What question did this study set out to answer?

The aim is to create a mobile solution for detecting sugarcane leaf diseases in real time to prevent crop losses.

February 23, 2026Open Access

Leaf-level sugarcane disease detection using CNN models: enhancing agricultural intelligence

Key Points

The aim is to create a mobile solution for detecting sugarcane leaf diseases in real time to prevent crop losses.
Developed a mobile and drone-compatible system for real-time monitoring.
Utilized MobileNetV2 for deep learning and CNN-based image analysis.
Applied soft-voting ensemble methods combining various classifiers like XGBoost, Random Forest.
Balanced the dataset from 400 to 1600 images through augmentation; 70/15/15 split for training, validation, and testing.
Conducted training with specific parameters including 25 epochs, Adam optimizer, and TensorFlow Lite for lightweight inference.
Achieved an ensemble accuracy of 96.2% and AUC-ROC of 98.0% on the test set.
Stand-alone MobileNetV2 performed with 94.8% accuracy and AUC of 97.0%.
Provided a comprehensive dataset and protocol for reproducibility and further research.

Abstract

The proposed solution will be mobile and can be used in the form of mobile or drone working systems where fields can be examined in their real time during the day so farmers and agronomists can take action. The system will help to prevent losses in crops, and sustainable agriculture in order to maximize the use of pesticides and monitor the well-being of plants in real time. Sugarcane leaf diseases (red rot, smut and leaf scald) can be detected early enough to prevent massive losses in yield. We have a transfer-learned deep-learning pipeline (MobileNetV2) with a soft-voting ensemble (combining CNN prediction with classical classifiers) (so-called) learned on CNN feature vectors (XGBoost, Random forest, Logistic regression) in this work, which is reproducible. We had 400 raw leaf images (?100/class) that with selected augmentation were balanced to 1600 images (400 images/class). The data was divided into training/validation/testing (70 15 15). The CNN (input 224 × 224, GlobalAveragePooling, Dense (128, ReLU), Dropout (0.5), Dense (4, softmax)) based on the Mo-bileNetV2 was trained with Adam (initial lr = 1e-4) in 25 epochs (batch size 32); the ensemble provided was the soft voting of the 5 classifiers. Accuracy = 96.2, Macro F1 = 96.0, and AUC-ROC = 98.0 of the ensemble on the held-out test set; standalone MobileNetV2 on the held-out test set Accuracy = 94.8 and AUC = 97.0. We also give complete dataset, preprocessing and training information to make sure it is reproducible. It has both lightweight inference path (TensorFlow Lite conversion) and low-quality image processing pipeline (quality assessment + enhancement) making it resistant to field and aerial data capture conditions. Important contributions: (1) a full-fledged MobileNetV2-pipetable pipeline to detect sugarcane leaf disease; (2) open dataset and augmentation protocol; (3) ablation and comparative experiments, which measure the contribution of the whole ensemble and each part.

Perguntar à IA

Bookmark

View Full Paper