The success of the supervised classification of remotely sensed images acquired over large geographical areas or at short time intervals strongly depends on the representativity of the samples used to train the classification algorithm and to define the model. When training samples are collected from an image or a spatial region that is different from the one used for mapping, spectral shifts between the two distributions are likely to make the model fail. Such shifts are generally due to differences in acquisition and atmospheric conditions or to changes in the nature of the object observed. To design classification methods that are robust to data set shifts, recent remote sensing literature has considered solutions based on domain adaptation (DA) approaches. Inspired by machine-learning literature, several DA methods have been proposed to solve specific problems in remote sensing data classification. This article provides a critical review of the recent advances in DA approaches for remote sensing and presents an overview of DA methods divided into four categories: 1) invariant feature selection, 2) representation matching, 3) adaptation of classifiers, and 4) selective sampling. We provide an overview of recent methodologies, examples of applications of the considered techniques to real remote sensing images characterized by very high spatial and spectral resolution as well as possible guidelines for the selection of the method to use in real application scenarios.

This paper investigates different batch-mode active-learning (AL) techniques for the classification of remote sensing (RS) images with support vector machines. This is done by generalizing to multiclass problem techniques defined for binary classifiers. The investigated techniques exploit different query functions, which are based on the evaluation of two criteria: uncertainty and diversity. The uncertainty criterion is associated to the confidence of the supervised algorithm in correctly classifying the considered sample, while the diversity criterion aims at selecting a set of unlabeled samples that are as more diverse (distant one another) as possible, thus reducing the redundancy among the selected samples. The combination of the two criteria results in the selection of the potentially most informative set of samples at each iteration of the AL process. Moreover, we propose a novel query function that is based on a kernel-clustering technique for assessing the diversity of samples and a new strategy for selecting the most informative representative sample from each cluster. The investigated and proposed techniques are theoretically and experimentally compared with state-of-the-art methods adopted for RS applications. This is accomplished by considering very high resolution multispectral and hyperspectral images. By this comparison, we observed that the proposed method resulted in better accuracy with respect to other investigated and state-of-the art methods on both the considered data sets. Furthermore, we derived some guidelines on the design of AL systems for the classification of different types of RS images.

Accurate spatial information of agricultural fields in smallholder farms is important for providing actionable information to farmers, managers, and policymakers. Very High Resolution (VHR) satellite images can capture such information. However, the automated delineation of fields in smallholder farms is a challenging task because of their small size, irregular shape and the use of mixed-cropping systems, which make their boundaries vaguely defined. Physical edges between smallholder fields are often indistinct in satellite imagery and contours need to be identified by considering the transition of the complex textural pattern between fields. In these circumstances, standard edge-detection algorithms fail to extract accurate boundaries. This article introduces a strategy to detect field boundaries using a fully convolutional network in combination with a globalisation and grouping algorithm. The convolutional network using an encoder-decoder structure is capable of learning complex spatial-contextual features from the image and accurately detects sparse field contours. A hierarchical segmentation is derived from the contours using the oriented watershed transform and by iteratively merging adjacent regions based on the average strength of their common boundary. Finally, field segments are obtained by adopting a combinatorial grouping algorithm exploiting the information of the segmentation hierarchy. An extensive experimental analysis is performed in two study areas in Nigeria and Mali using WorldView-2/3 images and comparing several state-of-the-art contour detection algorithms. The algorithms are compared based on the precision-recall accuracy assessment strategy which is tolerating small localisation errors in the detected contours. The proposed strategy shows promising results by automatically delineating field boundaries with F-scores higher than 0.7 and 0.6 on our two test areas, respectively, outperforming alternative techniques.

This paper aims at analyzing and comparing active learning (AL) and semisupervised learning (SSL) methods for the classification of remote sensing (RS) images. We present a literature review of the two learning paradigms and compare them theoretically and experimentally when addressing classification problems characterized by few training samples (w.r.t. the number of features) and affected by sample selection bias. Commonalities and differences are highlighted in the context of a conceptual framework used to describe the workflow of the two approaches. We point out advantages and disadvantages of the two approaches, delineating the boundary conditions on the applicability of the two paradigms with respect to both the amount and the quality of available training samples. Moreover, we investigate the integration of concepts that are in common between the two learning paradigms for improving state-of-the-art techniques and combining AL and SSL in order to jointly leverage the advantages of both approaches. In this framework, we propose a novel SSL algorithm that improves the progressive semisupervised support vector machine by integrating concepts that are usually considered in AL methods. We performed several experiments considering both synthetic and real multispectral and hyperspectral RS data, defining different classification problems starting from different initial training sets. The experiments are carried out considering classification methods based on support vector machines.

This letter investigates fully convolutional networks (FCNs) for the detection of informal settlements in very high resolution (VHR) satellite images. Informal settlements or slums are proliferating in developing countries and their detection and classification provides vital information for decision making and planning urban upgrading processes. Distinguishing different urban structures in VHR images is challenging because of the abstract semantic definition of the classes as opposed to the separation of standard land-cover classes. This task requires extraction of texture and spatial features. To this aim, we introduce deep FCNs to perform pixel-wise image labeling by automatically learning a higher level representation of the data. Deep FCNs can learn a hierarchy of features associated to increasing levels of abstraction, from raw pixel values to edges and corners up to complex spatial patterns. We present a deep FCN using dilated convolutions of increasing spatial support. It is capable of learning informative features capturing long-range pixel dependencies while keeping a limited number of network parameters. Experiments carried out on a Quickbird image acquired over the city of Dar es Salaam, Tanzania, show that the proposed FCN outperforms state-of-the-art convolutional networks. Moreover, the computational cost of the proposed technique is significantly lower than standard patch-based architectures.

This paper presents a novel technique for addressing domain adaptation (DA) problems with active learning (AL) in the classification of remote sensing images. DA models the important problem of adapting a supervised classifier trained on a given image (source domain) to the classification of another similar but not identical image (target domain) acquired on a different area. The main idea of the proposed approach is iteratively labeling and adding to the training set the minimum number of the most informative samples from the target domain, while removing the source-domain samples that do not fit with the distributions of the classes in the target domain. In this way, the classification system exploits already available information, i.e., the labeled samples of source domain, in order to minimize the number of target domain samples to be labeled, thus reducing the cost associated to the definition of the training set for the classification of the target domain. In addition, we define a convergence criterion that allows the technique to stop the iterative AL process on the target domain without relying on the availability of a test set for it. This is an important contribution, as in operational applications, it is not realistic to assume that a test set for the target domain is available. Experimental results obtained in the classification of very high resolution and hyperspectral images confirm the effectiveness of the proposed technique.