SeConDA : Self‐Training Consistency Guided Domain Adaptation for Cross‐Domain Remote Sensing Image Semantic Segmentation
ABSTRACT
Well‐trained remote sensing (RS) deep learning models often encounter a considerable decline in performance when applied to images that differ from the training data. This decline can be attributed to variations in imaging sensors, geographic location, imaging time, and radiation levels during image acquisition. Consequently, the widespread application of these models has been greatly impeded. An envisioned resolution to confront this challenge encompasses formulating a cross‐domain RS image semantic segmentation network integrated with self‐training consistency. This approach involves the generation of high‐quality pseudo‐labels for images in the target domain, which are then used to guide the training of the network. To enhance the model's ability to learn the data distributions of both the source and target domains, highly perturbed mixed samples are created by blending images from these domains. Additionally, adversarial training is incorporated to reduce the entropy of the model's predicted results, thereby mitigating the influence of noise present in the pseudo‐labels. As a result, this approach effectively extracts domain‐invariant features and minimizes the disparities between the distributions of the different domains. By employing the ISPRS and LoveDA datasets in a series of experiments conducted across varied scenarios, our empirical investigations evince the capacity of the proposed methodology to generalize the model to target domain data, which is achieved through the mitigation of disparities between domain distributions. It effectively alleviates the domain shift issues caused by differences in imaging locations and band combinations in RS image data and achieves state‐of‐the‐art results and validates its effectiveness.
