Start simple, measure everything
Deterministic splits, a frozen MobileNetV2 head, logging from day one: 0.9613 accuracy and 0.9501 F1 on PlantVillage.
0.96Lab accuracy
0.95Lab F1
From lab leaves to field photos
Abstract
A crop disease classifier covering 38 classes, trained on clean lab images, collapses on messy field photos: 0.95 F1 in the lab against 0.11 in the field. Twelve sprints of ablations across mixed domain training, backbone sweeps, routing and segmentation converged on one ResNet-50 recipe that holds 0.95 lab F1 with 0.65 field F1.
0.00Field F1 (PlantDoc)
0.00Lab F1 (PlantVillage)
0Disease classes
0Training sprints
Data
Two datasets, two worlds
Method
Twelve sprints, one rule
Every claim ships with a number. Checkpoints only survive if they beat what came before on both worlds.
S0 + S1
Baseline
S0 + S1
Baseline
S3
Robustness
S3
Robustness
Augment train, never test
Flips, rotation, brightness, blur and JPEG noise on the train set only; a run ships only if it beats the baseline.
S4
The field gap
S4
The field gap
Lab accuracy collapses outdoors
The baseline managed 0.11 F1 on PlantDoc. Fine tuning recovered 0.51 but forgot PlantVillage, dropping it to 0.31. The gap became the thesis.
0.11Field F1
0.31Lab F1
S5 + S6
Mixed training
S5 + S6
Mixed training
Concatenate domains, keep both
Joint PlantVillage + PlantDoc batches held 0.96 lab with 0.41 field and no forgetting. Eightfold field upsampling added nothing.
0.96Lab F1
0.41Field F1
S7
Capacity
S7
Capacity
Bigger backbones trade, never win
ResNet-50 cleared 0.65 in the field but forgot the lab at 0.29. No single head passed both gates, so the architecture had to change.
0.65Field F1
0.29Lab F1
S8 + S9 + S11
The routing detour
S8 + S9 + S11
The routing detour
A documented negative result
Two heads, two backbones, a learned router at 74%: routed scoring reached 0.39 against the single model's 0.63. Routing was shelved.
74%Router accuracy
0.39Routed score
S10
Segmentation
S10
Segmentation
Cutting leaves out hurts
GrabCut preprocessing dropped field F1 from 0.41 to 0.27. The same sweep found the real lever: backbone learning rate with repeat 20 sampling.
0.27Field F1
S12
The final recipe
S12
The final recipe
Cheap levers, then ship
Test time augmentation did nothing. Class balanced sampling cleared every gate: 0.661 field accuracy, 0.654 field F1, lab intact.
Shipped to production0.66Field accuracy
0.65Field F1
| Checkpoint | Lab acc | Field acc | Field F1 |
|---|---|---|---|
| Stage 1 baseline | 0.96 | 0.14 | 0.11 |
| Mixed training | 0.96 | 0.41 | 0.41 |
| Repeat 20 sweep | 0.99 | 0.63 | 0.56 |
| Shipped · class balanced | 0.96 | 0.66 | 0.65 |
Held out test splits, seed 42. Accuracy counts the top guess; F1 is macro.
Everything behind these numbers
Ablation logs, confusion matrices, plots and the shipped weights live in the results folder.
