In the present paper, a new deep neural network model, named, UrbanScape-Net, has been designed for improved land-cover classification. Motivated by the need to emphasize important features and capture intricate relationships in land-use scenes, spatial and self-attention mechanisms with the Xception architecture are integrated. The spatial attention module focuses on salient regions, the self-attention method captures intricate relationships within these regions and the dynamic convolutional layer with varying filter sizes extracts multi scale features. The Xception model, pretrained on ImageNet, has been employed as the foundational feature extractor. The experiment conducted on the UC Merced Land Use dataset demonstrated enhanced accuracy (of 98.83%) across 21 classes compared to state-of-the-art methods.

Unmanned Aerial Vehicles (UAV) have become ubiquitous across multiple disciplines and applications. However, as of late, aerial platforms have been predominant in curating hyperspectral, target detection, benchmark data for the Geoscience and Remote Sensing (GRS) community. We have noticed that there is an absence of drone-based standardized data sets which our work intends to fill. Considering the complexity and cost associated with aerial system data collections, we set to create a new drone-based standardized hyperspectral image (HSI) test data set. The added benefit of using drones can be related to their highly flexible revisit rates, thus, more and more researchers are using drones to collect their imagery. The collected data presented was calibrated and processed to create high resolution hyperspectral imagery, ground truth target masks, field spectral measurements (of two target objects), etc. The HSI data were analyzed with popular target detection algorithms (ACE, MF and SAM) for reference and general assessment of data set complexity. The results of the target detection analysis is reported so as to provide the GRS community with reference performance metrics, such as receiver operating characteristics (ROC) curves, so researchers can benchmark their own target detection algorithms. The goal is to have the complete data set available through the GRS Society website.

Objects on Earth with 3-dimensional geometries cast shadows onto the underlying background due to the obstruction of direct solar radiation. This phenomenon is considered for a multiclass subpixel target detection scenario, where the mixed pixels consist of materials from varying percentages of an illuminated target and partially shaded background. Hyperspectral data collected from UAS-based instruments and novel subpixel targets were used for the analyses. The novel targets enabled empirical observations of fully illuminated targets on partially shaded backgrounds. To assist in developing inferences on the impacts of detection, model reflectance spectra were generated for mixed pixels with fully illuminated and fully shaded background conditions. The modeling approach was validated with the empirical observations. The results imply the shadow effect is an important system parameter for subpixel target percentages of 20% or less, for the multiclass scenario and particular targets explored.

Classification of hyperspectral images is computationally expensive due to the presence of large number of spectral bands. Therefore, dimensionality reduction using selection of optimal set of bands is an essential task to speed up the subsequent classification process. Bands must be selected in such a way so that they are as much independent as possible without sacrificing classification accuracy. In this context, a supervised band selection approach is proposed combining mutual and neighborhood information of bands. For classification purpose, Support Vector Machine classifier is used. Overall classification accuracy is considered to assess the efficiency of the proposed method. Performance of the proposed technique is compared with several other Mutual Information based methods and the proposed method is found to be better as compared to others.

This paper presents a method for identifying clusters in hyperspectral images. A new mutual information based similarity index named k-MIDI is proposed. This index, k-MIDI is used for hierarchical clustering of random samples from the data. The remaining observations are assigned to the clusters based on their nearest neighbours.

In the realm of hyperspectral sub-pixel target detection, the Linear Mixing Model (LMM) is an established basis for analysis and modelling. However, its accuracy depends on several key assumptions, most notably that there exists no nonlinear mixing within a scene. The Forecasting and Analysis of Spectroradiometric System Performance (FASSP) model utilizes the LMM to perform system requirement analyses. To quantify the limitations of the LMM when nonlinear effects are present, this paper reviews the results of September 2022 data collect in which the spectra of several sub-pixel target panels were altered by shadowing and reflectance panels and compares it with those from FASSP. Overall, both forms of nonlinear effects reduce target detection performance and the FASSP model tends to overestimate target radiance and target detection performance relative to the empirical results when nonlinear effects are present.

The Digital Imaging and Remote Sensing (DIRS) lab in the Chester F. Carlson Center for Imaging Science at the Rochester Institute of Technology (RIT) focuses on the development of tools to extract information about the Earth from aerial and satellite imaging systems. Recent focus has been in the area of drone-based hyperspectral imaging and data acquisition. A strong in-lab and in-field calibration capability has allowed for the precise characterization of many aspects of hyperspectral imaging (HSI) performance, particularly in the field (i.e., vicarious calibration). This paper discusses our innovative approaches to in-field calibration along with our observations of hyperspectral (HS) instruments and data processing effects related to radiometric calibration, the sampled point spread function (SPSF), and geo-rectification.

We propose a weakly-supervised, multitask framework for training a convolutional neural network to solve the problem of cloud shadow mitigation given only cloud and shadow masks as labels. The network minimizes the Wasserstein distance between shadows and their proximal sunlit neighborhoods, generating a supervisory signal directly from within the input image. We extract further utility from the shadow mask through multitask learning by introducing an auxiliary task of shadow segmentation. Our approach is advantageous since it performs mitigation in an end-to-end framework which requires only a shadowed image for inference. We apply this process to the Landsat 8 OLI SPARCS validation data set and demonstrate plausible results.

Artificial neural networks are designed for classic classification problem, which is different than our goal of target detection. The objective of this paper is to develop an algorithm, based on a one-layer neural network, and assess its performance and utility as a target detection algorithm to detect a subpixel target in a hyperspectral image. The weights are estimated by maximizing the likelihood function of the output variable and are solved numerically using the gradient descent method with a variable step size based on the Lipschitz's constant for the objective function. Experimental results using hyperspectral data are presented so as to assess the performance of the proposed algorithm. Results demonstrated that a single-layer neural network, implemented using the gradient descent method with a variable step size, can detect subpixel objects in hyperspectral imagery.

With the popular use of remote sensing techniques, investigations into hyperspectral system designs and parameter trade-off studies have become more and more necessary. Analytical models based on statistical descriptions and energy propagation are certainly efficient methods to examine a large number of parameter trades and sensitive studies with low computational cost. In long wave Infrared (LWIR), an analytical version of a temperature/emissivity separation (TES) algorithm can be used to retrieve ground emissivity statistics. However, such a statistical analytical algorithm has not been fully developed, as far as we know. In this letter, a new statistical iterative spectrally smooth temperature/emissivity separation (S-ISSTES) algorithmic approach is proposed. The derivation and comparison of our statistical approach is discussed in detail. We show that it can retrieve first- and second-order statistics of surface spectra as well as the associated temperature from at-sensor radiance data. Experimental results using both real and synthetic data demonstrate the effectiveness of the proposed S-ISSTES algorithm.

With the popularity of (hyperspectral) remote sensing systems coupled with a myriad of applications, comes the need for investigations into hyperspectral system designs and parameter trade-off studies. Analytical models based on statistical descriptions and signal propagation are efficient methods to examine these parameter trade-off studies, as well as sensitivity studies, with low computational cost. In this paper, a newly developed long wave infrared (LWIR) statistical iterative spectrally smooth temperature/emissivity separation (S-ISSTES) algorithm has been integrated into a widely used full spectrum hyperspectral remote sensing system model known as the forecasting and analysis of spectroradiometric system performance (FASSP) model. This new tool now allows users to performLWIRfull system (i.e., from surface reflectance, to sensor, to retrieve emissivity, to detection analysis) trade studies. In this paper, we validate the LWIR model and detection performance of the new FASSP model followed by illustrating the usage of the full system model by demonstrating trade examples including useful parameter trade studies and subpixel detection sensitivity studies.

Defining resolution within satellite imagery is normally achieved through the observation of edge targets or is visually graded (e.g., National Imagery Interpretability Rating Scale (NIIRS)) for the level of detail observed. These methods are significantly disadvantaged by not directly measuring fundamental quantities related to the imaging system. Recently, ground mirror-based systems have been developed which can mimic an ideal point source observable by satellite systems allowing direct observation of an imaging system point response function (PRF). This fundamental quantity of an imaging system defines the end-to-end performance of the optics and detector. In this paper, we illustrate the use of the PRF in a new approach called the point-pair resolution technique (PPRT) which characterizes separability between two ideal point sources. We compare real and simulated point-pairs to demonstrate validity.

We demonstrate an enhanced ACE target detection algorithm for poorly illuminated and shadowed pixels in which image segmentation based on illumination is performed prior to target detection. This enhanced ACE detection method (ACE-shadow) relies on FLAASH-based scene atmospheric correction and MODTRAN-calculated direct to diffuse illumination ratios to model shadowed target spectra. Improvements to target detection were realized using our ACE-shadow detection algorithm because SNR characteristics of shadowed pixels dominate the segmented covariance.

We investigate applying convolutional neural network (CNN) architecture to facilitate aerial hyperspectral scene understanding and present a new hyperspectral data set, AeroRIT, which is large enough for CNN training. To date, the majority of hyperspectral airborne has been confined to various subcategories of vegetation and roads, and this scene introduces two new categories: buildings and cars. To the best of our knowledge, this is the first comprehensive large-scale hyperspectral scene with nearly seven-million pixel annotations for identifying cars, roads, and buildings. We compare the performance of the three popular architectures-SegNet, U-Net, and Res-U-Net, for scene understanding and object identification via the task of dense semantic segmentation to establish a benchmark for the scene. To further strengthen the network, we add squeeze and excitation blocks for better channel interactions and use self-supervised learning for better encoder initialization. Aerial hyperspectral image analysis has been restricted to small data sets with limited train/test splits capabilities, and we believe that AeroRIT will help advance the research in the field with a more complex object distribution to perform well on. The full data set, with flight lines in radiance and reflectance domains, is available for download at https://github.com/aneesh3108/AeroRIT. This data set is the first step toward developing robust algorithms for hyperspectral airborne sensing that can robustly perform advanced tasks such as vehicle tracking and occlusion handling.

This paper introduces a new and challenging hyperspectral data set to the remote sensing community called the “RIT2017 Data Set” which can be used for the assessment of target detection algorithms. This data set encompasses 90 targets in a background of up to 8 million pixels (or less if sub-setting). The same data set can also be used for atmospheric compensation studies for it has identical sets of large panels in both the sun and full shadow. This paper briefly introduces the data collection campaign, the target objects, and addresses the radiometric fidelity of the imaging spectrometer data, which showed very good results. Lastly, the data is atmospherically compensated using an in-scene technique, which also showed fairly good results.

Deep neural networks provide deep extracted features for image classification. As a high dimension data, hyperspectral image (HSI) feature extraction is unlike an RGB image whose feature representation could not be simply generated in the spatial domain. To take full advantage of HSI, a dual-channel convolutional neural network (CNN) is applied, 1D convolution for the spectral domain and 2D convolution for spatial domain. For pixel-wise classification of HSI, in our network model, one-dimensional customized DenseNet is for extracting the hierarchical spectral features and another customized DenseNet is applied to extract the hierarchical spatial-related feature. Furthermore, we experimentally tuned the several widen factors and dense-net growth rates to evaluate the impact of hyper-parameter. To compare our proposed method with HSI classification methods, we test other three DNNs based method in two real-world HSI dataset. The result demonstrated our approach outperformed the state-of-art method.

This paper examines the potential impacts of remote sensing, hyperspectral target detection from a physics-based modeling point of view. Often (atmospherically compensated) data is simply handed off to algorithm developers with the assumption that the data they are given accurately represents what transpired at the time of collection. In this paper, we discuss the various steps involved in processing raw collected data to a final product that can be utilized in target detection sceneries. We comment on where spectral content can be altered and discuss the physical modeling involved in atmospheric compensation and forward modeling. Impacts on detection performance are shown by employing a matched filter and hyperspectral imagery. Results show that having an incomplete model will alter detection results, for a given detection algorithm.

Accounting for radiative transfer within the atmosphere is usually necessary to accomplish target detection in airborne/satellite hyperspectral images. In this paper, two methods of accounting for the illumination and atmospheric effects-atmospheric compensation (AC) and forward modeling (FM)-are investigated in their application to target detection. Specifically, several crucial aspects are examined, such as the processing required, the computational complexity, and the flexibility accorded to an imperfect knowledge of acquisition conditions. Real ground-truthed hyperspectral data are employed in order to evaluate the operational applicability of such approaches in a target-detection scenario, as well as their impact on the processing-chain computational complexity. Results indicate that AC is recommended when accurate knowledge of the acquisition conditions is available, and the image has relatively uniform illumination and nonshadowed targets. Conversely, FM is preferred if scene conditions are not well known and when the targets may be subject to varying illumination conditions, including shadowing.

This paper develops (and applies) a hybrid target detector that incorporates structured backgrounds and physics based modeling together with a geometric infeasibility metric. More often than not, detection algorithms are usually applied to atmospherically compensated hyperspectral imagery. Rather than compensate the imagery, we take the opposite approach by using a physics based model to generate permutations of what the target might look like as seen by the sensor in radiance space. The development and status of such a method is presented and applied to the generation of target spaces. The generated target spaces are designed to fully encompass image target pixels while using a limited number of input model parameters. Additionally, a structured infeasibility projector (SIP) is developed which enables one to be more selective in rejecting false alarms. Results on HYDICE data show that the SIP algorithm, in conjunction with a physics based detector, outperforms results from the SAM and SMF algorithms for a target that is both fully sunlit and obscured by a tree canopy

In this paper, we compare three basis-vector selection methods as applied to subpixel target detection. This is a continuation of previous research in which a similar comparison was performed based on an AVIRIS image. Our goal is to find out to what extent our previous observations apply more broadly to other images, more specifically, a HYDICE image used in this paper. Our target detection approach is based on generating a radiance target region using a physical model to generate radiance spectra as observed under a wide range of atmospheric, illumination., and viewing conditions. The advantage of this approach is that the resulting target detection is invariant to those changing conditions. For the purpose of target detection, we use a structured model to describe each image spectra as a linear combination of the target and background basis-vectors, and then we apply a matched subspace detector. Finally, we find ROC curves to describe the relationship between the detection rate (DR) and the false alarm rate (FAR). Due to a large number of cases considered, we use summary metrics to represent our results. The obtained results are quite different from those obtained in (Bajorski et al., 2004) for the AVIRIS image. The best method for generating the background basis vectors in the AVIRIS image was the MaxD method, while the SVD method proved to be best for the HYDICE image used in this paper. Further research is needed to find out the reasons for these differences. It is not surprising that different methods are optimal for different types of data. However, it would be useful to be able to recognize the optimal method without assuming knowledge of the targets in the image