Chickpea Fusarium Wilt Detector
Hybrid lightweight CNN for on-device plant disease detection — MSc Thesis, Batangas State University (2024–2026)
| MSc Thesis Research | Batangas State University | 2024–2026 |
Published Paper: A Systematic Ablation of CBAM Depth in SqueezeNet for Optimising Cross-Domain Chickpea Disease Detection, Procedia Computer Science, Vol. 289, pp. 1259–1266, 2026 (ICCSCI 2026). DOI: 10.1016/j.procs.2026.08.346 (Lwin et al., 2026)
What It Does
A production-ready Android application that detects Chickpea Fusarium Wilt from smartphone photos — fully offline, instant results, with bilingual disease advisory (English + Myanmar Unicode). Designed for smallholder farmers in Myanmar with no internet connectivity.
Key Results
- 82.09% cross-domain accuracy on a real-world Myanmar field test set unseen during training
- Model size: 0.870 MB after quantization — compatible with 2 GB RAM devices and Android 9.0+
- Validated on 115 real physical Android devices via Google AI Edge Portal (private preview)
- Worst-case inference: 73.7 ms on low-tier CPU-only hardware
What I Built
- A novel hybrid CNN architecture combining SqueezeNet v1.1 with an attention mechanism — outperforming larger and more established architectures with a fraction of the parameters
- An INT8 quantization pipeline delivering over 3× model compression with zero accuracy degradation
- A personally collected field dataset from Myanmar farming communities, with expert agronomist annotation
- A bilingual Android app (Kotlin, Jetpack Compose, Material Design 3, TFLite) supporting SDG 1 and SDG 2
Why It Matters
Most plant disease AI models are designed for high-end hardware or require cloud connectivity. This work targets the opposite end: low-cost devices, no internet, field conditions — making AI-assisted disease detection accessible to farmers who need it most.
Tech Stack: PyTorch · TFLite/LiteRT · Kotlin · Jetpack Compose · Android Studio · Roboflow · Kaggle GPU · Google AI Edge Portal