📄 plant biology

Vision Normalizing Flows for the probability-informed detection of banana diseases from in-field images

This paper presents a lightweight, probabilistic image-recognition framework that leverages frozen DINOv3 embeddings and conditional normalizing flows to achieve near-perfect accuracy in detecting five major banana diseases from in-field images while providing interpretable uncertainty estimates for scalable deployment in smallholder farming systems.

Original authors: Prusokiene, A., Prusokas, A., Retkute, R.

Published 2026-07-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Prusokiene, A., Prusokas, A., Retkute, R.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but instead of a crime scene, you are looking at a banana leaf. In the world of farming, spotting a sick plant early is like finding a clue before the villain escapes; it stops the disease from spreading and saves the harvest. For a long time, computers trying to do this job have been like detectives who are too confident. They look at a picture and say, "That's definitely a banana with a specific disease!" even if the picture is actually of a different plant, a rock, or a banana that just looks a little weird because of the weather. They are trained to pick one answer from a list, and if the answer isn't on the list, they just guess anyway.

To fix this, scientists are using two powerful tools. First, there are "foundation models," which are like super-smart students who have read almost every picture in the world. They know what a leaf, a tree, or a banana looks like without needing to be taught specifically about banana diseases. Second, there are "normalizing flows." Think of these as a special kind of probability machine. Instead of just guessing a label, this machine asks, "How likely is it that this picture belongs to the 'sick banana' group?" and "How likely is it that this picture belongs to the 'healthy banana' group?" If the picture is a rock, the machine says, "Neither! This doesn't fit my map at all." This paper combines these two tools to create a detective that doesn't just guess, but actually knows when it's unsure.

The researchers behind this study, Alisa Prusokiene, Augustinas Prusokas, and Renata Retkute, wanted to build a system that could look at photos of banana plants taken in real fields and tell farmers if the plant is sick, healthy, or if the photo is just of something that isn't a banana at all. They focused on five nasty diseases that hurt banana farmers: Xanthomonas Wilt, Banana Bunchy Top Disease, Fusarium Wilt (also known as Panama disease), Yellow Sigatoka, and Black Sigatoka.

Here is how their "super-detective" works. They didn't try to teach a computer to learn everything from scratch. Instead, they used a pre-trained brain called DINOv3. Imagine DINOv3 as a genius who has already memorized the shapes and textures of millions of natural images. The researchers took this genius, froze its brain so it wouldn't change, and asked it to turn every banana photo into a long list of numbers (a 1,152-dimensional "embedding") that describes what the image looks like.

Then, they fed these number-lists into their special probability machine, the conditional normalizing flow. This machine learned to map out the "shape" of what a healthy banana looks like and the "shape" of what each of the five diseases looks like. When a new photo comes in, the machine doesn't just pick a winner; it calculates the exact mathematical likelihood of the photo belonging to each category.

The results were incredibly strong. When they tested the system on a set of images it had never seen before, it was almost perfect at telling the difference between sick and healthy plants. For every single disease, the model achieved an F1 score (a measure of accuracy) higher than 0.98. It also got "Average Precision" scores between 0.968 and 0.999, and "AUROC" scores (which measure how well it separates sick from healthy) between 0.997 and 1.000. In plain English, it was right almost all the time.

The most exciting part, however, is how the model handles confusion. The only time it really struggled was between Yellow Sigatoka and Black Sigatoka. These two diseases look very similar in the early stages, like twins wearing slightly different colored shirts. The model sometimes mixed them up, but even then, it knew it was confused. Because the model calculates probabilities, it can tell you, "I think this is Yellow Sigatoka, but I'm not 100% sure because it looks a lot like Black Sigatoka." This is a huge improvement over older systems that would just pick one and pretend to be certain.

The model also proved it could tell the difference between a banana plant and a completely different object. If you showed it a picture of a dog, a building, or a random bush, the system didn't force a banana disease diagnosis. Instead, it flagged the image as "out of distribution," meaning "This isn't a banana, so I can't diagnose it." This is crucial for real-world use, where a farmer might accidentally snap a photo of the wrong thing.

The researchers found that by using these two complementary "log-likelihood ratios"—one comparing the disease to a healthy banana, and another comparing the banana to non-banana images—they could create a two-dimensional map. On this map, sick bananas, healthy bananas, and random objects all land in different, clearly separated zones. This allows the system to not only diagnose the disease but also to flag uncertain predictions for a human expert to double-check.

In short, this paper suggests that by combining a frozen, pre-trained vision model with a probability-based flow, we can create a diagnostic tool that is both highly accurate and honest about its own uncertainty. It doesn't just say "It's sick"; it says "It's sick, and here is exactly how sure I am," while also knowing when to say, "Wait, that's not even a banana." This approach offers a promising path for building reliable, scalable tools that can help smallholder farmers in tropical regions protect their crops without needing expensive lab equipment.

Technical Summary: Vision Normalizing Flows for Probabilistic Banana Disease Detection

Problem Statement
Banana diseases, including Xanthomonas Wilt (BXW), Banana Bunchy Top Disease (BBTD), Fusarium Wilt, Yellow Sigatoka, and Black Sigatoka, cause severe production losses in smallholder farming systems. Accurate in-field diagnosis is hindered by variable symptom expression across cultivars and growth stages, as well as morphologically overlapping foliar signs. While laboratory methods (e.g., PCR, ELISA) offer high sensitivity, they are often inaccessible to smallholders due to cost and infrastructure requirements. Existing computer vision solutions, primarily based on Convolutional Neural Networks (CNNs), operate under a "closed-set" assumption. These discriminative models assign every input to a training class, often producing high-confidence predictions even for out-of-distribution (OOD) images (e.g., non-banana vegetation, atypical stress, or novel diseases), thereby failing to quantify uncertainty or recognize when an image falls outside the training distribution.

Methodology
The authors propose a probabilistic image-recognition framework that decouples visual feature extraction from probabilistic inference. The architecture consists of two main components:

Frozen Vision Foundation Model: The system utilizes the DINOv3 (Self-Distillation with No Labels) vision foundation model as a frozen feature extractor. Instead of fine-tuning the backbone, the model extracts 1,152-dimensional embeddings from input images. These embeddings are constructed by concatenating the classification token, an averaged register token, and an averaged patch token from the ViT-S variant of DINOv3. This approach leverages general-purpose self-supervised representations without task-specific adaptation.
Conditional Normalizing Flow: The extracted embeddings are fed into a conditional normalizing flow implemented using the PyTorch library Zuko. This flow models the conditional probability density $p_\theta(z | y)$ $p_{θ} (z ∣ y)$ , where $z$ $z$ is the image embedding and $y$ $y$ is the disease class conditioning vector.
- The flow employs eight autoregressive neural spline transformations parameterized by masked fully connected networks with SiLU activations.
- These transformations utilize monotonic rational-quadratic spline (RQS) bijections to map the complex multimodal distribution of DINOv3 embeddings onto a simple multivariate Gaussian latent distribution.
- Class conditioning is achieved via an eight-dimensional vector supplied to every transformation.
- During inference, the model evaluates the log-likelihood of a single image embedding under each of the eight disease/condition classes (five diseases, healthy banana, non-banana vegetation, and general natural imagery).

Training Strategy
The model was trained on a combined dataset of over 100,000 images sourced from four public datasets (Ouedraogo et al., Mduma and Elinisa, Garcin et al., and Unsplash). To address class imbalance and improve generalization, the authors employed:

Dirichlet Mixup: Training batches consisted of 4,096 latent vectors generated by linearly combining latent representations from different classes based on Dirichlet-distributed weights.
Augmentation: Extensive image augmentation (geometric, environmental, camera, and noise effects) was applied to the training set prior to feature extraction.
Optimization: The model was optimized using the Prodigy optimizer and the Schedule-Free framework, minimizing the average conditional negative log-likelihood (NLL) without manual learning rate scheduling.

Key Results
Evaluated on an independent test set, the framework demonstrated:

High Classification Accuracy: The model achieved multi-class accuracy near-perfect levels. F1 scores exceeded 0.98 for all five diseases. Average precision ranged from 0.968 to 0.999, and AUROC values were between 0.997 and 1.000.
Specific Performance: Fusarium Wilt achieved the highest performance (F1 = 0.9927, AUROC = 0.9998). The primary source of confusion was between Yellow Sigatoka and Black Sigatoka, a biologically plausible ambiguity due to overlapping early-infection symptoms.
Uncertainty and OOD Detection: Unlike softmax classifiers, the normalizing flow provided explicit likelihood estimates. The system successfully distinguished diseased plants, healthy plants, and OOD images (non-banana vegetation and general scenes) using two complementary log-likelihood ratios:
1. A disease ratio (disease class vs. healthy banana).
2. A plant ratio (banana vs. non-banana).
Diagnostic Space: These ratios define a two-dimensional diagnostic space that cleanly separates classes and flags uncertain predictions (samples near decision boundaries) for confirmatory testing.

Significance and Claims
The paper claims that this approach addresses a critical gap in agricultural AI: the inability of current systems to recognize when they are uncertain or when an input is outside their training distribution. Key contributions include:

Uncertainty-Aware Diagnostics: By estimating explicit conditional probability densities rather than decision boundaries, the system provides interpretable measures of prediction confidence and image relevance. This allows for the identification of ambiguous cases that require expert review or laboratory confirmation.
Efficiency and Transferability: The method achieves high performance without disease-specific fine-tuning of the vision backbone, suggesting that general-purpose foundation models capture sufficient visual features for disease discrimination. This reduces the dependency on large, disease-specific annotated datasets.
Scalability: The inference process relies on frozen embeddings and lightweight flow evaluation, making it compatible with smartphone deployment for smallholder farming systems and disease surveillance programs.

The authors note limitations, including the reliance on visible symptoms (unable to detect latent infections), the use of retrospective public datasets rather than prospective field trials, and the underrepresentation of mixed infections or novel diseases in the training data. They conclude that while the system does not replace molecular diagnostics, it offers a scalable tool for prioritizing scarce diagnostic resources and improving surveillance coverage.

More like this