A FlutterFlow Prototype Mobile System Approach for Apple Leaf Disease Classification using a Hybrid Segmentation and Region-based CNN Detection
Year: 2026
Type: conference-paper
Abstract: Apple leaf diseases significantly affect crop yield and food security, underscoring the need for accessible and real-time diagnostic tools. This study proposes an integrated end-to-end prototype system that combines a custom Convolutional Neural Network (CNN) classifier trained on the PlantVillage apple leaf dataset with a cloud-based API and a FlutterFlow-based mobile application for automated disease classification. The methodology involves training a hybrid segmentation and region-based detection pipeline, deploying it via a FastAPI server hosted on Hugging Face Spaces, and integrating it into a mobile interface developed using FlutterFlow. The pipeline incorporates HSV-based image segmentation and contour-based region detection to isolate leaf areas of interest prior to classification, complemented by a 23-feature handcrafted descriptor extraction module spanning color statistics, Gray-Level Co-occurrence Matrix (GLCM) texture properties, and morphological shape attributes. End-to-end system testing was conducted within the FlutterFlow development environment to evaluate functionality, response time, and prediction consistency. The CNN model achieved 92.00% accuracy in isolation, with a weighted precision of 92.72%, recall of 92.00%, and F1-score of 91.65% on a four-class held-out test set comprising Apple_healthy, Apple_scab, Black_rot, and Cedar_apple_rust. However, end-to-end evaluation of the full hybrid pipeline — combining HSV segmentation, contour-based region detection, and CNN inference — yielded a lower accuracy of 71.43%, a gap attributable to a training–inference distributional mismatch that is analyzed in detail and identified as the primary limitation of this study. System testing confirmed that the mobile prototype successfully delivers annotated detection outputs with bounding box visualization and per-class confidence scores, with response behavior that was subjectively prompt during testing; quantitative latency, throughput, and localization (IoU/mAP) benchmarking were not performed and are identified as necessary future work. The findings demonstrate that the proposed framework provides a functional prototype with a clearly documented performance ceiling, indicating potential for future deployment as a decision-support tool for precision agriculture pending the improvements identified herein.
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: hybrid
OpenAlex ID: W7224929019