SAAIL Lab SAAIL Lab
DSA 2026 · Oral & Poster Published · DSA 2026

Few-Shot Transfer Learning for Cassava Leaf Disease Detection Under East African Field Conditions

Innocent Nyalala 1,2 , Patrick Vincent Ndowo 1 , Shivam Vyas 1 , Twamaa Makame 1 , Surabhi Gudla 1 , Khamis Haji 1
1School of Engineering and Science, Indian Institute of Technology Madras Zanzibar, Tanzania
2Wadhwani School of Data Science and AI (WSDSAI), Indian Institute of Technology Madras, Chennai, India
Data Science Africa (DSA) 2026 Workshop, Makerere University, Kampala, Uganda (Jul 2026)
Few-Shot Transfer Learning for Cassava Leaf Disease Detection Under East African Field Conditions

Summary: Few-shot scaling curves across four pretrained backbones (ResNet-50, EfficientNet-B3, MobileNetV3, ViT-B/16) on 21,397 Ugandan field cassava images at k ∈ {10, 25, 50, 100} shots/class.

Abstract

Cassava feeds over 500 million people across Sub-Saharan Africa, yet annual crop losses of up to 40% from Cassava Brown Streak Disease and Cassava Mosaic Disease often go undetected due to scarce diagnostic infrastructure. We present a systematic evaluation of few-shot transfer learning on 21,397 smartphone-captured Ugandan field images spanning five disease classes. Four pretrained backbones (ResNet-50, EfficientNet-B3, MobileNetV3, ViT-B/16) are evaluated under linear probing and full fine-tuning at k in {10, 25, 50, 100} labeled images per class with a shared hyperparameter configuration. Full fine-tuning of ViT-B/16 reaches 63.6% accuracy at k = 50, a +28.5 pp gain over training from scratch.

Keywords: few-shot learningtransfer learningcassava diseaseEast AfricaViT-B/16DSA 2026

Key Findings & Contributions

68.8%
ViT-B/16 (k=100)
+29.6 pp over scratch (39.2%)
63.6%
ViT-B/16 (k=50)
+28.5 pp at 250 total images
k = 50
Optimal Field Budget
~2 field collection days
21,397
Field Cassava Images
5 disease conditions · Uganda
  1. Evaluated four pretrained backbones (ViT-B/16, EfficientNet-B3, ResNet-50, MobileNetV3) under linear probing vs. full fine-tuning across k in {10, 25, 50, 100} shots/class on 21,397 Ugandan field images.
  2. Demonstrated that full fine-tuning outperforms linear probing by +18 to +20 pp at small k, proving ImageNet features require end-to-end spatial adaptation for field cassava symptoms.
  3. Identified Healthy leaf confusion (F1: 0.49, 25% misclassified as CBB) vs. CMD (F1: 0.77) and established k = 50 labels/class (250 images total, +0.053 pp/image marginal ROI) as the optimal collection budget.

Key Empirical Figure

ViT-B/16 linear probe versus full fine-tuning accuracy scaling across k in {10, 25, 50, 100} shots per class on iCassava Uganda, highlighting the +20.3 pp adaptation gain at optimal k = 50.
ViT-B/16 linear probe versus full fine-tuning accuracy scaling across k in {10, 25, 50, 100} shots per class on iCassava Uganda, highlighting the +20.3 pp adaptation gain at optimal k = 50.

Benchmark Results

Few-Shot Top-1 Accuracy (%) on Ugandan Field Cassava (21,397-Image Pool)

Method & Adaptationk = 10k = 25k = 50k = 100
ViT-B/16 (Full Fine-Tuning)50.658.063.668.8
ResNet-50 (Full Fine-Tuning)45.558.062.368.6
EfficientNet-B3 (Full Fine-Tuning)40.552.763.568.4
MobileNetV3 (Full Fine-Tuning)41.356.662.266.3
ViT-B/16 (Linear Probe only)30.032.943.350.3
ResNet-50 (Linear Probe only)28.934.145.249.8
Scratch Baseline (No Pretraining)26.932.135.139.2

BibTeX

@inproceedings{nyalala2026fewshot,
  title={Few-Shot Transfer Learning for Cassava Leaf Disease Detection Under East African Field Conditions},
  author={Nyalala, Innocent and Vincent, Patrick and Vyas, Shivam and Makame, Twamaa and Gudla, Surabhi and Haji, Khamis},
  booktitle={Data Science Africa (DSA) 2026 Workshop},
  address={Makerere University, Kampala, Uganda},
  year={2026},
  month={July}
}
APA Citation

Nyalala, I., Vincent, P., Vyas, S., Makame, T., Gudla, S., & Haji, K. (2026, July). Few-Shot Transfer Learning for Cassava Leaf Disease Detection Under East African Field Conditions. Data Science Africa (DSA) 2026 Workshop, Makerere University, Kampala, Uganda.