Semantic segmentation of high-resolution Unmanned Aerial Vehicle (UAV) remote sensing images plays a crucial role in environmental monitoring, urban planning, agricultural assessment, and disaster management. Semantic segmentation methods that are based on deep learning have demonstrated superior performance; however, they rely on large amounts of annotated data, and thus their performance significantly degrades in small-sample scenarios. To obtain better performance on small-scale remote sensing semantic segmentation datasets, methods combining knowledge distillation and semi-supervised learning are proposed. These methods use models pre-trained on large-scale natural image datasets (such as ImageNet) to guide the training of student models on target datasets directly, achieving significant performance gains. However, the feature distribution of natural image datasets differs significantly from that of remote sensing image datasets. Therefore, student models, directly guided by teacher models pre-trained on natural image datasets, often struggle to obtain the optimal performance, especially when few samples are labeled in the target domain. Whether introducing a medium-scale remote sensing dataset as an intermediate domain between natural image datasets and the target remote sensing dataset can further improve model performance is a question worth exploring. This study proposed a few-shot remote sensing image semantic segmentation method that combined multi-stage knowledge distillation (MKD) and semi-supervised learning (SSL) to progressively bridge domain gaps and leverage unlabeled data. The experimental results on the Erhai UAV dataset (EH) show that the proposed MKD + SSL method achieves a mean IoU of 77.05% with only 880 labeled samples, outperforming the widely used single-stage KD method by + 3.06% mIoU, with per-class IoU gains up to +(2.17% − 5.21%). On the Cityscapes benchmark, our framework further surpasses state-of-the-art methods such as UniMatch, achieving a + 1.5% and + 1.4% improvement in mIoU under 1/16 and 1/8 labeled settings, respectively. These results demonstrate that the proposed method effectively enhances segmentation accuracy in few-shot settings and generalizes well across diverse datasets, with wide practical value.

Camera traps are widely used in wildlife surveys but often generate numerous empty images due to false triggers, which are time-consuming to filter manually. Microsoft’s MegaDetector (MD) is commonly applied for empty image filtering; however, setting its confidence threshold is challenging because a high threshold increases the false negative rate (FNR), whereas a low threshold raises the false positive rate (FPR), making it difficult to achieve a satisfactory trade-off between the two. To address this dilemma, we proposed three MD-based methods with different performance and computational requirements. For FNR-sensitive users, we proposed MD_CS, which integrated context image similarity to achieve a very low FNR while controlling the FPR. For FPR-sensitive users, we proposed MD_CL, which used MD to generate supervisory signals for training a classification model, thereby reducing the FPR while controlling the FNR. For users requiring both low FNR and low FPR, we proposed MD_CS_CL, an ensemble method that combined the strengths of MD_CS and MD_CL. Experimental results on multiple datasets showed that these three methods achieved a significantly better balance between FNR and FPR compared to the threshold adjustment method of the MD method. Specifically, MD_CS maintained an FNR of 1.8–5.9% while controlling the FPR at 4.9–8.5%; MD_CL kept the FPR at 0.4–2.6% while controlling the FNR at 3.8–8.4%; MD_CS_CL achieved an optimal balance with an FNR of 1.4–4.0% and an FPR of 0.3–0.9%. On a server equipped with an RTX 4090 GPU, the three methods processed 6.7, 5.4, and 4.0 images per second, respectively, and none required manual annotation. Consequently, the proposed methods reduce labor costs, improve work efficiency, and provide reliable technical support for large-scale wildlife monitoring by achieving a superior FNR–FPR trade-off compared to MD’s traditional threshold regulation.