Quantifying and mitigating explanation drift under INT8 quantization in plant disease image classification

Abstract

Abstract Deep learning image classifiers are routinely compressed to INT8 for edge deployment, yet the effect of quantization on post-hoc saliency explanations has not been quantified at scale. We define explanation drift as the systematic change in a saliency map between a full-precision (FP32) model and its fake-quantized INT8 simulation, and measure it with three metrics reported against explicit random-overlap baselines: top- k Intersection-over-Union, top- k Dice, and Spearman rank correlation. We evaluate four explanation methods (Grad-CAM, Grad-CAM++, Integrated Gradients, LIME) on three architectures spanning distinct design families (EfficientNetV2-S, ResNet-50, MobileNetV3-Large) fine-tuned on the Paddy Disease Classification dataset (10 classes, 10,407 images), yielding 3,240 paired saliency comparisons per operating point on a stratified 320-image subset of the validation split. A central methodological finding is that the choice of Grad-CAM target layer, not quantization, governs whether attribution appears to survive: for all three architectures the last module returned by a naive “final layer” heuristic emits a $$1\times 1$$ spatial tensor, which forces a constant saliency map and produces spurious attribution collapse in up to $$61.7\%$$ of images. With architecture-appropriate target layers resolved by an explicit spatial-dimension criterion, no collapse occurs in any of the 3,240 comparisons, and drift is substantial but bounded: mean top- k IoU ranges from 0.366 to 0.790 across architecture–method pairs (4.5–9.7 $$\times $$ the random-overlap baseline of 0.081) with median Spearman $$\rho = 0.850$$ . Drift remains strongly method- and architecture-dependent, with Integrated Gradients least preserved and MobileNetV3-Large most affected among architectures. Calibration choice dominates deployment fidelity: on the full validation split, MinMax calibration in ONNX Runtime reduces FP32–INT8 prediction agreement to 0.644 and 0.641 for EfficientNetV2-S and MobileNetV3-Large, whereas percentile calibration restores it to 0.927 and 0.915. Explanation-aware quantization-aware training with a teacher–student CAM consistency loss improves MobileNetV3-Large (top- k IoU $$+16.3\%$$ ) but degrades EfficientNetV2-S ( $$-8.5\%$$ to $$-26.0\%$$ ) relative to post-training quantization; a $$\lambda = 0$$ control isolates the cause, showing that the direction of these changes is present in standard QAT, while the consistency term improves top- k IoU in all nine architecture–method cells relative to that control ( $$+0.6\%$$ to $$+14.4\%$$ , six significant after Holm correction). Stored full-validation CAM summaries differ from the primary subset by at most 1.4 percentage points in IoU; matching per-image records are unavailable, limiting independent verification. Finally, the auxiliary ground-truth localisation summaries (with incompletely verified execution provenance) for annotated SIIM-ACR chest radiographs report low mean FP32 and INT8 overlap with expert masks, illustrating why explanation similarity under quantization must not be read as explanation quality . Analysis code, figure-generation software, and derived result tables are openly archived.