An adaptation of CLIP that starts with a small labeled seed set and recursively mines more examples scored 99.09% on UCM and 95.07% on NWPU. In the paper's comparison with named semi-supervised baselines, those results were reported as 3.22% and 6.21% improvements over DARP, respectively. The document is labeled arXiv:2608.25485v1 and dated 26 Aug 2026.
A very small labeled foothold
The paper asks whether semi-supervised recursive label mining can adapt CLIP to remote-sensing scene classification when annotated samples are scarce. The proposed system, called SE-CLIP, uses the labeled examples as a starting point and tries to expand its support set with similarity-ranked images.
The evaluation used two benchmark datasets. UCM contains 2,100 aerial images in 21 land-use categories. NWPU contains 31,500 remote-sensing images across 45 classes, with 700 images per class. Images were resized to 256 by 256 pixels. For each dataset, 80% of the data was used for training, five seeds per class were labeled, the remaining training images were left unlabeled, and 20% was held out for testing. Accuracy was the metric, and results were reported as averages across five runs with different random seeds.
SE-CLIP adapts CLIP with LoRA, a parameter-efficient low-rank update, and fixed text anchors for the classes. It uses a cross-entropy alignment objective based on temperature-scaled cosine similarities between image and text representations.
Training begins with a 10-iteration warm-up on expert seeds. After that, the system ranks unlabeled samples by similarity within each class, migrates selected samples into the support set, and appends them without re-evaluating them. In ordinary terms, it uses high-scoring unlabeled images as additional pseudo-labeled examples in later training.
Tuning the mining process
The mining budget, k, changed the outcome. With LoRA rank r set to 4, UCM reached its best tested accuracy, 98.81%, at k = 1. NWPU moved from 88.56% at k = 1 to 94.92% at k = 5, then slipped slightly to 94.77% at k = 7. In other words, the smaller benchmark favored a very small budget, while the larger one improved further before leveling off. These figures are means with standard deviations across five runs; confidence intervals were not reported.
Changing LoRA's rank also altered the reported endpoints. From r = 2 to r = 16, accuracy rose from 98.71% to 99.09% on UCM and from 94.41% to 95.07% on NWPU. The paper therefore found higher accuracy at the higher tested rank on both benchmarks, but this remains a sensitivity finding within the tested range.
What the ablations revealed
An ablation comparison showed a large difference between the full procedure and warm-up-only training. Full recursive discovery reached 99.09% on UCM and 95.07% on NWPU. The warm-up-only version reached 87.81% and 76.34%, respectively. The comparison shows that the initial labeled-seed warm-up alone did not match the full procedure under these settings.
Per-class balanced mining also recorded substantially higher accuracy than global mining. The balanced version reached 98.62% on UCM and 94.92% on NWPU, compared with 83.19% and 42.33% for global mining. The text also reports a strict standard deviation of zero across optimization iterations for local selection.
Training curves showed accuracy rising after top-k mining began following the 10-iteration warm-up, then plateauing around iteration 60 on UCM and between 160 and 170 on NWPU. t-SNE plots, which visualize the arrangement of image embeddings, showed tighter and better-separated semantic clusters for SE-CLIP than for zero-shot CLIP. The cluster finding is qualitative because no numerical cluster-separation measure was reported.
The paper reports a relatively small trainable footprint for the r = 16 LoRA configuration: 1,769,472 trainable parameters, or 1.16% of the model, with 1.12 GB of GPU memory at batch size 16. It reports inference latency of 10.26 milliseconds per batch. These resource measurements have no repeated-run uncertainty in the reported analysis.
Where the evidence stops
High scores on two benchmarks do not answer how the method would fare elsewhere. The evaluation used only UCM and NWPU, so generalization to other datasets, sensors, or domains was not shown. The authors also say the approach needs expressive text vocabularies to form discriminative anchors and remains limited on fine-grained scenes with high inter-class structural similarity.
The reported standard deviations capture variation across the five random-seed runs, but confidence intervals were not given. The comparison with DARP likewise does not report standard deviations or confidence intervals, so the size of those gaps remains tied to the stated benchmark protocol.
The authors conclude that target-directed low-rank adaptation of CLIP is a more effective alternative to traditional unimodal classifiers for complex remote-sensing data distributions. On the evidence supplied here, that conclusion is bounded by the UCM and NWPU tests and by the method's difficulty with fine-grained, visually similar classes.
The acknowledgment says the research was supported by the Ongoing Research Funding program, ORF-2026-995, at King Saud University in Riyadh, Saudi Arabia.
Paper data and sources
Original title: Semi-Supervised Adaptation of Vision-Language Models for Image Classification
Authors: Mohamed L. Mekhalfi, Mohamad M. Al Rahhal, Yakoub Bazi et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text