Decoder-Aware Self-Supervised Continual Pretraining and Uncertainty-Guided Pseudo-Labeling for Wheat Organ Segmentation
Bibliographic record
Abstract
Accurate segmentation of wheat organs—spikes, leaves, and stems—is critical for plant phenotyping but remains a challenge due to domain variability and limited labeled data. In this work, we propose a label-efficient segmentation framework that utilizes only 99 labeled samples from the GWFSS dataset. Our approach begins with a continual decoder-aware pretraining (DeCon-ML), where a ConvNeXt-L encoder (initialized with ImageNet1K weights) is paired with a Feature Pyramid Network (FPN) decoder and pretrained on 64,368 unlabeled wheat images in two stages: training the decoder while keeping the encoder frozen, followed by joint encoder-decoder pre-training. For fine-tuning, we train ConvNeXt-L models on the labeled set with 99 samples, initialized with ImageNet- 1K and GWFSS-pretrained weights, and ensemble them to leverage complementary representations. We further enhance generalization by incorporating pseudo-labels, chosen based on feature-space similarity to the labeled set and filtering out low-confidence predictions through uncertainty estimation. We also incorporate a BEiT-L model trained only on the training set and ensembled with ConvNext-L models to achieve our best results. Our proposed approach achieves a mean Intersection over Union (mIoU) of 73.61 and 67.88, respectively, on the validation and test set of the GWFSS dataset, effectively segmenting all four classes (spikes, leaves, stems, background) under varying conditions. This study demonstrates how combining continual pretraining, similarity-aware candidate selection and uncertainty-guided pseudo-labeling can significantly improve semantic segmentation with minimal supervision in agricultural vision. Code is available at https://github.com/tapu1996/DeCon-UGPL-GWFSS.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.002 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.002 | 0.001 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".