DeepLeafAnalyzer A Hybrid Deep Learning Framework for ROI Optimized Multi Scale Leaf Disease Detection

Early, accurate diagnosis of crop leaf diseases is essential for effective disease management and sustainable agricultural production. While deep learning algorithms perform well in laboratory settings, where they are often exposed to known, consistent lighting and imaging conditions, much of this performance can break down in field settings due to lighting variations, background noise, occlusion, symptom variability by size, and a lack of interpretability. To address these problems, this paper introduces a Hybrid Deep Learning Framework (HybridLeafNetAnalyzer) that integrates the HybridLeafDiseaseNet for region-aware, multi-scale leaf disease classification. The proposed framework uses a dual-stage Region of Interest (ROI) optimisation technique: first, U-Net segmentation removes irrelevant background information; second, YOLO localisation identifies the disease-relevant area for further classification. Moreover, it adopts an LSTM-based feature-sequence encoder for modelling the ordered spatial representations, a Feature Pyramid Network (FPN) and Atrous Spatial Pyramid Pooling (ASPP) for multi-scale representation, and Vision Transformer (ViT)-style self-attention and a Channel-Spatial Attention Module (CSAM) for feature refinement contextually. This framework was tested on two imaging datasets: PlantVillage and FieldPlant, which were collected under controlled and field conditions, respectively, and AI Challenger. Under the reported experimental protocol, DeepLeafAnalyzer outperforms the evaluated baseline models (VGG16, ResNet50, MobileNetV2, and EfficientNet-B0), achieving 98.42% accuracy, 98.55% precision, 98.31% recall, and an F1 score of 98.43%. The accuracy drop when ablating ROI optimisation was the highest, at 4.56% points. Representative Grad-CAM visualisations supported qualitative interpretability and generally corresponded to visually disease-relevant regions, suggesting that predictions often aligned with visually relevant disease regions. We did not use confidence-based categorisation as a measure of disease severity, but included it as an auxiliary indicator of predictive certainty. In summary, DeepLeafAnalyzer is an interpretable, region-aware framework for leaf disease classification under the various imaging conditions considered. The implementation code supporting this study is publicly available at https://github.com/santoshi148/DeepLeafAnalyzer .

Authors

Publication Details

Journal
International Journal of Computational Intelligence Systems
Published
2026-11-21
DOI
https://doi.org/10.1007/s44196-026-01691-4
Primary Topic
Smart Agriculture and AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

DeepLeafAnalyzer A Hybrid Deep Learning Framework for ROI Optimized Multi Scale Leaf Disease Detection

Suresh Kumar Mandala, Santoshi Chinta, Raghavendar Raju Lingala
International Journal of Computational Intelligence Systems
Smart Agriculture and AI
article

DeepLeafAnalyzer A Hybrid Deep Learning Framework for ROI Optimized Multi Scale Leaf Disease Detection

Suresh Kumar Mandala, Santoshi Chinta, Raghavendar Raju Lingala
article en

Abstract

Early, accurate diagnosis of crop leaf diseases is essential for effective disease management and sustainable agricultural production. While deep learning algorithms perform well in laboratory settings, where they are often exposed to known, consistent lighting and imaging conditions, much of this performance can break down in field settings due to lighting variations, background noise, occlusion, symptom variability by size, and a lack of interpretability. To address these problems, this paper introduces a Hybrid Deep Learning Framework (HybridLeafNetAnalyzer) that integrates the HybridLeafDiseaseNet for region-aware, multi-scale leaf disease classification. The proposed framework uses a dual-stage Region of Interest (ROI) optimisation technique: first, U-Net segmentation removes irrelevant background information; second, YOLO localisation identifies the disease-relevant area for further classification. Moreover, it adopts an LSTM-based feature-sequence encoder for modelling the ordered spatial representations, a Feature Pyramid Network (FPN) and Atrous Spatial Pyramid Pooling (ASPP) for multi-scale representation, and Vision Transformer (ViT)-style self-attention and a Channel-Spatial Attention Module (CSAM) for feature refinement contextually. This framework was tested on two imaging datasets: PlantVillage and FieldPlant, which were collected under controlled and field conditions, respectively, and AI Challenger. Under the reported experimental protocol, DeepLeafAnalyzer outperforms the evaluated baseline models (VGG16, ResNet50, MobileNetV2, and EfficientNet-B0), achieving 98.42% accuracy, 98.55% precision, 98.31% recall, and an F1 score of 98.43%. The accuracy drop when ablating ROI optimisation was the highest, at 4.56% points. Representative Grad-CAM visualisations supported qualitative interpretability and generally corresponded to visually disease-relevant regions, suggesting that predictions often aligned with visually relevant disease regions. We did not use confidence-based categorisation as a measure of disease severity, but included it as an auxiliary indicator of predictive certainty. In summary, DeepLeafAnalyzer is an interpretable, region-aware framework for leaf disease classification under the various imaging conditions considered. The implementation code supporting this study is publicly available at https://github.com/santoshi148/DeepLeafAnalyzer .

International Journal of Computational Intelligence Systems
Smart Agriculture and AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon