Source-linked AI summary
PlantDoc: A Dataset for Visual Plant Disease Detection
Davinder Singh, Naman Jain, Pranjali Jain, Pratik Kayal, Sudhakar Kumawat, Nipun Batra
TL;DR
Plant disease detection needs scalable methods that work beyond controlled laboratory imagery, especially under natural field conditions. The paper introduces PlantDoc, a real-world dataset and benchmark, finding that fine-tuning on it reduces classification error by up to 31% and exposes the limited efficacy of models trained on controlled datasets.
Problem
Plant disease detection lacks sufficiently large-scale, non-controlled datasets, while existing vision solutions often rely on high-resolution images with plain backgrounds.
Method
The paper curates PlantDoc, a dataset of 2,598 images spanning 13 plant species and 27 disease/healthy classes, and benchmarks classification and object-detection models.
Results
Up to 31%: fine-tuning models on PlantDoc reduces classification error, while models trained on controlled datasets show low accuracy on real-case images.
Takeaways & Limitations
PlantDoc provides a first step toward computer-vision-based scalable plant disease detection in non-controlled environments.
Takeaways & Limitations
The dataset may contain mislabeled images, and more images per class may be needed to train highly accurate disease-detection models.
Abstract
from arXiv · showhide
India loses 35% of the annual crop yield due to plant diseases. Early detection of plant diseases remains difficult due to the lack of lab infrastructure and expertise. In this paper, we explore the possibility of computer vision approaches for scalable and early plant disease detection. The lack of availability of sufficiently large-scale non-lab data set remains a major challenge for enabling vision based plant disease detection. Against this background, we present PlantDoc: a dataset for visual plant disease detection. Our dataset contains 2,598 data points in total across 13 plant species and up to 17 classes of diseases, involving approximately 300 human hours of effort in annotating internet scraped images. To show the efficacy of our dataset, we learn 3 models for the task of plant disease classification. Our results show that modelling using our dataset can increase the classification accuracy by up to 31%. We believe that our dataset can help reduce the entry barrier of computer vision techniques in plant disease detection.
1 INTRODUCTION
PlantDoc addresses the difficulty of detecting plant diseases in real-world conditions by providing a computer-vision dataset collected from non-controlled environments. The dataset is benchmarked to assess its utility for scalable disease detection.
- 35% of crop production in India is lost annually to pests and disease, threatening food security and creating economic, social, and environmental impacts.
- Farmers often lack accessible expertise and laboratory infrastructure for timely identification of diseased leaves.
- Computer vision offers scalable, cost-effective image classification suitable for rapid inference on consumer smartphones.
- PlantDoc contains 2,598 images across 13 plant species and 27 healthy-or-diseased classes collected under natural backgrounds and lighting.
- Fine-tuning models on PlantDoc reduces classification error by up to 31%, while controlled-dataset models perform poorly in real scenarios.
2 RELATED WORK
Prior plant-disease vision research includes specialized techniques and datasets, but existing resources are often narrow, private, or laboratory-controlled. PlantDoc is positioned as a publicly available dataset of real-life plant images.
- Plant-disease research spans detection techniques and datasets intended to advance the field.
- Earlier image-based studies targeted particular crops, and their datasets were not publicly released, limiting reproducibility.
- PlantVillage was the only public plant-disease dataset identified by the authors and reported 99.35% accuracy using automated disease-detection systems.
- PlantVillage images were collected in laboratory setups, unlike the real cultivation-field conditions targeted by PlantDoc.
3 THE PLANTDOC DATASET
PlantDoc was curated from internet images to represent plant disease detection under real-world conditions, with manual filtering, verification, and leaf-level bounding-box annotation. Cropped-PlantDoc additionally provides leaf crops derived from those annotations.
- Data Collection: PlantDoc was created from internet images to capture real-life farm conditions without the enormous effort required for large-scale field collection.
- Data Collection: Four users filtered images using website metadata and APSNet guidelines, removed unsuitable or duplicate images, and excluded classes with fewer than 50 images.
- Data Collection: Every image was checked by two individuals to reduce labeling errors, using visual factors such as disease color, area, density, and species shape.
- Data Collection: Leaf bounding boxes were annotated for all leaves, including mixed healthy and diseased images, with coordinates and class labels stored in XML files.
- Cropped-PlantDoc Dataset: Cropping 2,598 images by their bounding boxes produced 9,216 leaf images in the Cropped-PlantDoc dataset.
4 BENCHMARKING PLANTDOC DATASET
The paper benchmarks PlantDoc through classification and leaf-detection experiments designed for non-controlled image settings. It compares training setups and architectures to assess performance on real-world plant images.
- Two benchmark tasks evaluate plant image classification and leaf detection on the PlantDoc dataset.
- Training uses stochastic gradient descent with momentum 0.9, categorical crossentropy, 0.001 learning rate, augmentation, and 100 × 100 image inputs.Pre-trained models use Keras weights trained on ImageNet.
- The benchmark includes comparisons between controlled PlantVillage data and real-world PlantDoc images.The experiments test whether models trained on controlled data perform effectively in non-controlled settings.
- Transfer learning, classification architectures, and object-detection models are evaluated across uncropped and cropped PlantDoc settings.The experiments include VGG16, several CNN architectures, and Faster R-CNN with InceptionResNetV2 and MobileNet.
- The benchmarking uses mean average precision for object detection, with higher mAP indicating better performance.The comparison includes scores on the COCO dataset because no plant-disease evaluation existed for the domain.
5 RESULTS AND DISCUSSION
Results indicate that controlled-dataset training performs poorly on real-world images, while PlantDoc supports stronger classification and detection performance. The models also focus on disease-related visual features and leaf shape.
- Saliency and activation maps show attention to disease-correlated blemishes and to leaf shape for distinguishing species.Examples include lines in Corn Leaf Blight and spots in Tomato Bacterial Spots.
- Training on PlantVillage or ImageNet initially produces low accuracy on real-world plant images.The reported failure is associated with background noise, multiple leaf classes, and low-resolution leaves.
- Faster R-CNN with InceptionResNetV2 achieves the best reported leaf-detection performance, with an mAP of 38.9.
- Tomato Bacterial leaf spot and Septoria can look similar and are difficult to label using visual features alone.
6 APPLICATION BUILDING
The authors adapt the detection approach to a mobile application using a lower-complexity MobileNet-based object-detection model for real-time CPU inference.
- A MobileNet object-detection network built on SSD is used to support real-time bounding-box and class prediction on a mobile CPU.The application targets reduced complexity while retaining effective accuracy and competitive performance.
7 LIMITATIONS
The dataset has annotation-quality and coverage limitations that constrain highly accurate disease detection.
- Some images may be wrongly classified because the dataset lacks extensive domain expertise during curation.The paper illustrates potentially ambiguous examples involving visually similar diseases.
- Highly accurate disease-detection models may require more images in each class.The authors describe PlantDoc as a feasible direction given limited public real-life field data.
8 CONCLUSIONS AND FUTURE WORK
The paper presents PlantDoc as a new dataset for visual plant disease detection and evaluates its relevance to scalable computer-vision-based detection. The authors identify controlled-dataset limitations and suggest image segmentation as future work.
- PlantDoc is introduced as an entirely new dataset for plant disease detection.
- Benchmark experiments show that models learned on controlled datasets lack efficacy for real-world plant-disease images.
- The paper frames PlantDoc as an important first step toward scalable computer-vision-enabled plant disease detection.
- Image segmentation to extract leaves from images is proposed as a way to potentially enhance the dataset's utility.