Table 6 Ablation study isolating the effect of contrastive pretraining using identical ConvNeXt-Tiny architecture and supervised training schedule.

From: PlantCLR: contrastive self-supervised pretraining for generalizable plant disease detection

Dataset

Variant

Acc. (%)

Macro-F1 (%)

PlantVillage

V1: Supervised only

98.21

98.05

V2: +SimCLR (PlantCLR)

99.10

99.04

Cassava

V1: Supervised only

93.74

93.52

V2: +SimCLR (PlantCLR)

96.83

96.70

  1. PV, PlantVillage (in-domain); CS, Cassava (cross-dataset adaptation).
  2. Significant values are in bold.
Back to article page