Semi-Supervised Building Footprint Generation With Feature and Output Consistency Training

Mapping dispersed houses in rural areas of Colombia by exploiting planet satellite images with convolutional neural networks

“…Moreover, the use of a semantic segmentation approach with VHR images also makes it challenging to apply this methodology to map large spatial extensions in a reasonable amount of time. In this 3 research works 8,16,17 , they used PlanetScope imagery to map individual houses. In 8 , Gordana and Onur explored a high spatial resolution PlanetScope dataset for extracting houses inventory information in order to prevent serious damages and to determine the possible loses from earthquakes.…”

“…The approach aimed to overcome the challenges of existing methods, such as limited contextual information. Li et al 17 proposed a semi-supervised approach to houses footprint generation that uses features and outputs consistency training. The method aimed to improve the accuracy of houses footprint generation while reducing the need for manual labeling of training data.…”

A Residual-Inception U-Net (RIU-Net) Approach and Comparisons with U-Shaped CNN and Transformer Models for Building Segmentation from High-Resolution Satellite Images

“…As an important feature in the urban environment, the mapping of buildings has significant importance for different applications such as urban mapping, population estimation, land cover/land use analysis, cadastral and topographic map production, change detection, and disaster management [ 3 , 4 , 5 , 6 ]. Nonetheless, obtaining reliable and accurate building maps from high-resolution satellite images is still challenging due to various reasons such as complex backgrounds [ 7 , 8 ], similarities between the background and the buildings [ 9 ], noise in data [ 4 ], heterogeneity in data structures [ 7 , 8 ], diversity in roof types [ 10 ], and characteristics (size, shape, color, etc.) of buildings [ 11 ], and other topological difficulties.…”

“…If the problems encountered according to the methods used are briefly mentioned, the conventional methods that were used in the early studies generally use manually extracted features. In these models, the extraction process requires prior knowledge, which leads to a poor model generalization ability and is costly and time-consuming [ 7 ]. Deep learning methods that later replaced these methods in the following years have their own concerns.…”

“…Over the years, deep learning methods have gradually replaced conventional methods in image segmentation studies and achieved breakthroughs [ 13 ]. With the rapid development in computing powers and the increasing availability of data sources, deep learning methods, particularly CNNs, have started to be widely used as they surpass the traditional methods in terms of efficiency and accuracy [ 7 ]. Different from the conventional methods, CNNs have the capacity to extract features directly from the inputs and make predictions using sequential convolutions with fully connected layers [ 4 ].…”