Semi-supervised Deep Learning via Transformation Consistency Regularization for Remote Sensing Image Semantic Segmentation

Deep convolutional neural networks have gotten a lot of press in the last several years, especially domains like computer vision and remote sensing (RS). However, achieving superior performance with deep highly depends on massive number accurately labeled training samples. In real-world applications, gathering large samples is time consuming labor intensive, for pixel-level data annotation. This dearth labels land-cover classification pressing RS domain because high-precision high-quality are extremely difficult to acquire, but unlabeled readily available. this study, we offer new semisupervised semantic labeling framework segmentation high-resolution images take advantage limited amount examples numerous Our model uses transformation consistency regularization encourage consistent network predictions under different random transformations or perturbations. We try three transforms compute loss analyze their performance. Then, present technique by using hybrid regularization. A weighted sum losses, which contains supervised term computed an unsupervised data, may be used update parameters our technique. comprehensive experiments two datasets confirmed that suggested approach utilized latent information from obtain more precise outperformed existing algorithms terms further demonstrated strategy has potential partially tackle problem image segmentation.