Scene-level classification of aerial images is challenging due to varying object scales, positions, inter-class similarity, and intra-class diversity. Convolutional Neural Networks (CNNs) are effective for extracting useful feature maps, while Capsule Networks (CapsNets) are adept at recognizing the pose information of objects in an image. This work uses both models to ensure that aerial scenes are accurately classified. Additionally, considering that CapsNets typically require many parameters and floating-point operations, we aim to make our model lightweight. Using adaptive fusion of features extracted from MobileNetv2, a lightweight CNN model, enables the capsule network to learn faster, resulting in improved performance. With more than 58% parameter reduction for the datasets UC Merced, AID, and SIRI-WHU, the proposed approach, termed as AdMobiCaps, achieves significantly better performance compared to traditional CapsNet showcasing its potential in aerial scene classification.

Graph neural networks (GNNs) have become the prevailing methodology for addressing graph data-related tasks, permeating critical domains like recommendation systems and drug development. The necessity for trustworthiness and interpretability of GNNs has risen to the forefront, especially given their direct impact on end users' lives. To address this need, we present Gen-GraphEx, a model-agnostic, model-level explanation method that prioritizes user centricness by eliminating the need for having access to the hidden layers of the GNN model it seeks to explain. Given a particular class label, Gen-GraphEx learns a graph generative model (GGM) that produces explanation graphs that not only contain discriminative patterns that the GNN has learned for that class but also lie in distribution with real graphs that belong to that class according to the GNN. Unlike existing state-of-the-art models, Gen-GraphEx also has the unique ability to interpolate the GGMs of two target classes to generate instances that lie near the decision boundary of the two classes giving a deeper insight into the model's decision-making. Its advantages over existing methods in the literature also include nonreliance on another subsequent deep learning module for explanation generation, ability to generate graphs with various node and edge features, and being more computationally efficient. Extensive validation and thorough comparative analysis of the proposed approach is carried out across an array of real and synthetic datasets that consistently demonstrate its exceptional performance and competitiveness ranking alongside state-of-the-art model-level explainers. Our code is available at https://github.com/amisayan/Gen-GraphEx.

Recently, convolutional neural networks (CNNs) have achieved great success by attaining state-of-the-art accuracies for aerial scene classification. However, there is a serious lack of good explanations for understanding the decision-making process of such black-box models. To establish the trustworthiness of the classifiers, various local and global explainers are commonly used nowadays. These primarily provide us with explanations in terms of the most influential features leading to the model decision. However, developing an explainer showing causal relationships among these features can offer more visibility, interpretability, and trustworthiness to the model users or stakeholders. To the best of our knowledge, this area is still underexplored in the context of scene-level classification of aerial images. We address this issue by proposing a novel causality-based CNN explainer based on gradient-weighted class activation mapping (Grad-CAM) and generative flow network (GFlowNet). Our proposed model is termed causal grad-CAM (CG-CAM), where Grad-CAM is used to highlight the relevant regions (layer-wise) in an aerial scene, and the GFlowNet is utilized to generate the directed acyclic graph (DAG) representing causal relationships between the feature maps at different layers of a deep CNN classifier, to achieve better understandability. Experimentation using the benchmark UCMerced and NWPU-RESISC45 datasets demonstrates the effectiveness of our CG-CAM-based explanations for aerial scene classification.

Graph Neural Networks (GNNs) learn node representations by aggregating a node's feature vector with its neighbors. They perform well across a variety of graph tasks. However, to enhance the reliability and trustworthiness of these models during use in critical scenarios, it is of essence to look into the decision making mechanisms of these models rather than treating them as black boxes. Our model-centric method gives insight into the kind of information learnt by GNNs about node neighborhoods during the task of node classification. We propose a neighborhood generator as an explainer that generates optimal neighborhoods to maximize a particular class prediction of the trained GNN model. We formulate neighborhood generation as a reinforcement learning problem and use a policy gradient method to train our generator using feedback from the trained GNN-based node classifier. Our method provides intelligible explanations of learning mechanisms of GNN models on synthetic as well as real-world datasets and even highlights certain shortcomings of these models.

Catastrophic forgettingof previously acquired knowledge is a major setback suffered by the neural networks (NNs) when these are trained on tasks in sequential fashion. The NN variants, including the deep network models, as commonly used in remote sensing data prediction are also not free from this limitation. The issue becomes more prominent when the prediction is performed over data collected from spatial zones with a large degree of subregional variations or heterogeneity. In order to tackle this problem, in this article, we propose a multilayered adaptive recurrent incremental network (MARINE) model, offering a spatial heterogeneity-aware self-adaptation scheme that resolves the catastrophic forgetting issue in an autonomous manner. Typically, the proposed MARINE is equipped with an intrinsic mechanism of clustering the spatial subregions as per their heterogeneity levels and auto-constructing the recurrent network layers in an ensemble fashion so that the knowledge acquired about one group of heterogeneity level does not overwrite that acquired about other groups. With respect to spatio-temporal prediction of normalized difference vegetation index (NDVI) time series, as derived from MODIS Terra satellite remote sensing imagery, we demonstrate that our proposed MARINE achieves competitive results when compared with the state-of-the-arts. More significantly, in the presence of higher degree of spatial heterogeneity, MARINE outperforms others by betteravoiding the catastrophic forgetting issue.

Short-term forecast of power-load consumption provides significant insights into cost-effective load transmission/distribution planning and eventual smooth operation of the power system. With the recent advancements of smart grid technology, the short-term load forecasting has become a challenging task since the relevant data show prominent spatial dependency which must be taken into account at the time of data-driven predictive analytics. However, in several cases, the spatial information (latitude and longitude) of the locations are often found to be missing in the historical load data. This paper attempts to address this issue by proposing a semantics-driven mechanism for deriving possible spatial relationships among the consumer locations followed by short-term load forecasting based on spatio-temporal graph convolutional network. The key idea remains here in indirectly estimating the spatial influences based on the underlying semantics of load usage patterns at the consumer end. The proposed semantically-enhanced load forecasting approach, designated as SELFIE, is evaluated using publicly available benchmark load datasets over the spatial regions of USA. Rigorous comparative study demonstrates the effectiveness of our proposed SELFIE achieving a minimum of 25% performance improvement over the state-of-the-art techniques under the scenarios of missing spatial information.

Stream classification models for non-stationary environments often assume the immediate availability of data labels. However, in a practical scenario, it is quite natural that the data labels are available only after some temporal lag. This paper explores how a stream classifier model can be made adaptive to such label latency scenario. We propose SkipE-RNN, a self-evolutionary recurrent neural network with dynamically evolving skipped-recurrent-connection for the best utilization of previously observed label information while classifying the current data. When the data label is unavailable, SkipE-RNN uses an auto-learned mapping function to find the best match from the already known data labels and updates the network parameter accordingly. Later, upon availability of true data label, if the previously mapped label is found to be incorrect, SkipE-RNN employs a regularization technique along with the parameter updating process, so as to penalize the model. In addition, SkipE-RNN has inborn power of self-adjusting the network capacity by growing/pruning hidden nodes to cope with the evolving nature of data stream. Rigorous empirical evaluations using synthetic as well as real-world datasets reveal effectiveness of SkipE-RNN in both finitely delayed and infinitely delayed data label scenarios.

Neural network models are quite popular among the various machine learning approaches for prediction of derived remote sensing image time series. However, the existing models are mostly based on multi-pass parameter learning strategy and these use fixed network architectures which need to be determined through rigorous empirical study. Eventually, their performance deteriorates during online prediction of such data, as commonly encountered in various real-life scenarios. In order to address this issue, this paper proposes OPAL, an online prediction model based on autonomous learning approach. The autonomous learning of OPAL is achieved by employing a self-evolutionary recurrent neural network, whereas its single-pass learning makes it fit for online prediction environment. Experimentation with normalized difference vegetation index (NDVI) data, derived from MODIS Terra satellite imagery, shows that proposed OPAL is able to attain state-of-the-art accuracy even with single-pass learning and without requiring empirical adjustment of network architecture.

With the increasing demand of product supply, manufacturers are in urgent need of online tool condition monitoring (TCM) without compromising with the maintenance cost in terms of time as well as man-power requirement. However, the existing machine learning models for TCM are mostly offline and not suitable for the non-stationary environment of the machining settings. Moreover, the access of the ground truth always imposes a shutdown of the machining process and the existing models are severely affected by such delay in receiving labelled samples. In order to tackle these issues, we propose SERMON as a novel learning model based on a pair of self-evolving mutually-operative recurrent neural networks. The proposed SERMON is well-equipped with features for automated and real-time monitoring of machine fault status even in the finite/infinite label delay scenario. The experimental evaluation of SERMON using real-world dataset on 3D-printing process demonstrates its effectiveness in online fault detection under non-stationary as well as delayed label context of the machining process. Additional comparative study on large-scale benchmark streaming datasets further exhibits the scalability power of SERMON.

In this paper, we propose MUSE-RNN, a multilayer self-evolving recurrent neural network model for real-time classification of streaming data. Unlike the existing approaches, MUSE-RNN offers special treatment towards capturing temporal aspects of data stream through its novel recurrent learning approach based on the teacher forcing policy. Novelties here are twofold. First, in contrast to the traditional RNN models, MUSE-RNN has intrinsic ability to self-adjust its capacity by growing and pruning hidden nodes as well as layers, to handle the ever-changing characteristics of data stream. Second, MUSERNN adopts a unique scoring-based layer adaptation mechanism, which makes it capable of recalling prior tasks, with minimum exploitation of network parameters. The performance of MUSERNN is evaluated in comparison with a number of state-of-theart techniques, using seven popular data streams and continual learning problems under prequential test-then-train protocol. Experimental results demonstrate the effectiveness of MUSERNN in stream classification scenario.

With the recent explosion of data navigating in motion, there is a growing research interest for analyzing streaming data, and consequently, there are several recent works on data stream analytics. However, exploring the potentials of traditional recurrent neural network (RNN) in the context of streaming data classification is still a little investigated area. In this paper, we propose a novel variant of RNN, termed as FERNN, which features single-pass learning capability along with self-evolution property. The online learning capability makes FERNN fit for working on streaming data, whereas the self-organizing property makes the model adaptive to the rapidly changing environment. FERNN utilizes hyperplane activation in the hidden layer, which not only reduces the network parameters to a significant extent, but also triggers the model to work by default as per teacher forcing mechanism so that it automatically handles the vanishing/exploding gradient issues in traditional RNN learning based on back-propagation-through-time policy. Moreover, unlike the majority of the existing autonomous learning models, FERNN is free from normal distribution assumption for streaming data, making it more flexible. The efficacy of FERNN is evaluated in terms of classifying six publicly available data streams, under the prequential test-then-train protocol. Experimental results show encouraging performance of FERNN attaining state-of-the-art classification accuracy with fairly reduced computation cost.

A proper assessment of reservoir water dynamics is of utmost importance for the development of any region. However, the reservoir water dynamics is not merely a periodic event. Rather, it is a result of complex interplay among various water balancing components, especially the meteorological factors. So, the key objectives of this research work are to model the influence of spatial variability of meteorological variables on the hydrological processes in a reservoir, and to utilize this knowledge of spatial variability to aid in prediction of reservoir dynamics. In this regard, we propose FORWARD, a forecasting model based on spatial Bayesian network (SpaBN) which has inherent capability of efficiently modeling the spatial impact of meteorological and topographical factors distributed over the watershed. The forecasting efficiency of FORWARD has been compared with a set of linear and non-linear prediction techniques with respect to a case study on forecasting daily live capacity of the Mayurakshi reservoir in India. The experimental results show the superiority of FORWARD over the others.

This paper proposes a novel data-driven model (BESTED), based on spatial Bayesian network with incorporated exponential smoothing mechanism, for predicting precipitation time series on daily basis. In BESTED, the spatial Bayesian network helps to efficiently model the influence of spatially distributed variables. Moreover, the incorporated exponential smoothing mechanism aids in tuning the network inferred values to compensate for the unknown factors, influencing the precipitation rate. Empirical study has been carried out to predict the daily precipitation in West Bengal, India, for the year 2015. The experimental result demonstrates the superiority of the proposed BESTED model, compared to the other benchmarks and state-of-the-art techniques.

Natural systems, like the hydrological, climatological, atmospheric, or any other environmental processes, are extremely complex as well as dynamic in nature. It is therefore difficult to forecast, analyze, and quantify these processes by using simple empirical equations. Modeling and forecasting of reservoir water dynamics are not exceptions in this respect, as these involve various challenges due to the effect of meteorological factors, natural processes of stream flow, climatic change, and so on. The intent of our present work is to propose a novel forecasting model, FORWARD, that handles some of these issues in complex reservoir dynamics. FORWARD is based on a variant of spatial Bayesian network (SpaBN), having inherent capability of modeling impact of spatial variability of meteorological factors over the river catchment. The forecasting efficiency of FORWARD has been compared with four other linear and non-linear techniques based on six different statistical performance measures. The experimental results show the superiority of FORWARD over the other techniques. Though FORWARD has been demonstrated with respect to a case study on forecasting reservoir live capacity, the model possesses a generic structure that can also be applied in other domains by introducing minimal augmentation.

Meteorological parameters are often considered as crucial factors for climatological pattern analysis. Predictions of these parameters have been studied extensively in the field of remote sensing and GIS. It is one of the most critical steps involved in most of the meteorological data mining process. Spatial interpolation is an efficient technique to yield minimal error in prediction. From existing literatures, it is evident that the land-use/land-cover (LULC) distribution of the terrain influences these parameters in a varying manner and it is important to model their behaviour for climatological analyses. However, this semantic LULC knowledge of the terrain is generally ignored in the prediction process of the meteorological parameters. Recently, we have proposed a new spatial interpolation technique, namely semantic kriging (SemK) [3,5,7], which considers the semantic LULC knowledge for land-atmospheric interaction modeling and incorporates it into the existing interpolation process for better accuracy. However, the a-priori correlation analysis of SemK ignores the effect of other nearby LULC classes on each other. This article presents a new variant of SemK, namely a-posterior probabilistic Bayesian SemK (BSemK), which extends the a-priori correlation analysis of SemK with a-posterior probabilistic analysis. The proposed approach provides more accurate estimation of the parameters. Experimentation with LST data advocates the efficacy of the proposed approach compared to the a-priori SemK and other existing interpolation techniques.

Spatial autocorrelation (SA), describing correlation of a particular feature/phenomenon with itself across space, is one of the major properties of any spatial data. Among the various measures of SA proposed till date, the Moran's index (I) is the most common as well as significant one. However, measuring Moran's I, which needs to deal with spatial weight between each pair of spatial data objects, becomes almost unfeasible in case of large-scale raster data, like geostationary satellite data, containing several millions of pixels. This paper proposes a method based on the Hadoop MapReduce framework for computing Moran's I in large-scale raster data. The main contribution of the work lies in the implementation of the Mapper and Reducer processes for a cost effective estimation of Moran's I, considering both rook case and queen case of spatial contiguity. The key feature of these algorithms is an efficient manipulation of the spatial weight matrix, and thereby reducing the overall memory and time requirement. The experimentation shows a promising result in this regard.

With the advent of advanced remote sensing technologies in past few decades, acquiring higher resolution satellite images has become easier and cheaper in recent days. However, on the other hand, it has offered a big challenge to the remote sensing community in smart image interpretation from such huge volume of data. Deep learning, which offers efficient algorithms for extracting multiple levels of feature abstractions, may be suitable to serve the purpose. This letter presents a deep learning approach (Deep-STEP) for spatiotemporal prediction of satellite remote sensing data. The proposed learning architecture is derived from a deep stacking network, consisting of a stack of multilayer perceptron, each of which models the spatial feature of the associated region at a particular time instant. The proposed method has been demonstrated on normalized difference vegetation index (NDVI) data sets, derived from satellite remote sensing imagery, containing several thousands to millions of pixels/records. The experimental results (related to NDVI prediction) reveal that the proposed architecture exhibits fairly satisfactory performance with promising learning capabilities.