This work aims at addressing two issues simultaneously: data compression at input space and semantic segmentation. Semantic segmentation of remotely sensed multi- or hyperspectral images through deep learning (DL) artificial neural networks (ANN) delivers as output the corresponding matrix of pixels classified elementwise, achieving competitive performance metrics. With technological progress, current remote sensing (RS) sensors have more spectral bands and higher spatial resolution than before, which means a greater number of pixels in the same area. Nevertheless, the more spectral bands and the greater number of pixels, the higher the computational complexity and the longer the processing times. Therefore, without dimensionality reduction, the classification task is challenging, particularly if large areas have to be processed. To solve this problem, our approach maps an RS-image or third-order tensor into a core tensor, representative of our input image, with the same spatial domain but with a lower number of new tensor bands using a Tucker decomposition (TKD). Then, a new input space with reduced dimensionality is built. To find the core tensor, the higher-order orthogonal iteration (HOOI) algorithm is used. A fully convolutional network (FCN) is employed afterwards to classify at the pixel domain, each core tensor. The whole framework, called here HOOI-FCN, achieves high performance metrics competitive with some RS-multispectral images (MSI) semantic segmentation state-of-the-art methods, while significantly reducing computational complexity, and thereby, processing time. We used a Sentinel-2 image data set from Central Europe as a case study, for which our framework outperformed other methods (included the FCN itself) with average pixel accuracy (PA) of 90% (computational time ∼90s) and nine spectral bands, achieving a higher average PA of 91.97% (computational time ∼36.5s), and average PA of 91.56% (computational time ∼9.5s) for seven and five new tensor bands, respectively.

Remote sensing (RS) spectral time series provide a substantial source of information for the regular and cost-efficient monitoring of the Earth’s surface. Important monitoring tasks include land use and land cover classification, change detection, forest monitoring and crop type identification, among others. To develop accurate solutions for RS-based applications, often supervised shallow/deep learning algorithms are used. However, such approaches usually require fixed-length inputs and large labeled datasets. Unfortunately, RS images acquired by optical sensors are frequently degraded by aerosol contamination, clouds and cloud shadows, resulting in missing observations and irregular observation patterns. To address these issues, efforts have been made to implement frameworks that generate meaningful representations from the irregularly sampled data streams and alleviate the deficiencies of the data sources and supervised algorithms. Here, we propose a conceptually and computationally simple representation learning (RL) approach based on autoencoders (AEs) to generate discriminative features for crop type classification. The proposed methodology includes a set of single-layer AEs with a very limited number of neurons, each one trained with the mono-temporal spectral features of a small set of samples belonging to a class, resulting in a model capable of processing very large areas in a short computational time. Importantly, the developed approach remains flexible with respect to the availability of clear temporal observations. The signal derived from the ensemble of AEs is the reconstruction difference vector between input samples and their corresponding estimations, which are averaged over all cloud-/shadow-free temporal observations of a pixel location. This averaged reconstruction difference vector is the base for the representations and the subsequent classification. Experimental results show that the proposed extremely light-weight architecture indeed generates separable features for competitive performances in crop type classification, as distance metrics scores achieved with the derived representations significantly outperform those obtained with the initial data. Conventional classification models were trained and tested with representations generated from a widely used Sentinel-2 multi-spectral multi-temporal dataset, BreizhCrops. Our method achieved $77.06\%$ overall accuracy, which is $\sim 6\%$ higher than that achieved using original Sentinel-2 data within conventional classifiers and even $\sim 4\%$ better than complex deep models such as OmnisCNN. Compared to extremely complex and time-consuming models such as Transformer and long short-term memory (LSTM), only a $3\%$ reduction in overall accuracy was noted. Our method uses only $6.8k$ parameters, i.e., $\sim 400x$ fewer than OmnicsCNN and $\sim 27x$ fewer than Transformer. The results prove that our method is competitive in terms of classification performance compared with state-of-the-art methods while substantially reducing the computational load.

Remote sensing imaging datasets for classification generally present high levels of imbalance between classes of interest. This work presented a study of a set of performance evaluation metrics for an imbalance dataset. In this work, a support vector machine (SVM) was used to perform the classification of seven classes of interest in a popular dataset called Salinas-A. The performance evaluation of the classifier was performed using two types of metrics: 1) Metrics for multi-class classification, and 2) Metrics based on the binary confusion matrix. In the results, a comparison of the scores of each metric is developed, some being more optimistic than others due to the bias that they present given the imbalance. In addition, our case study helps to conclude that the Matthews correlation coefficient (MCC) presents the lowest bias in imbalanced cases and is regarded to be robust metric. These results can be extended to any imbalanced dataset taking into account the equations developed by Luque.

Bathymetry is a method of quantifying depths to study the topography of water bodies, including oceans, seas, rivers and lakes. The measurement of bathymetry by means of satellite images is one of the fundamental investigations in the field of remote sensing (RS) of marine environment, which has a lot of applications for the coastal environment and its monitoring. The precise determination of depth water is essential for various purposes, such as the monitoring of underwater topography, the movement of deposited sediments and the production of maritime maps for navigation. Remote sensing allows the bathymetry modeling at spatial scales that are impossible to achieve with traditional methods. Bathymetry can be estimated using RS with several techniques, each with its own capacity for depth detection, accuracy, error, strengths, advantages, disadvantages and the best application environment. Before that, a GUI interface is developed in Matlab that contains enough data to be able to compute the bathymetry in multispectral images from the satellite LANDSAT 8, with the intention of being able to analyze how flooded an area will be.