Adaptive Gated DenseNet121 for Paddy Disease Classification: Performance Gains and Limits of Sparsity-Based Pruning

Abstract

Abstract DenseNet121 is effective for rice disease recognition, but its dense feature reuse keeps many connections active during inference. We evaluated whether learnable layer gates could improve paddy disease classification and expose removable DenseNet connections. Using the Kaggle Paddy Disease Classification dataset, we trained an ImageNet-pretrained DenseNet121 baseline and an Adaptive Gated DenseNet121 on 8,325 training images and validated on 2,082 images across ten classes. The baseline achieved 96.78% accuracy and 96.64% macro F1-score. The adaptive model achieved 97.21% accuracy and 96.89% macro F1-score, with 97.16% macro precision and 96.65% macro recall. Pruning at threshold 0.825 reduced parameters from 6.96 million to 6.84 million while retaining 96.39% F1-score, but slightly higher thresholds caused severe collapse. Adaptive gating therefore improved classification and enabled limited pruning, but did not yet provide robust mobile compression.

Code and data availability

The paper uses the public Kaggle Paddy Disease Classification image dataset for all phenotyping measurements and states that the authors' PyTorch implementation, Kaggle notebooks, and report-generation scripts are publicly available in a GitHub repository. Both are paper-specific, public, and actionable.