Foliagen: Framework for Foliage Image Generation from Individual Crop Leaf Images

ICLR 2026 Conference Submission20684 Authors

Keywords: Foliage Dataset, Leaf Disease, Image Classification, Transfer Learning, Image Generation

TL;DR: This paper introduces a framework, called Foliagen, to generate diseased foliage image from individual diseased images with customizable disease rate.

Abstract: While machine learning (ML)-based crop disease classifiers mostly targeted individual leaf images, real-world applications call for disease classification on crop foliage images instead, because they usually rely on cameras mounted on unmanned aerial vehicles to capture foliage images across vast crop fields for automated disease identification. We found that known state-of-the-art (SOTA) classifiers on the only real-world soybean foliage image dataset all exhibited unsatisfactory performance, despite the dataset being modest-sized and including just two soybean disease categories (among many). Hence, it is desirable to make available large foliage image datasets with common crop disease categories for better evaluating and possibly improving SOTA crop disease classifiers on foliage images. This paper introduces a framework that generates crop foliage images utilizing available datasets of individual leaf images, termed Foliagen (short for foliage generation). A generated foliage image dataset can be arbitrarily sized, with each image emulating the natural distribution of diseased leaves with a specified disease rate. Being annotated by design, such generated datasets are valuable for (1) evaluating the SOTA classifiers when applied to practical use and (2) pre-training general SOTA classifiers, making it possible to effectively fine-tune them using any real-world foliage image dataset for improved classification performance. The Foliagen framework is exemplified by generating foliage image datasets for soybean and tomato. Our evaluation results indicate that five SOTA classifiers on generated datasets with nine disease categories achieve accuracy up to 87\% for soybean and 86\% for tomato under $\gamma$ = 5\%, and that they all exhibit less than 92\% in classifying the real soybean foliage image dataset (with just two disease categories). Foliagen makes it possible to generate crop foliage image datasets to evaluate future disease classifiers objectively, aiming at in-field applications.

Primary Area: datasets and benchmarks

Submission Number: 20684