Scene classification from remote sensing images provides new possibilities for potential application of high spatial resolution imagery. How to efficiently implement scene recognition from high spatial resolution imagery remains a significant challenge in the remote sensing domain. Recently, convolutional neural networks (CNN) have attracted tremendous attention because of their excellent performance in different fields. However, most works focus on fully training a new deep CNN model for the target problems without considering the limited data and time-consuming issues. To alleviate the aforementioned drawbacks, some works have attempted to use the pretrained CNN models as feature extractors to build a feature representation of scene images for classification and achieved successful applications including remote sensing scene classification. However, existing works pay little attention to exploring the benefits of multilayer features for improving the scene classification in different aspects. As a matter of fact, the information hidden in different layers has great potential for improving feature discrimination capacity. Therefore, this paper presents a fusion strategy for integrating multilayer features of a pretrained CNN model for scene classification. Specifically, the pretrained CNN model is used as a feature extractor to extract deep features of different convolutional and fully connected layers; then, a multiscale improved Fisher kernel coding method is proposed to build a mid-level feature representation of convolutional deep features. Finally, the mid-level features extracted from convolutional layers and the features of fully connected layers are fused by a principal component analysis/spectral regression kernel discriminant analysis method for classification. For validation and comparison purposes, the proposed approach is evaluated via experiments with two challenging high-resolution remote sensing data sets, and shows the competitive performance compared with fully trained CNN models, fine-tuning CNN models, and other related works.

To investigate the performance of extreme gradient boosting (XGBoost) in remote sensing image classification tasks, XGBoost was first introduced and comparatively investigated for the spectral-spatial classification of hyperspectral imagery using the extended maximally stable extreme-region-guided morphological profiles (EMSER_MPs) proposed in this study. To overcome the potential issues of XGBoost, meta-XGBoost was proposed as an ensemble XGBoost method with classification and regression tree (CART), dropout-introduced multiple additive regression tree (DART), elastic net regression and parallel coordinate descent-based linear regression (linear) and random forest (RaF) boosters. Moreover, to evaluate the performance of the introduced XGBoost approach with different boosters, meta-XGBoost and EMSER_MPs, well-known and widely accepted classifiers, including support vector machine (SVM), bagging, adaptive boosting (AdaBoost), multi class AdaBoost (MultiBoost), extremely randomized decision trees (ExtraTrees), RaF, classification via random forest regression (CVRFR) and ensemble of nested dichotomies with extremely randomized decision tree (END-ERDT) methods, were considered in terms of the classification accuracy and computational efficiency. The experimental results based on two benchmark hyperspectral data sets confirm the superior performance of EMSER_MPs and EMSER_MPs with mean pixel values within region (EMSER_MPsM) compared to that for morphological profiles (MPs), morphological profile with partial reconstruction (MPPR), extended MPs (EMPs), extended MPPR (EMPPR), maximally stable extreme-region-guided morphological profiles (MSER_MPs) and MSER_MPs with mean pixel values within region (MSER_MPsM) features. The proposed meta-XGBoost algorithm is capable of obtaining better results than XGBoost with the CART, DART, linear and RaF boosters, and it could be an alternative to the other considered classifiers in terms of the classification of hyperspectral images using advanced spectral-spatial features, especially from generalized classification accuracy and model training efficiency perspectives.

Land Use and Land Cover (LULC) classification is vital for environmental and ecological applications. Sentinel-2 is a new generation land monitoring satellite with the advantages of novel spectral capabilities, wide coverage and fine spatial and temporal resolutions. The effects of different spatial resolution unification schemes and methods on LULC classification have been scarcely investigated for Sentinel-2. This paper bridged this gap by comparing the differences between upscaling and downscaling as well as different downscaling algorithms from the point of view of LULC classification accuracy. The studied downscaling algorithms include nearest neighbor resampling and five popular pansharpening methods, namely, Gram-Schmidt (GS), nearest neighbor diffusion (NNDiffusion), PANSHARP algorithm proposed by Y. Zhang, wavelet transformation fusion (WTF) and high-pass filter fusion (HPF). Two spatial features, textural metrics derived from Grey-Level-Co-occurrence Matrix (GLCM) and extended attribute profiles (EAPs), are investigated to make up for the shortcoming of pixel-based spectral classification. Random forest (RF) is adopted as the classifier. The experiment was conducted in Xitiaoxi watershed, China. The results demonstrated that downscaling obviously outperforms upscaling in terms of classification accuracy. For downscaling, image sharpening has no obvious advantages than spatial interpolation. Different image sharpening algorithms have distinct effects. Two multiresolution analysis (MRA)-based methods, i.e., WTF and HFP, achieve the best performance. GS achieved a similar accuracy with NNDiffusion and PANSHARP. Compared to image sharpening, the introduction of spatial features, both GLCM and EAPs can greatly improve the classification accuracy for Sentinel-2 imagery. Their effects on overall accuracy are similar but differ significantly to specific classes. In general, using the spectral bands downscaled by nearest neighbor interpolation can meet the requirements of regional LULC applications, and the GLCM and EAPs spatial features can be used to obtain more precise classification maps.

Automated extraction of buildings from earth observation (EO) data has long been a fundamental but challenging research topic. Combining data from different modalities (e.g., high-resolution imagery (HRI) and light detection and ranging (LiDAR) data) has shown great potential in building extraction. Recent studies have examined the role that deep learning (DL) could play in both multimodal data fusion and urban object extraction. However, DL-based multimodal fusion networks may encounter the following limitations: (1) the individual modal and cross-modal features, which we consider both useful and important for final prediction, cannot be sufficiently learned and utilized and (2) the multimodal features are fused by a simple summation or concatenation, which appears ambiguous in selecting cross-modal complementary information. In this paper, we address these two limitations by proposing a hybrid attention-aware fusion network (HAFNet) for building extraction. It consists of RGB-specific, digital surface model (DSM)-specific, and cross-modal streams to sufficiently learn and utilize both individual modal and cross-modal features. Furthermore, an attention-aware multimodal fusion block (Att-MFBlock) was introduced to overcome the fusion problem by adaptively selecting and combining complementary features from each modality. Extensive experiments conducted on two publicly available datasets demonstrated the effectiveness of the proposed HAFNet for building extraction.

Accurate extraction of buildings using high spatial resolution imagery is essential to a wide range of urban applications. However, it is difficult to extract semantic features from a variety of complex scenes (e.g., suburban, urban and urban village areas) because various complex man-made objects usually appear heterogeneous with large intra-class and low inter-class variations. The automatic extraction of buildings is thus extremely challenging. The fully convolutional neural networks (FCNs) developed in recent years have performed well in the extraction of urban man-made objects due to their ability to learn state-of-the-art features and to label pixels end-to-end. One of the most successful FCNs used in building extraction is U-net. However, the commonly used skip connection and feature fusion refinement modules in U-net often ignore the problem of feature selection, and the ability to extract smaller buildings and refine building boundaries needs to be improved. In this paper, we propose a trainable chain fully convolutional neural network (CFCN), which fuses high spatial resolution unmanned aerial vehicle (UAV) images and the digital surface model (DSM) for building extraction. Multilevel features are obtained from the fusion data, and an improved U-net is used for the coarse extraction of the building. To solve the problem of incomplete extraction of building boundaries, a U-net network is introduced by chain, which is used for the introduction of a coarse building boundary constraint, hole filling, and "speckle" removal. Typical areas such as suburban, urban, and urban villages were selected for building extraction experiments. The results show that the CFCN achieved recall of 98.67%, 98.62%, and 99.52% and intersection over union (IoU) of 96.23%, 96.43%, and 95.76% in suburban, urban, and urban village areas, respectively. Considering the IoU in conjunction with the CFCN and U-net resulted in improvements of 6.61%, 5.31%, and 6.45% in suburban, urban, and urban village areas, respectively. The proposed method can extract buildings with higher accuracy and with clearer and more complete boundaries.

In this article, the graphics processing unit (GPU)-accelerated CatBoost (GPU-CatBoost) algorithm for hyperspectral image classification is first introduced and comparatively studied using diverse features. To further foster the classification performance from both accurate and efficient viewpoints, an ensemble version of GPU-CatBoost, the GPU-accelerated CatBoost-Forest (GPU-CatBF) algorithm, is proposed by adopting the parallelized minimum redundancy maximum relevance (mRMR) ensemble (PmRMRE) feature selection (FS) algorithm. To evaluate the performance and suitability of mRMR and PmRMRE, 11 other state-of-the-art FS algorithms are comprehensively investigated. Experimental results on three widely acknowledged hyperspectral benchmarks showed that: 1) GPU-CatBoost is also an advanced ensemble learning (EL) algorithm for hyperspectral image classification using diverse features; 2) mRMR and PmRMRE have advanced properties for highly discriminative features and band selection, and the best results are achieved by PmRMRE in most cases in terms of both the robustness and computational efficiency; and 3) GPU-CatBF always outperforms CatBoost and GPU-CatBoost, while compatible and even better results are reachable without losing much computational efficiency in contrast with other selected decision tree-based EL algorithms.