Rafi, Muhammad Ahsanur (2026) Potato Leaf Disease Classification using Ttransfer Learning Based MobileNetV3Small Architecture. Undergraduate thesis, UPN Veteran Jawa Timur.
![[img]](https://repository.upnjatim.ac.id/style/images/fileicons/text.png) |
Text (COVER)
21081010305.-cover.pdf Download (2MB) |
|
Text (BAB 1)
21081010305.-bab1.pdf Download (1MB) |
|
Text (BAB 2)
21081010305.-bab2.pdf Restricted to Repository staff only until 23 July 2029. Download (10MB) |
|
Text (BAB 3)
21081010305.-bab3.pdf Restricted to Repository staff only until 23 July 2029. Download (12MB) |
|
Text (BAB 4)
21081010305.-bab4.pdf Restricted to Repository staff only until 23 July 2029. Download (14MB) |
|
Text (BAB 5)
21081010305.-bab5.pdf Download (706kB) |
|
Text (DAFTAR PUSTAKA)
21081010305.-daftarpustaka.pdf Download (2MB) |
|
Text (LAMPIRAN)
21081010305.-lampiran.pdf Restricted to Repository staff only Download (1MB) |
Abstract
Leaf disease infections (early and late blight) in potatoes have the potential to reduce harvest quality. To overcome the limitations of manual visual identification, this research develops a deep learning classification model using the efficient MobileNetV3-Small architecture based on transfer learning. Preprocessing leverages a naturally balanced dataset, pixel standardization to [-1.0, 1.0] to match the pre-trained ImageNet weights, and dynamic augmentation (Gaussian blur, noise, lighting variation) to simulate real-world environmental disturbances. The hyperparameter configuration was optimized using Bayesian Optimization. Training was executed in two phases: feature extraction (Adam optimizer) and fine-tuning of the final 40 layers (SGD Nesterov). The test results show that this architecture is efficient, with an internal accuracy reaching 98.15%. However, the cross-dataset (external) evaluation recorded a drop in accuracy to 84.67%, confirming the occurrence of domain shift. Although robust in detecting infected specimens, the model exhibited a hypersensitivity bias toward healthy leaves (Healthy), in which foreign elements of the real environment were misinterpreted as pathology. This research concludes that this lightweight architecture has potential for field diagnostics, but the robustness of its inference absolutely requires enrichment of the physical dataset’s diversity during training in order to suppress false positives caused by external visual variation.
| Item Type: | Thesis (Undergraduate) | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Contributors: |
|
||||||||||||
| Subjects: | Q Science > QA Mathematics > QA76.6 Computer Programming Q Science > QA Mathematics > QA76.87 Neural computers |
||||||||||||
| Divisions: | Faculty of Computer Science > Departemen of Informatics | ||||||||||||
| Depositing User: | Ahsanur Rafi Muhammad | ||||||||||||
| Date Deposited: | 24 Jul 2026 06:59 | ||||||||||||
| Last Modified: | 24 Jul 2026 08:21 | ||||||||||||
| URI: | https://repository.upnjatim.ac.id/id/eprint/57961 |
Actions (login required)
 |
View Item |