Convolutional neural network (CNN) has demonstrated impressive ability to represent hyperspectral images and achieve promising results in image classification. However, traditional CNN models can only operate convolution on regular square regions with fixed size weights, thus, they cannot universally adapt the distinct local various object distributions geometric appearances. Therefore, their classification performances are still be improved, especially class boundaries. To alleviate this...

In the literature of remote sensing, deep models with multiple layers have demonstrated their potentials in learning abstract and invariant features for better representation classification hyperspectral images. The usual supervised models, such as convolutional neural networks, need a large number labeled training samples to learn model parameters. However, real-world image task provides only limited samples. This paper adopts another popular model, i.e., belief networks (DBNs), deal this...

Machine learning methods have achieved good performance and been widely applied in various real-world applications. They can learn the model adaptively be better fit for special requirements of different tasks. Generally, a machine system is composed plentiful training data, process, an accurate inference. Many factors affect among which diversity process important one. The help each procedure to guarantee total learning: data ensures that provide more discriminative information model,...

Recently, researchers have shown the powerful ability of deep methods with multilayers to extract high-level features and obtain better performance for hyperspectral image classification. However, a common problem traditional models is that learned might be suboptimal because limited number training samples, especially large intraclass variance low interclass variance. In this paper, novel convolutional neural networks (CNNs) multiscale convolution (MS-CNNs) are proposed address by...

In hyperspectral image (HSI) classification, spatial context has demonstrated its significance in achieving promising performance. However, conventional context-based methods simply assume that spatially neighboring pixels should correspond to the same land-cover class, so they often fail correctly discover contextual relations among complex situations, and thus leading imperfect classification results on some irregular or inhomogeneous regions such as class boundaries. To address this...

With complex building composition and imaging condition, urban areas show versatile characteristics in remote sensing optical images. It demonstrates that multiple features should be utilized to characterize areas. On the other hand, since levels of development neighboring are not statistically independent, each area site depend on those sites. In this paper, we present a conditional random fields (CRFs) ensemble model incorporate learn their contextual information. This involves two...

Hyperspectral images exhibit strong dependencies across spatial and spectral neighbors, which have been proved to be very useful for hyperspectral image classification. State-of-the-art classification algorithms use the in a heuristic way or probabilistic frameworks but impose unreasonable assumptions on observed data. In this paper, we formulate conditional random field (CRF) replace such heuristics of images. Moreover, because avoiding explicit modeling data, proposed method can...

Deep models with multiple layers have demonstrated their potential in learning abstract and invariant features for better representation classification of remote sensing images. Moreover, metric (ML) is usually introduced into the deep to further increase discrimination representations. However, usual ML methods treat training samples each batch stochastic gradient descent-based procedure independently, thus, they neglect important contextual (structural) information samples. In this paper,...

Although 3D point cloud classification neural network models have been widely used, the in-depth interpretation of activation neurons and layers is still a challenge. We propose novel approach, named Relevance Flow, to interpret hidden semantics networks. It delivers class activated in intermediate back-propagation manner, associates with input points visualize each layer. Specially, we reveal that has learned plane-level part-level layers, utilize normal IoU evaluate consistency both...

Denoising of hyperspectral imagery in the domain imaging spectroscopy by conditional random fields (CRFs) is addressed this work. For denoising imagery, strong dependencies across spatial and spectral neighbors have been proved to be very useful. Many available image algorithms adopt multidimensional tools deal with problems thus naturally focus on use dependencies. However, few them were specifically designed In paper, we propose a multiple-spectral-band CRF (MSB-CRF) simultaneously model...

Gaussian process (GP) classifiers represent a powerful and interesting theoretical framework for the Bayesian classification of hyperspectral images. However, collection labeled samples is time consuming costly data, training available are often not enough an adequate learning GP classifier. Moreover, computational cost performing inference using scales cubically with size set. To address limitations image classification, reducing label keeping set in moderate size, this paper introduces...

Despite much advance obtained in hyperspectral image sensors, they are still very sensitive to the noise, and thus cause captured data carry enough noise degrade classification results. The traditional approach first resorts denoising then feeds denoised into a classifier. However, such straightforward approach, treating separately, suffers greatly from neglecting their impacts on each other. This paper presents new simultaneous method pursuit of cleanest for optimal sense given task...

Nowadays, deep learning methods, especially the convolutional neural networks (CNNs), have shown impressive performance on extracting abstract and high-level features from hyperspectral image. However, general training process of CNNs mainly considers pixelwise information or samples' correlation to formulate penalization while ignores statistical properties spectral variability each class in These sample-based penalizations would lead uncertainty due imbalanced limited number samples. To...

Recently, graph convolutional network (GCN) has progressed significantly and gained increasing attention in hyperspectral image (HSI) classification due to its impressive representation power. However, existing GCN-based methods do not give full consideration the multiscale spatial information, since convolution operations are governed by fixed neighborhood. As a result, their performances can be limited, particularly regions with diverse land cover appearances. In this article, we develop...

Although deep learning has received extensive attention and achieved excellent performance in various scenarios, it suffers from adversarial examples to some extent. In particular, physical attack poses a greater threat than digital attack. However, existing research paid less the of object detection UAV remote sensing images (RSIs). this work, we carefully analyze universal patch for multi-scale objects field sensing. There are two challenges faced by an RSIs. On one hand, number is more...

Unsupervised learning of a convolutional neural network (CNN) is feasible method to represent and classify remote sensing images, where labeling the observed data prepare training samples highly expensive time-consuming task. In this letter, we propose an unsupervised feature fusion formulate easy-to-train but effective CNN representation images. The efficiency effectiveness are derived from following two aspects. First, proposed trains deep through each layer in greedy layer-wise manner,...

Hyperspectral images exhibit strong dependencies across spatial and spectral neighbors, which have been proved to be very useful for hyperspectral image classification. The recently defined conditional random field (CRF) can effectively model use the classification of in a unified probabilistic framework. However, order computationally tractable, usual CRFs are limited incorporate only pairwise potentials. Thus, capture interactions neglect higher dependencies, potentially high-level...

Hyperspectral image classification has attracted extensive research efforts in the recent decades. The main difficulty lies few labeled samples versus high dimensional features. spectral-spatial method using Markov random field (MRF) been shown to perform well improving performance. Moreover, active learning (AL), which iteratively selects most informative unlabeled and enlarges training set, widely studied proven useful remotely sensed data. In this paper, we focus on combination of MRF AL...

A practical hyperspectral target characterization task estimates a signature from imprecisely labeled training data. The imprecisions arise the characteristics of real-world tasks. First, accurate pixel-level labels on data are often unavailable. Second, subpixel targets and occluded cause samples to contain mixed multiple types. To address these imprecisions, this paper proposes new method produce diverse signatures under instance learning (MIL) framework. proposed uses only bag-level...