Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Grapevine Disease Prediction Using Climate Variables from Multi-Sensor Remote Sensing Imagery via a Transformer Model

Weiying Zhao, Natalia Efremova

TL;DR

This study tackles block-level grapevine disease prediction by integrating climate variables from multi-sensor remote sensing imagery with a Transformer-based TabPFN framework. The method leverages the posterior predictive distribution $p(y|x, D_{train})$, approximated offline, to produce fast online predictions $q_ heta(y|x_{test}, D_{train})$ on new data. Experiments on 76 Australian vineyards (627 blocks) show that TabPFN-based PFNClassifier delivers competitive performance against gradient-boosted trees, particularly under class imbalance, and enables per-block disease risk maps for targeted interventions. The work demonstrates a scalable, data-efficient approach for precision viticulture with potential environmental and productivity benefits, and points to future enhancements via phenology and temporal climate features.

Abstract

Early detection and management of grapevine diseases are important in pursuing sustainable viticulture. This paper introduces a novel framework leveraging the TabPFN model to forecast blockwise grapevine diseases using climate variables from multi-sensor remote sensing imagery. By integrating advanced machine learning techniques with detailed environmental data, our approach significantly enhances the accuracy and efficiency of disease prediction in vineyards. The TabPFN model's experimental evaluations showcase comparable performance to traditional gradient-boosted decision trees, such as XGBoost, CatBoost, and LightGBM. The model's capability to process complex data and provide per-pixel disease-affecting probabilities enables precise, targeted interventions, contributing to more sustainable disease management practices. Our findings underscore the transformative potential of combining Transformer models with remote sensing data in precision agriculture, offering a scalable solution for improving crop health and productivity while reducing environmental impact.

Grapevine Disease Prediction Using Climate Variables from Multi-Sensor Remote Sensing Imagery via a Transformer Model

TL;DR

This study tackles block-level grapevine disease prediction by integrating climate variables from multi-sensor remote sensing imagery with a Transformer-based TabPFN framework. The method leverages the posterior predictive distribution , approximated offline, to produce fast online predictions on new data. Experiments on 76 Australian vineyards (627 blocks) show that TabPFN-based PFNClassifier delivers competitive performance against gradient-boosted trees, particularly under class imbalance, and enables per-block disease risk maps for targeted interventions. The work demonstrates a scalable, data-efficient approach for precision viticulture with potential environmental and productivity benefits, and points to future enhancements via phenology and temporal climate features.

Abstract

Early detection and management of grapevine diseases are important in pursuing sustainable viticulture. This paper introduces a novel framework leveraging the TabPFN model to forecast blockwise grapevine diseases using climate variables from multi-sensor remote sensing imagery. By integrating advanced machine learning techniques with detailed environmental data, our approach significantly enhances the accuracy and efficiency of disease prediction in vineyards. The TabPFN model's experimental evaluations showcase comparable performance to traditional gradient-boosted decision trees, such as XGBoost, CatBoost, and LightGBM. The model's capability to process complex data and provide per-pixel disease-affecting probabilities enables precise, targeted interventions, contributing to more sustainable disease management practices. Our findings underscore the transformative potential of combining Transformer models with remote sensing data in precision agriculture, offering a scalable solution for improving crop health and productivity while reducing environmental impact.
Paper Structure (6 sections, 1 equation, 4 figures, 1 table)

This paper contains 6 sections, 1 equation, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Experiments
  4. Dataset
  5. Experimental results and discussion
  6. Conclusion

Figures (4)

  • Figure 1: Flowchart of the proposed disease forecasting framework. The TabPFN method hollmann2023tabpfnpicard2024fast is used as an example. The TabPFN learns to approximate the PPD of a given prior in the offline stage to yield predictions on a new dataset in a single forward pass in the online stage.
  • Figure 2: Total blocks affected by different kinds of diseases. The diseases have imbalanced distribution.
  • Figure 3: ROC curve results comparison of the methods. The experimental results are based on the same training and testing databases.
  • Figure 4: Disease probability maps for 10 blocks in early 2021 in Australia. The probability of disease is displayed as a visual heat map (green = low probability, yellow = high probability).