A robust and light-weight transfer learning-based architecture for accurate detection of leaf diseases across multiple plants using less amount of images.

Abstract

Leaf diseases are a global threat to crop production and food preservation. Detecting these diseases is crucial for effective management. We introduce LeafDoc-Net, a robust, lightweight transfer-learning architecture for accurately detecting leaf diseases across multiple plant species, even with limited image data. Our approach concatenates two pre-trained image classification deep learning-based models, DenseNet121 and MobileNetV2. We enhance DenseNet121 with an attention-based transition mechanism and global average pooling layers, while MobileNetV2 benefits from adding an attention module and global average pooling layers. We deepen the architecture with extra-dense layers featuring swish activation and batch normalization layers, resulting in a more robust and accurate model for diagnosing leaf-related plant diseases. LeafDoc-Net is evaluated on two distinct datasets, focused on cassava and wheat leaf diseases, demonstrating superior performance compared to existing models in accuracy, precision, recall, and AUC metrics. To gain deeper insights into the model's performance, we utilize Grad-CAM++.

Plant phenotyping relevance

植物葉の病徴を画像から分類する深層学習手法を開発し、複数データセットで性能評価しており、植物フェノタイピング手法が研究の中心である。

Code and data availability

The paper evaluates LeafDoc-Net on two publicly available Mendeley datasets (cassava leaf disease and wheat leaf disease) that constitute the paper's phenotyping image inputs. Both are explicitly linked in the data availability statement with URLs matching allowed_urls. No author analysis code or trained model deposit,