Remote sensing image super-resolution (SR) refers to a technique improving the spatial resolution, which in turn benefits to the subsequent image interpretation, e.g., target recognition, classification, and change detection. In popular sparse representation-based methods, due to the complex imaging conditions and unknown degradation process, the sparse coefficients of low-resolution (LR) observed images are hardly consistent with the real high-resolution (HR) counterparts, which leads to unsatisfactory SR results. To address this problem, a novel coupled sparse autoencoder (CSAE) is proposed in this paper to effectively learn the mapping relation between the LR and HR images. Specifically, the LR and HR images are first represented by a set of sparse coefficients, and then, a CSAE is established to learn the mapping relation between them. Since the proposed method leverages the feature representation ability of both sparse decomposition and CSAE, the mapping relation between the LR and HR images can be accurately obtained. Experimentally, the proposed method is compared with several state-of-the-art image SR methods on three real-world remote sensing image datasets with different spatial resolutions. The extensive experimental results demonstrate that the proposed method has gained solid improvements in terms of average peak signal-to-noise ratio and structural similarity measurement on all of the three datasets. Moreover, results also show that with larger upscaling factors, the proposed method achieves more prominent performance than the other competitive methods.

Timely, spatially complete, and reliable forest above-ground biomass (AGB) data are a prerequisite to support forest management and policy formulation. Traditionally, forest AGB is spatially estimated by integrating satellite images, in particular, optical data, with field plots from forest inventory programs. However, field data are limited in remote and unmanaged areas. In addition, optical reflectance usually saturates at high-density biomass level and is subject to cloud contaminations. Thus, this study aimed to develop a deep learning based workflow for mapping forest AGB by integrating Landsat 8 and Sentinel-1A images with airborne light detection and ranging (LiDAR) data. A reference AGB map was derived from the wall-to-wall LiDAR data and field measurements. The LiDAR plots—stratified random samples of forest biomass extracted from the LiDAR simulated strips in the reference map—were adopted as a surrogate for traditional field plots. In addition to the deep learning model, i.e., stacked sparse autoencoder network (SSAE), five different prediction techniques including multiple stepwise linear regressions, K-nearest neighbor, support vector machine, back propagation neural networks, and random forest were individually used to establish the relationship between LiDAR-derived forest biomass and the satellite predictors. Optical variables (Landsat 8 OLI), SAR variables (Sentinel-1A), and their combined variables were individually input to the six prediction models. Results showed that the SSAE model had the best performance for the forest biomass estimation. The combined optical and microwave dataset as explanatory variables improved the modeling performance compared to either the optical-only or microwave-only data, regardless of prediction algorithms. The best mapping accuracy was obtained by the SSAE model with inputs of optical and microwave integrated metrics that yielded $R^{2}$ of 0.812, root mean squared error (RMSE) of 21.753 Mg/ha, and relative RMSE (RMSE$_{r}$) of 14.457%. Overall, the SSAE model with inputs of combined Landsat 8 OLI and Sentinel-1A information could result in accurate estimation of forest biomass by using the stratification-sampled and LiDAR-derived AGB as ground reference data. The modeling workflow has the potential to promote future forest growth monitoring and carbon stock assessment across large areas.

Super-resolution for satellite video attaches much significance to earth observation accuracy, and the special imaging and transmission conditions on the video satellite pose great challenges to this task. The existing deep convolutional neural-network-based methods require pre-processing or post-processing to be adapted to a high-resolution size or pixel format, leading to reduced performance and extra complexity. To this end, this paper proposes a five-layer end-to-end network structure without any pre-processing and post-processing, but imposes a reshape or deconvolution layer at the end of the network to retain the distribution of ground objects within the image. Meanwhile, we formulate a joint loss function by combining the output and high-dimensional features of a non-linear mapping network to precisely learn the desirable mapping relationship between low-resolution images and their high-resolution counterparts. Also, we use satellite video data itself as a training set, which favors consistency between training and testing images and promotes the method’s practicality. Experimental results on “Jilin-1” satellite video imagery show that this method demonstrates a superior performance in terms of both visual effects and measure metrics over competing methods.

Point cloud classification is quite challenging due to the influence of noise, occlusion, and the variety of types and sizes of objects. Currently, most methods mainly focus on subjectively designing and extracting features. However, the features rely on prior knowledge, and it is also difficult to accurately characterize the complex objects of point clouds. In this paper, we propose a concise multi-scale convolutional network (MSNet) for adaptive and robust point cloud classification. Both the local feature and global context are incorporated for this purpose. First, around each point, the spatial contexts of different sizes are partitioned as voxels of different scales. A voxel-based MSNet is then simultaneously applied at multiple scales to adaptively learn the discriminative local features. The class probability of a point cloud is predicted by fusing the features together across multiple scales. Finally, the predicted class probabilities of MSNet are optimized globally using the conditional random field (CRF) with a spatial consistency constraint. The proposed method was tested with data sets of mobile laser scanning (MLS), terrestrial laser scanning (TLS), and airborne laser scanning (ALS) point clouds. The experimental results show that the proposed method was able to achieve appreciable classification accuracies of 83.18%, 98.24%, and 97.02% on the MLS, TLS, and ALS data sets, respectively. The results also demonstrate that the proposed network has a strong generalization capability for classifying different kinds of point clouds under complex urban environments.

Cloud detection of remote sensing imagery is quite challenging due to the influence of complicated underlying surfaces and the variety of cloud types. Currently, most of the methods mainly rely on prior knowledge to extract features artificially for cloud detection. However, these features may not be able to accurately represent the cloud characteristics under complex environment. In this paper, we adopt an innovative model named Fuzzy Autoencode Model (FAEM) to integrate the feature learning ability of stacked autoencode networks and the detection ability of fuzzy function for highly accurate cloud detection on remote sensing imagery. Our proposed method begins by selecting and fusing spectral, texture, and structure information. Thereafter, the proposed technique established a FAEM to learn the deep discriminative features from a great deal of selected information. Finally, the learned features are mapped to the corresponding cloud density map with a fuzzy function. To demonstrate the effectiveness of the proposed method, 172 Landsat ETM+ images and 25 GF-1 images with different spatial resolutions are used in this paper. For the convenience of accuracy assessment, ground truth data are manually outlined. Results show that the average RER (ratio of right rate and error rate) on Landsat images is greater than 29, while the average RER of Support Vector Machine (SVM) is 21.8 and Random Forest (RF) is 23. The results on GF-1 images exhibit similar performance as Landsat images with the average RER of 25.9, which is much higher than the results of SVM and RF. Compared to traditional methods, our technique has attained higher average cloud detection accuracy for either different spatial resolutions or various land surfaces.

The objective of photogrammetry is to extract information from imagery. With the increasing interaction of sensing and computing technologies, the fundamentals of photogrammetry have undergone an evolutionary change in the past several decades. Numerous theoretical progresses and practical applications have been reported from traditionally different but related multiple disciplines, including computer vision, photogrammetry, computer graphics, pattern recognition, remote sensing and machine learning. This has gradually extended the boundary of traditional photogrammetry in both theory and practice. This paper introduces a new, holistic theoretical framework to describe various photogrammetric tasks and solutions. Under this framework, photogrammetry is generally regarded as a reversed imaging process formulated as a unified optimization problem. Depending on the variables to be determined through optimization, photogrammetric tasks are mostly divided into image space tasks, image-object space tasks and object space tasks, each being a special case of the general formulation. This paper presents representative solution approaches for each task. With this effort, we intend to advocate an imminent and necessary paradigm change in both research and learning of photogrammetry.