A New Multispectral Data Augmentation Technique Based on Data Imputation

Abstract

Deep Learning (DL) has been recently introduced into the hyperspectral and multispectral image classification landscape. Despite the success of DL in the remote sensing field, DL models are computationally intensive due to the large number of parameters they need to learn. The high density of information present in remote sensing imagery with high spectral resolution can make the application of DL models to large scenes challenging. Methods such as patch-based classification require large amounts of data to be processed during the training and prediction stages, which translates into long processing times and high energy consumption. One of the solutions to decrease the computational cost of these models is to perform segment-based classification. Segment-based classification schemes can significantly decrease training and prediction times, and also offer advantages over simply reducing the size of the training datasets by randomly sampling training data. The lack of a large enough number of samples can, however, pose an additional challenge, causing these models to not generalize properly. Data augmentation methods are used to generate new synthetic samples based on existing data to increase the classification performance. In this work, we propose a new data augmentation scheme using data imputation and matrix completion methods for segment-based classification. The proposal has been validated using two high-resolution multispectral datasets from the literature. The results obtained show that the proposed approach successfully increases the classification performance across all the scenes tested and that data imputation methods applied to multispectral imagery are a valid means to perform data augmentation. A comparison of classification accuracy between different imputation methods applied to the proposed scheme was also carried out.

Nearest Neighboring Self-Supervised Learning for Hyperspectral Image Classification

“…Nalepa et al proposed two ways for DA, i.e., shift the values in each band relative to the average band value and multiply principal components (PCs) by random factors [51]. Other works explored DA when samples behave as local patches, such as spatial random occlusion [52], rotate and flip [53,54], patch cleaning and imputation based on the border between superpixels [55,56]. Derived from the traditional augmentation technique, namely rotate and flip, Acción et al applied heterogeneous operations of these two techniques on inner and outer patches [57].…”

Generation Of Dense Urban Features Using Conditional GAN

“…Apart from improving accuracy and reducing overfitting [13][14][15], data augmentation using synthetic data can improve the ability of deep learning frameworks to perform well on unseen and unobserved inputs [16], known as generalization ability [17]. In the domain of remote sensing, synthetic data augmentation has been utilized in semantic segmentation [18,19], remote sensing image classification [11,20,21], detection and recognition of detection models [22], data translation [23], target detection [22,24], and others. Goodfellow et al (2014) introduced the generative adversarial network (GAN) that learns to produce synthetic examples with the matching characteristics of the data distribution on which they are trained [25].…”

Generation Of Dense Urban Features Using Conditional GAN

“…Wang, Zhang, Leng, & Leung, 2019), data augmentation using synthetic data can improve the ability of deep learning frameworks to perform well on unseen and unobserved inputs (Bowles et al, 2018), known as generalization ability (Tang, Sun, & Shen, 2021). In the domain of remote sensing, synthetic data augmentation has been utilized in semantic segmentation (Bittner, Ferreira, Andrada, Bird, & Portugal, 2022;Parekhji, Pandya, & Kanani, 2021), remote sensing image classification (Acción, Argüello, & Heras, 2021;W. Wang, Liu, & Mou, 2021;S.…”