paddy pest species across 58,400 field images, with 2023 Kharif-season field trials on 12 farms in three agro-climatic zones covering tropical, subtropical and temperate conditions. Cross-dataset evaluation uses the eight rice-related classes of IP102 and Dhan-Pest, showing the model keeps its lead after a dataset change. Deployment figures span server A100 down to Jetson Nano, Raspberry Pi with Coral TPU and phone NNAPI, with total per-image time from 62.8 ms to 376.9 ms on edge devices.
Coverage is limited to 14 major paddy pest species, with expansion to beneficial insects and disease vectors still planned. Accuracy drops by about 5% under extreme lighting such as night-time or heavy rain. The concept attribution module depends on expert-annotated concept masks, which are labor-intensive, and weakly supervised concept discovery is still being explored. High-resolution images are impractical to store on-chain long term, so only hashes go on-chain while originals sit in off-chain IPFS. Blockchain throughput begins to degrade beyond a 2,000 TPS send rate.