PaperlayerView exact revision on arXiv

Source-linked AI summary

An In-field Automatic Wheat Disease Diagnosis System

Jiang Lu, Jie Hu, Guannan Zhao, Fenghua Mei, Changshui Zhang

arXiv:1710.08299v1cs.CV

TL;DR

Crop disease diagnosis needs practical methods for difficult in-field imagery, where existing approaches may depend on controlled images or omit localization. This paper proposes a weakly supervised deep multiple instance learning system with the WDD2017 dataset to identify wheat diseases and localize disease areas from image-level labels. The system improves recognition over conventional CNN architectures under matched parameter counts, maintains localization, and is packaged as a real-time mobile app.

  • Problem

    Existing crop-disease methods may rely on controlled images, while in-field wheat diagnosis requires simultaneous disease identification and localization without expensive manual annotations.

  • Method

    The paper develops DMIL-WDDS, a weakly supervised deep multiple instance learning framework evaluated on the newly collected in-field WDD2017 wheat disease dataset.

  • Results

    The proposed framework improves recognition over conventional CNN architectures under the same deep-model parameter count while maintaining localization of corresponding disease areas.

  • Takeaways & Limitations

    WDD2017 provides an in-field benchmark, and DMIL-WDDS is designed into a mobile app to support agricultural disease diagnosis.

  • Takeaways & Limitations

    The authors identify mixed cases involving multiple disease types or multiple crops as difficult challenges requiring a more robust future framework.

Abstract

from arXiv · show

Crop diseases are responsible for the major production reduction and economic losses in agricultural industry world- wide. Monitoring for health status of crops is critical to control the spread of diseases and implement effective management. This paper presents an in-field automatic wheat disease diagnosis system based on a weakly super- vised deep learning framework, i.e. deep multiple instance learning, which achieves an integration of identification for wheat diseases and localization for disease areas with only image-level annotation for training images in wild conditions. Furthermore, a new in-field image dataset for wheat disease, Wheat Disease Database 2017 (WDD2017), is collected to verify the effectiveness of our system. Under two different architectures, i.e. VGG-FCN-VD16 and VGG-FCN-S, our system achieves the mean recognition accuracies of 97.95% and 95.12% respectively over 5-fold cross-validation on WDD2017, exceeding the results of 93.27% and 73.00% by two conventional CNN frameworks, i.e. VGG-CNN-VD16 and VGG-CNN-S. Experimental results demonstrate that the proposed system outperforms conventional CNN architectures on recognition accuracy under the same amount of parameters, meanwhile main- taining accurate localization for corresponding disease areas. Moreover, the proposed system has been packed into a real-time mobile app to provide support for agricultural disease diagnosis.

1. Introduction

The paper targets automatic wheat disease identification and localization in difficult in-field images, using weak supervision to avoid costly manual annotations. It introduces DMIL-WDDS and the WDD2017 dataset as a practical basis for evaluating this integrated task.

  • Motivation: Manual and sensor-based crop disease diagnosis can be time-consuming, subjective, or dependent on expensive bulky equipment.Image-based methods offer a more accessible alternative using common cameras and consumer-level storage devices.
  • Challenges: In-field wheat diagnosis must handle complex backgrounds, changing capture conditions, multiple leaves or disease areas, disease-stage variation, and similar disease appearances.These challenges are illustrated through samples from WDD2017.
  • Objective: The proposed objective is to identify wheat disease categories and locate corresponding disease areas simultaneously in in-field images using only image-level training annotations.The twofold task is modeled as weakly supervised learning to avoid expensive and laborious manual annotation.
  • Proposed system: DMIL-WDDS combines deep learning and multiple instance learning with a fully convolutional network for local feature extraction and disease estimation.The system separates mobile-client and computing-server components and returns disease categories with corresponding disease-area positions.
  • Contributions: The framework integrates wheat disease identification and localization and outperforms conventional CNN recognition architectures under the same deep-model parameter count.The paper reports this result across the proposed contribution and system description.
  • Contributions: WDD2017 is a collected in-field wheat disease dataset introduced to demonstrate the system’s effectiveness and provide a benchmark for subsequent research.The dataset contribution complements the weakly supervised diagnosis framework.

2. Related work

Prior crop-disease methods commonly relied on controlled or idealized imagery, handcrafted preprocessing, or disease identification without localization. The paper positions its in-field system and WDD2017 dataset as responses to these practical gaps.

  • Existing limitations: Many earlier image-based crop-disease methods used pure backgrounds or controlled environments, limiting practical application in field conditions.Some approaches also required leaf segmentation before disease identification.
  • Existing limitations: Naive leaf segmentation becomes problematic when images contain multiple leaves or multiple infected regions.This creates a second difficulty for applying conventional preprocessing to complex images.
  • Deep learning gap: Deep learning approaches classified plant diseases from controlled-condition images but generally did not address where diseases were located.The cited approaches were therefore not applicable to wild environments and omitted localization.
  • Weakly supervised learning: Multiple instance learning has been combined with deep learning to reduce manual annotation and achieve object localization under weak labels.Prior work included assumptions about desired objects lying among region proposals.
  • Fully convolutional networks: The paper exploits a fully convolutional network for lower-cost instance-level disease estimation, equivalent to sliding-window processing over the whole image.FCNs were originally proposed for semantic segmentation.
  • Paper positioning: DMIL-WDDS addresses in-field wheat disease identification and localization, while WDD2017 supplies an in-field dataset intended as a benchmark for subsequent work.The paper describes this combination as closer to practical agricultural situations.

3. Materials and methods

The system combines in-field wheat disease data with a weakly supervised deep multiple instance learning framework to identify diseases and localize affected areas from image-level labels. It uses fully convolutional spatial predictions, MIL aggregation, and bounding-box approximation for diagnosis and localization.

  • Wheat Disease Database 2017 (WDD2017): WDD2017 includes complex backgrounds, varying capture conditions, disease stages, and similar appearances between different wheat diseases.Each image almost exclusively contains the annotated disease type, while the collection retains primitive in-field capture information.
  • Aggregation by multiple instance learning (MIL): MIL reduces annotation effort by assigning labels to bags rather than individual instances, treating a bag as positive when it contains at least one positive instance.The framework aggregates instance-level class probabilities into bag-level estimates using functions such as Max, Avg, or Softmax.
  • Fully convolutional network (FCN): The FCN converts fully connected layers into convolutional representations and produces spatial score maps whose locations correspond to disease estimates for receptive fields.The modified VGG-CNN-VD16 fully connected channels are changed from {4096, 4096, 1000} to {1024, 1024,C}.
  • Bounding boxes approximation (BBA): Disease localization thresholds the predicted disease score map, extracts contours, and generates scaled-down bounding boxes around the corresponding contours.The BBA step reduces overlap introduced by the FCN before outputting boxes for disease areas.
  • DMIL-WDDS: Training uses image-level one-hot labels and minimizes mean-square error between image-level labels and predictions while regularizing model parameters.At inference, MIL produces class score maps for identification, which are up-sampled into heat maps before BBA generates disease-area boxes.

4. Experiments

The experiments evaluate conventional CNN baselines and DMIL-WDDS models on WDD2017 using five-fold cross-validation, with matched model-parameter comparisons and specified training settings.

  • 4. Experiments: The experiments build several models to estimate recognition accuracy and show disease localization results on WDD2017.The implementation uses Theano and runs on a Geforce GTX 1080 GPU.
  • 4.1. Models: DMIL-WDDS is evaluated against VGG-CNN-S and VGG-CNN-VD16 using corresponding VGG-FCN-S and VGG-FCN-VD16 models.The paired models have the same amount of parameters within each architecture family.
  • 4.2. Dataset and settings: WDD2017 images are resized to 224 × 224 × 3 for conventional CNNs and 832 × 832 × 3 for DMIL-WDDS.The dataset is uniformly split into five folds for cross-validation.
  • 4.2. Dataset and settings: DMIL-WDDS models are trained for 20 epochs with batch size 2 and initial learning rate 0.00005, while conventional CNNs use 60 epochs, batch size 45, and initial learning rate 0.0001.Both model groups use Nesterov Momentum SGD with momentum 0.9.
  • 4.2. Dataset and settings: Five-fold cross-validation reports class-wise and total test accuracy, with hyperparameters selected using a random validation set from the training folds.The reported accuracies are averaged across folds.

5. Results and discussion

DMIL-WDDS improves wheat disease recognition over matched conventional CNN architectures while also localizing disease areas in challenging in-field images. Soft aggregation generally provides the strongest recognition and more precise localization, and the framework is presented as extensible but not yet targeted at mixed multi-disease or multi-crop cases.

  • 5.1. Wheat disease identification: Soft-agg achieves the best recognition performance for both DMIL-WDDS models, Avg-agg ranks second, and Max-agg performs worst.The comparison covers total accuracies across five-fold cross-validation.
  • 5.1. Wheat disease identification: 95.12% VGG-FCN-S accuracy exceeds 93.27% VGG-CNN-VD16 accuracy, while VGG-FCN-S with Soft-agg reaches 66.57% for Powdery Mildew versus 2.00% for VGG-CNN-S.The authors relate these observations to fine feature extraction for local areas in the whole image.
  • 5.2. Contrast test and feature visualization: Masking diseased areas changes a correctly recognized Stripe Rust image to Healthy Wheat, indicating sensitivity to disease areas rather than other wheat regions.The contrast test uses the VGG-FCN-VD16 DMIL-WDDS model.
  • 5.2. Contrast test and feature visualization: Feature maps progress from pixel-level patterns in Block 1 to distinct features of four diseased areas in Block 4.The visualization is based on a raw Stripe Rust image processed by VGG-FCN-VD16.
  • 5.3. Localization for disease areas: Soft-agg gives more precise localization than Max-agg's partial localization and Avg-agg's overly general localization.The authors describe Soft-agg as a trade-off between Max-agg and Avg-agg.
  • 5.3. Localization for disease areas: DMIL-WDDS with VGG-FCN-VD16 and Soft-agg localizes multiple Smut areas despite hands and illumination changes in the image.The same examples retain correct class-level predictions.

6. Conclusion

DMIL-WDDS integrates wheat disease recognition and localization in-field using image-level labels, outperforming conventional CNNs and supporting mobile diagnosis.

  • DMIL-WDDS improves wheat disease recognition over conventional CNN architectures under the same amount of deep-model parameters.The framework also outperforms a deeper conventional CNN model when using a shallow VGG-FCN-S model.
  • Softmax aggregation provides higher recognition accuracy than alternative aggregation functions within the MIL framework.
  • Softmax-based DMIL-WDDS produces more accurate localization of disease areas than two other aggregation functions.
  • The proposed diagnosis system is implemented as a mobile app to support agricultural disease diagnosis.
Loading 1710.08299v1…