Implementing Parallel and Streaming Wavelet Neural Networks for Classification and Regression Using Apache Spark

Implementing Parallel and Streaming Wavelet Neural Networks for Classification and Regression Using Apache Spark Student's Name Institutional Affiliation Course Number and Name Instructor's Name Assignment Due Date Dataset Description The dataset selected for this project is the SpaceNet 7 Satellite Image Dataset, which contains high-resolution satellite images. The dataset used in this case is multi-temporal since the images obtained are taken over different periods of the year and can be used to assess changes in space and time (Van Etten et al. 2021). Every image is created of spectral bands, and source data contains visible and non-visible peculiarities. This dataset consists of two images that show outcomes of the same area taken during different times but within a year to capture the spatial and temporal features (Deshmukh et al., 2023). Several spectral data bands are included on each image, which contain both visible and non-visible features. Wavelet neural networks (WNNs) are antenna networks that extract spatial and frequency information (Van Etten & Hogan, 2021). The dataset's number and complexity demonstrate the Apache Spark platform's efficacy in big data (Van Etten et al., 2021). Yet another reason for its selection was its application in land use maps, change detection, and loss from disasters, directly impacting urban and peri-urban areas and environmental management and policies. This dataset pertains to classification and regression tasks (Hafner et al., 2024). Classification involves ascertaining land use types, including urban, forest, and water. In contrast, regression aims to estimate changes in a particular form of coverage of a specific area within a given time (Deshmukh et al., 2023). The effectiveness of parallel and streaming WNNs makes this concept a perfect test benchmark. Task Details The project explored the scenario of applying WNNs for distributed and real-time computation over large datasets using Apache Spark. The main challenge is that WNNs were not developed for distributed computing landscapes. This project closes that gap by proposing using Apache Spark for parallel processing and streaming. The objective is to implement WNNs that are scalable and efficient when managing large-scale satellite imagery both in batch mode and streaming setups. Batch processing works on the principle of processing a large amount of data simultaneously while streaming deals with the input data by mimicking real-time scenarios (Venkatesh et al., 2022). Therefore, using the two approaches guarantees models are credible and can be applied in different settings. Concerns that have guided this project include how WNNs can be integrated to work with the parallel and streaming aspects of Apache Spark, whether this integration enhances the performance and scalability of the WNNs, and under which circumstances WNNs can be used to surpass conventional machine learning algorithms. All these questions require a background knowledge of the data set, the wavelet mathematics on which the data set is based, and the inner workings of the Apache Spark computing platform. Methods Dataset Pre-processing The images were pre-processed, and numerical feature vectors were extracted before feeding the feature into the neural network because the network does not accept raw images. High-resolution satellite images from the SpaceNet 7 dataset were used in the experiments. The dataset set at the beginning of the experiment was of the order of 40,000 samples of size 1024 x 1024 × 3 RGB channel. These images offered a diverse set of features that could then construct classifiers. However, data had to be pre-processed, making it suitable for use in WNN and the architecture of Apache Spark. The image data was pre-processed to increase the speed of computations within the proposed neural network as well as the quality of learning. First, the images were resized from the original size to 256×256 to decrease the time needed for calculations without losing some spatial information. Otherwise, this resized shape was compatible with successful conversion with wavelet transformations and parallel processing. These obtained images were then downscaled, and wavelet transformations were used to get a set of spatial and frequency domain numerical descriptors for the image (Eduru et al., 2024). These transformations split the images into coefficients indicating the presence of various scales and frequencies, but, in principle, the feature vectors captured a wealth of information that the images contained. These wavelet coefficients were spread into the one-dimensional structure, and the last feature vector of image size was 65536. This transformed data was aimed to scale the data for input into the neural network, but at the same time, other aspects of the data should maintain features that are useful for the analysis (Xian, 2020). After the feature extraction process, the dataset was divided into training and testing data sets. A typical 80:20 split was used to form a robust assessment matrix in which the dataset contained 32000 images for training and 8000 for testing. Apache Spark was also used in pre-processing, and distributed data handling was used for batch transformation and big data management. The data obtained was stored in a distributed format of Apache Parquet for fast access to data during experimental work and training neural networks. The pre-processing approach allowed automation of the dataset preparation process for batch and stream evaluations and addressed computational cost and data integrity for the WNNs (Eduru et al., 2024). Model Architecture and Neural Network Design To classify the feature vectors after the wavelet transformation, the project used Multi-Layer Perceptron (MLP). The network architecture defined as the input layer took the feature vectors of size 65,536, with two hidden layers of 512 neurons each. These layers employed ReLU activation functions to non-linear transform features, further enhancing the model's pattern recognition capabilities. The last fully connected layer was equal to the number of unique types of land use in the dataset. A softmax layer for probability-based classification of the image continued the network. Dropout regularization was done at 30 %, along with batch normalization, which made the training more generalized, changing it from volatility to stability. Following these techniques, overfitting was minimized, and variance in optimization behavior across epochs was positively controlled. For more efficient training, the Adam optimizer was performed, and for the final classification, a definite cross entropy was used; the model was trained for 50 epochs and a batch size of 128. This is why the starting point for the classification performance was the MLP model, which, though less complex t...