Source-linked AI summary
Deep Learning for Classification and Severity Estimation of Coffee Leaf Biotic Stress
J. G. M. Esgario, R. A. Krohling, J. A. Ventura
TL;DR
Coffee biotic stresses threaten productivity and crop quality, motivating practical identification and severity estimation from leaf images. The study develops a CNN-based multi-task system with augmentation, achieving strong classification and severity-estimation accuracies across evaluated datasets.
Problem
Coffee pests and diseases can reduce photosynthesis, yield, and product quality, while plant diseases threaten productivity and small farmers’ livelihoods.
Method
The study builds a larger coffee-leaf dataset and evaluates CNNs in a multi-task system for automatic biotic-stress classification and severity estimation, including standard augmentation and mixup.
Results
Multi-task learning achieved average accuracies of 94.05% for biotic-stress classification and 84.76% for severity estimation, while symptom classification reached 96.63%.
Takeaways & Limitations
ResNet50 provided the best overall tradeoff among computational cost, performance, and reliability, and multi-task learning improved effectiveness for related tasks.
Takeaways & Limitations
The diagnosis setting remains constrained by complex backgrounds, uncontrolled image capture, poorly defined symptom boundaries, and the need for representative, reliably labeled datasets.
Abstract
from arXiv · showhide
Biotic stress consists of damage to plants through other living organisms. Efficient control of biotic agents such as pests and pathogens (viruses, fungi, bacteria, etc.) is closely related to the concept of agricultural sustainability. Agricultural sustainability promotes the development of new technologies that allow the reduction of environmental impacts, greater accessibility to farmers and, consequently, increase on productivity. The use of computer vision with deep learning methods allows the early and correct identification of the stress-causing agent. So, corrective measures can be applied as soon as possible to mitigate the problem. The objective of this work is to design an effective and practical system capable of identifying and estimating the stress severity caused by biotic agents on coffee leaves. The proposed approach consists of a multi-task system based on convolutional neural networks. In addition, we have explored the use of data augmentation techniques to make the system more robust and accurate. The experimental results obtained for classification as well as for severity estimation indicate that the proposed system might be a suitable tool to assist both experts and farmers in the identification and quantification of biotic stresses in coffee plantations.
1. Introduction
Coffee plantations face biotic stresses that reduce crop performance, while disease diagnosis and severity estimation remain important but challenging. This work motivates deep-learning and multi-task approaches using more representative datasets to identify and quantify coffee leaf stress.
- Biotic agents such as pests and pathogens impair plant functions, threaten food security, and can cause substantial crop losses, especially for small farmers.
- Severity estimation matters alongside causal-agent identification because symptom severity supports yield prediction and recommendations for control treatments.
- Coffee stresses including leaf miner, rust, brown leaf spot, and cercospora leaf spot can cause defoliation, reduced photosynthesis, and lower yield and product quality.
- Automatic diagnosis remains difficult because field images vary in backgrounds and capture conditions, while symptoms vary by disease stage and may overlap across disorders.
- Representative datasets are difficult and costly to construct because image labeling often requires specialists, while socially sourced data may be unreliable.
- The proposed extension introduces a larger coffee-leaf dataset and convolutional neural networks for more robust automatic classification and severity estimation.
2. Materials and methods
The study builds coffee-leaf datasets and CNN systems for biotic-stress classification and severity estimation, including a shared multi-task architecture. It also evaluates standard augmentation and mixup to improve generalization from limited data.
- 2.1. Image dataset: The dataset contains 1747 coffee-leaf images, including healthy leaves and leaves affected by leaf miner, rust, brown leaf spot, or cercospora leaf spot.Labeling was assisted by a specialist.
- 2.1. Image dataset: Two datasets were generated: original whole-leaf images and cropped symptom images containing a single stress.The symptom dataset includes 2722 images after adding 575 images from Barbedo (2019).
- 2.1. Image dataset: Whole-leaf images were labeled by predominant stress and severity, with 62 visually ambiguous multi-stress leaves excluded.Severity labels ranged from healthy (< 0.1%) to very high (> 15%).
- 2.2. Data augmentation: Standard augmentation used mirroring, rotation, and color variation, while mixup linearly combined two randomly selected images and their one-hot labels.Mixup generates {x̃, ỹ} = λ{x_i, y_i} + (1 − λ){x_j, y_j}, with λ sampled from a beta distribution.
- 2.3. Deep learning architectures: The study selected common CNN architectures and adapted them for multi-task learning by adding a parallel fully connected layer while sharing convolutional layers.The shared layers learn joint features for stress classification and severity estimation.
- 2.3. Deep learning architectures: Multi-task learning reuses most of the network architecture, requiring only one model and making learning substantially faster.
3. Experiments
The experiments evaluate CNN-based biotic-stress classification and severity estimation using leaf and symptom images, multiple architectures, multi-task learning, and augmentation. ResNet50 and multi-task learning achieved strong results, while class imbalance and severity-label structure affected errors and mixup performance.
- 3.1. Experimental setup: Images were segmented, cropped, and resized to 224x224x3 so leaves occupied more of the input and small symptoms were easier for CNNs to capture.A fixed threshold in the HSV color space was used to fit a bounding box around the leaf.
- 3.1. Experimental setup: Experiments used online augmentation, transfer learning from ImageNet-pretrained networks, and 70-15-15 training, validation, and test splits.Models were selected using the lowest validation loss and evaluated with accuracy, precision, and recall.
- 3.2. Leaf dataset results: ResNet50 achieved 84.13% accuracy for severity estimation, with most errors involving high and very high severity classes because of dataset imbalance.Errors were concentrated near the main diagonal, so adjacent severity mistakes were considered less severe for this ordinal task.
- 3.3. Symptom dataset results: The symptom dataset reached 96.63% average accuracy, as isolated symptom regions reduced spurious elements and allowed the network to focus on symptom details; ResNet50 again performed best.VGG16 performed well on the Leaf dataset but poorly on the Symptom dataset.
- 3.4. Visualization: The t-SNE visualization showed well-defined class clusters overall, although cercospora leaf spot had the greatest overlap with other classes.Standard augmentation produced a very linear relationship between severity labels and their severity range.
- 3.3. Symptom dataset results: Mixup reduced within-class feature variance and improved recall slightly for classification, but worsened severity estimation by favoring intermediate labels over extreme ones.The visualization analysis links this behavior to mixing images with ordinal severity labels.
4. Conclusion
The study developed deep-learning approaches for classifying biotic stress and estimating severity in coffee leaves, using a new coffee-leaf image dataset. ResNet50 produced the best results, while the dataset’s limited stress coverage remains a key scope boundary.
- 4. Conclusion: The work addressed biotic-stress classification and severity estimation for important coffee diseases and pests using leaf images.A new dataset of coffee-leaf images was developed for the experiments.
- 4. Conclusion: ResNet50 obtained the best results among the CNN architectures evaluated.The study also found that multi-task learning could improve performance and computational cost for tasks in the same domain.
- 4. Conclusion: The dataset has low representativity because it covers only the main biotic stresses affecting coffee trees.The authors suggest that adding more stress types and images may help alleviate this limitation.
- 4. Conclusion: An application developed in the authors’ laboratory is being used to embed the technology for practical use, with initial tests described as promising.Further results from those tests were planned for future reporting.