In this study, a deep learning (DL)-based approach is proposed for the detection and reconstruction of buildings from a single aerial image. The pre-required knowledge to reconstruct the 3D shapes of buildings, including the height data as well as the linear elements of individual roofs, is derived from the RGB image using an optimized multi-scale convolutional–deconvolutional network (MSCDN). The proposed network is composed of two feature extraction levels to first predict the coarse features, and then automatically refine them. The predicted features include the normalized digital surface models (nDSMs) and linear elements of roofs in three classes of eave, ridge, and hip lines. Then, the prismatic models of buildings are generated by analyzing the eave lines. The parametric models of individual roofs are also reconstructed using the predicted ridge and hip lines. The experiments show that, even in the presence of noises in height values, the proposed method performs well on 3D reconstruction of buildings with different shapes and complexities. The average root mean square error (RMSE) and normalized median absolute deviation (NMAD) metrics are about 3.43 m and 1.13 m, respectively for the predicted nDSM. Moreover, the quality of the extracted linear elements is about 91.31% and 83.69% for the Potsdam and Zeebrugge test data, respectively. Unlike the state-of-the-art methods, the proposed approach does not need any additional or auxiliary data and employs a single image to reconstruct the 3D models of buildings with the competitive precision of about 1.2 m and 0.8 m for the horizontal and vertical RMSEs over the Potsdam data and about 3.9 m and 2.4 m over the Zeebrugge test data.

Forest canopy cover (FCC) is an important ecological parameter of forest ecosystems, and is correlated with forest characteristics, including plant growth, regeneration, biodiversity, light regimes, and hydrological properties. Here, we present an approach of combining Sentinel-2 data, high-resolution aerial images, and machine learning (ML) algorithms to model FCC in the Hyrcanian mixed temperate forest, Northern Iran. Sentinel-2 multispectral bands and vegetation indices were used as variables for modeling and mapping FCC based on UAV ground truth to a wider spatial extent. Random forest (RF), support-vector machine (SVM), elastic net (ENET), and extreme gradient boosting (XGBoost) were the ML algorithms used to learn and generalize on the remotely sensed variables. Evaluation of variable importance indicated that vegetation indices including NDVI, NDVI-A, NDRE, and NDI45 were the dominant predictors in most of the models. Model accuracy estimation results showed that among the tested models, RF (R² = 0.67, RMSE = 18.87%, MAE = 15.35%) and ENET (R² = 0.63, RMSE = 20.04%, MAE = 16.44%) showed the best and the worst performance, respectively. In conclusion, it was possible to prove the suitability of integrating UAV-obtained RGB images, Sentinel-2 data, and ML models for the estimation of FCC, intended for precise and fast mapping at landscape-level scale.

Global Digital Elevation Models (GDEM) are considered very attractive for current research and application areas due to their free and wide range accessibility. The ASTER Global Digital Elevation Model exhibits the highest spatial resolution data of all global DEMs and it is generated for almost the whole globe. Unfortunately, ASTERGDEM data include many artifacts and height errors that decrease the quality and elevation accuracy significantly. This study provides a method for quality improvement of the ASTER GDEM data by correcting systematic height errors using ICESat laser altimetry data and removing artifacts and anomalies based on a segment-based outlier detection and elimination algorithm. Additionally, elevation errors within water bodies are corrected using a water mask produced from a high-resolution shoreline data set. Results indicate that the accuracy of the corrected ASTER GDEM is significantly improved and most artifacts are appropriately eliminated. Nevertheless, artifacts containing lower height values with respect to the neighboring ground pixels are not entirely eliminated due to confusion with some real non-terrain 3D objects. The proposed method is particularly useful for areas where other high quality DEMs such as SRTM are not available.

The expansion of the railway industry has increased the demand for the three-dimensional modeling of railway tracks. Due to the increasing development of UAV technology and its application advantages, in this research, the detection and 3D modeling of rail tracks are investigated using dense point clouds obtained from UAV images. Accordingly, a projection-based approach based on the overall direction of the rail track is proposed in order to generate a 3D model of the railway. In order to extract the railway lines, the height jump of points is evaluated in the neighborhood to select the candidate points of rail tracks. Then, using the RANSAC algorithm, line fitting on these candidate points is performed, and the final points related to the rail are identified. In the next step, the pre-specified rail piece model is fitted to the rail points through a projection-based process, and the orientation parameters of the model are determined. These parameters are later improved by fitting the Fourier curve, and finally a continuous 3D model for all of the rail tracks is created. The geometric distance of the final model from rail points is calculated in order to evaluate the modeling accuracy. Moreover, the performance of the proposed method is compared with another approach. A median distance of about 3 cm between the produced model and corresponding point cloud proves the high quality of the proposed 3D modeling algorithm in this study.

High resolution Digital Surface Models (DSMs) produced from airborne laser-scanning or stereo satellite images provide a very useful source of information for automated 3D building reconstruction. In this paper an investigation is reported about extraction of 3D building models from high resolution DSMs and orthorectified images produced from Worldview-2 stereo satellite imagery. The focus is on the generation of 3D models of parametric building roofs, which is the basis for creating Level Of Detail 2 (LOD2) according to the CityGML standard. In particular the building blocks containing several connected buildings with tilted roofs are investigated and the potentials and limitations of the modeling approach are discussed. The edge information extracted from orthorectified image has been employed as additional source of information in 3D reconstruction algorithm. A model driven approach based on the analysis of the 3D points of DSMs in a 2D projection plane is proposed. Accordingly, a building block is divided into smaller parts according to the direction and number of existing ridge lines for parametric building reconstruction. The 3D model is derived for each building part, and finally, a complete parametric model is formed by merging the 3D models of the individual building parts and adjusting the nodes after the merging step. For the remaining building parts that do not contain ridge lines, a prismatic model using polygon approximation of the corresponding boundary pixels is derived and merged to the parametric models to shape the final model of the building. A qualitative and quantitative assessment of the proposed method for the automatic reconstruction of buildings with parametric roofs is then provided by comparing the final model with the existing surface model as well as some field measurements.

Convolutional neural networks (CNNs) have demonstrated excellent performance in hyperspectral image (HSI) classification. However, tuning some critical hyperparameters of a CNN--such as the receptive field (RF) size--presents a major challenge due to the presence of features with different scales in HSIs. Contrary to the conventional design of CNNs, which fixes the RF size, it has been proven that the RF size is modulated by the stimulus and hence, depends on the scene being considered. Such a dilemma has been rarely considered in CNN design. In this letter, a new multibranch selective kernel network (MSKNet) is introduced, in which the input image is convolved using different RF sizes to create multiple branches so that the effect of each branch is adjusted by an attention mechanism according to the input contrast. As a result, our newly developed MSKNet is capable of modeling different scales. Our experimental results, conducted on three widely used HSIs, reveal that the MSKNet can outperform state-of-the-art CNNs in the context of HSI classification problems. The source code of our newly developed MSKNet is available from: https://github.com/mhaut/MSKNet-HSI