XpressWeed: Meta-Inspired Few-Shot Adaptation for Plant Weed Segmentation Using Texture Priors

1 Overview

XpressWeed is a texture-prior driven segmentation framework for weed identification in precision agriculture. It reframes the problem: instead of treating plant leaves as rigid objects, XpressWeed models them as textures — leveraging the fact that each species exhibits distinctive vein patterns, surface roughness, and micro-structures that remain stable under deformation, occlusion, and changing illumination.

A Fully Convolutional Network (FCN) is pre-trained on collages from 12 leaf texture classes, building robust and species-agnostic feature representations. A Model-Agnostic Meta-Learning (MAML) strategy then adapts this model to 4 completely unseen weed species using only 80 images — achieving 85% accuracy versus 48% for traditional MAML on the same data.

Fig. 1 — Feature landscapes showing how texture priors create well-separated class regions (left), how MAML alone struggles with new data points (top-right), and how XpressWeed efficiently discriminates unseen classes by combining both (bottom-right).

2 The Problem

Weeds compete with primary crops for nutrients, water, and sunlight — causing significant yield loss worldwide. While deep learning has been applied to weed segmentation, two fundamental obstacles remain unsolved:

3 Proposed Method

XpressWeed combines supervised texture pre-training with MAML-based few-shot adaptation. The two-stage design separates large-scale representation learning from lightweight few-shot adaptation, achieving both robustness and computational efficiency.

Fig. 2 — Workflow of the XpressWeed framework: supervised learning builds robust texture priors, followed by MAML inner/outer loop adaptation to new weed species.

Stage 1 — Supervised Texture Pre-training

The FCN is trained on texture-prior images constructed from PlantVillage leaf patches arranged into collages. Overlapping leaves and background-removed variants simulate real farmland conditions where boundaries are difficult to separate. Weighted cross-entropy loss handles severe class imbalance. Meta-learning is deliberately not used at this stage — it would require expensive second-order differentiation over thousands of images.

Stage 2 — MAML-Based Adaptation

After pre-training, MAML adapts the model to 4 unseen weed species with only 80 labeled patches. Each episode uses 3-way classification, 12 support samples per class (inner loop update), and 10 query samples per class (outer loop meta-update). The n-way episodic setup implicitly mimics contrastive learning — same-class embeddings cluster together while different classes are pushed apart — without explicit contrastive losses.

4 Network Architecture

The backbone is an encoder–decoder design inspired by U-Net, modified to emphasize texture discrimination and multi-scale feature fusion. It is intentionally lightweight — MAML's second-order gradient updates require smaller models to converge efficiently.

Fig. 3 — Architecture of the proposed CNN model. Input: 512×768×3 → Bottleneck: 8×12×512 → Output: 512×768×N_classes. Scroll to see full architecture.

5 Stage 1: Pre-training Results

The FCN was trained on a PlantVillage-derived dataset of 12 leaf texture classes arranged into collages. The design intentionally introduces challenging conditions — overlapping leaves, background-removed variants, and mixed textures — to simulate real farmland boundaries.

Fig. 4 — Dataset for supervised texture pre-training: leaf collages from 12 PlantVillage species with overlapping leaves and sharp boundaries.

Pre-training Metrics

Training Accuracy96%

Validation Accuracy92%

Mean IoU92%

mAP@[0.50:0.95]0.90

✅

All 12 texture classes achieve AP ≈ 0.99, confirming the model learns strong, discriminative texture representations before adaptation.

Fig. 5 — Per-class Precision–Recall curves for the supervised pre-training model. All 12 classes show AP ≈ 0.99, confirming strong discrimination.

Fig. 5b — Mean Average Precision across IoU thresholds [0.50:0.95] = 0.90.

6 Stage 2: MAML Adaptation Results

The pre-trained model is adapted to 4 unseen weed species using a real Indian farmland dataset. 20 high-resolution (2112×1600) images were cropped into 80 patches (512×768) for MAML training. The task is 3-way classification, 12-shot, over 25 epochs.

Fig. 6 — Sample images and pixel-level masks from the farmland dataset. 6 classes: background, soybean (crop), + 4 unseen weed species.

Fig. 7 — Pixel-wise class distribution in the MAML adaptation dataset — highly imbalanced, mirroring real farmland conditions.

XpressWeed vs. Traditional MAML

Fig. 8 — Training set results: input image · ground truth mask · model prediction. Strong segmentation of soybean (crop) and weed species.

Fig. 9 — Held-out validation results: the adapted model generalises to unseen farmland images with 85% accuracy.

Per-Class Average Precision

Fig. 10 — Per-class Precision–Recall curves for the MAML-adapted model. Macro mAP = 0.783, Micro AUPRC = 0.962. Background and crop (seen species) near-perfect; Lamber (fewest annotated pixels) is the weakest weed class.

7 Comparison with Related Work

8 Code & Repository

Class Color Mapping

# RGB mask colors → integer class IDs (farmland adaptation dataset)
color2class = {
    (0,   0,   0):   0,   # Background
    (61,  61,  245): 1,   # Plant data (Soybean — seen in pre-training)
    (28,  101, 232): 2,   # Kena        (unseen weed)
    (255, 53,  94):  3,   # Lavhala     (unseen weed)
    (255, 106, 77):  4,   # Lamber      (unseen weed — fewest pixels)
    (238, 164, 15):  5,   # Obscure Morning Glory (unseen weed)
}

MAML Episode Structure

# Episode configuration
N_WAY       = 3      # classes per episode
K_SUPPORT   = 12     # support samples per class (inner loop)
K_QUERY     = 10     # query  samples per class (outer loop)
N_EPOCHS    = 25

for epoch in range(N_EPOCHS):
    for episode in sample_episodes():
        support_x, support_y = episode['support']   # (36, 512, 768, 3)
        query_x,   query_y   = episode['query']     # (30, 512, 768, 3)

        # Inner loop: fast adaptation on support set
        fast_weights = inner_loop_update(model, support_x, support_y)

        # Outer loop: meta-gradient update on query set
        meta_loss = compute_loss(fast_weights, query_x, query_y)
        meta_update(model, meta_loss)

Data Augmentation

# Augmentations applied consistently to both images AND masks:
#  1. Random horizontal flip           (p = 0.5)
#  2. Random zoom  0.9× – 1.1×         → crop-centre or reflect-pad
#  3. Random shadow / highlight (RSH)  (p = 0.85, triangular regions)
#  4. Color jitter                     → hue, saturation, brightness
#  5. Gaussian blur                    → simulates out-of-focus leaves
#  6. Gaussian noise                   → simulates sensor noise

aug_imgs, aug_masks = augment_images_and_masks(
    images_batch, masks_batch,
    target_size=(512, 768),
    rsh_prob=0.85
)

9 Citation

If you use XpressWeed in your research, please cite:

@inproceedings{kethineni2026xpressweed,
  title     = {XpressWeed: Meta-Inspired Few-Shot Adaptation for
               Plant Weed Segmentation Using Texture Priors},
  author    = {Kethineni, Kiran K. and Kanukuntla, Rishi Raj and
               Mohanty, Saraju P. and Kougianos, Elias},
  booktitle = {Proceedings of the IEEE SusTech 2026},
  year      = {2026},
  institution = {University of North Texas}
}

Related Work from Our Group