Spatial sampling to enhance meta learning for semantic segmentation of build-ups in remote sensing

We propose a novel approach to enhance semantic segmentation model adaptability for remote sensing applications by integrating spatial sampling strategies with meta learning frameworks. Our method leverages the Moran’s I statistic to characterize spatial distribution patterns (dispersed, random, or clustered) of samples and organizes the training data into M groups for Model-Agnostic Meta-Learning (MAML) with K -shot learning. This spatial sampling scheme addresses the important role of spatial distribution in training effectiveness and model optimization. We formulate meta-learning tasks for built-up area segmentation across diverse spatial distributions while employing random binary class labels to minimize overfitting and memorization. We evaluate our approach using pretrained SegFormer and SwinUNet architectures, comparing their performance against meta-learning variants and standard UNet on the Gaofen Image Datas (GID) dataset. The meta-learning UNet achieves a query Mean Intersection over Union (MIoU) of 0.75, while statistical analysis demonstrates that fine-tuning pretrained Meta UNet models yields robust performance with an MIoU of 0.80, compared to 0.76-0.78 for other models including pretrained visual Transformers. Visual analysis reveals that the pretrained Meta UNet produces more spatially coherent segmentation results with reduced fragmentation compared to baseline models. These findings highlight the potential of spatial sampling to enhance meta-learning approaches for semantic segmentation of land cover features in remotely sensed imagery. • Spatial patterns of training samples significantly impact deep model training. • Leveraged Moran’s I to group spatial patterns of samples for enhanced meta-learning. • Designed multiscale Moran convolutions to quantify spatial distributions of samples. • Spatially-aware sampling improves MAML generalization across spatial contexts. • Superior performance over vanilla Transformers in build-up segmentation tasks.