Deep learning-based segmentation of very high-resolution (VHR) satellite images is a significant task providing valuable information for various geospatial applications, specifically for land use/land cover (LULC) mapping. The segmentation task becomes more challenging with the increasing number and complexity of LULC classes. In this research, we generated a new benchmark dataset from VHR Worldview-3 images for twelve distinct LULC classes of two different geographical locations. We evaluated the performance of different segmentation architectures and encoders to find the best design to create highly accurate LULC maps. Our results showed that the DeepLabv3+ architecture with an ResNeXt50 encoder achieved the best performance for different metric values with an IoU of 89.46%, an F-1 score of 94.35%, a precision of 94.25%, and a recall of 94.49%. This design could be used by other researchers for LULC mapping of similar classes from different satellite images or for different geographical regions. Moreover, our benchmark dataset can be used as a reference for implementing new segmentation models via supervised, semi- or weakly-supervised deep learning models. In addition, our model results can be used for transfer learning and generalizability of different methodologies.

Scanned historical maps are available from different sources in various scales and contents. Automatic geographical feature extraction from these historical maps is an essential task to derive valuable spatial information on the characteristics and distribution of transportation infrastructures and settlements and to conduct quantitative and geometrical analysis. In this research, we used the Deutsche Heereskarte 1:200,000 Türkei (DHK 200 Turkey) maps as the base geoinformation source to construct the past transportation networks using the deep learning approach. Five different road types were digitized and labeled to be used as inputs for the proposed deep learning-based segmentation approach. We adapted U-Net++ and ResneXt50_32×4d architectures to produce multi-class segmentation masks and perform feature extraction to determine various road types accurately. We achieved remarkable results, with 98.73% overall accuracy, 41.99% intersection of union, and 46.61% F1 score values. The proposed method can be implemented in DHK maps of different countries to automatically extract different road types and used for transfer learning of different historical maps.

With the advancement of satellite technology, a considerable amount of very high-resolution imagery has become available to be used for the Land Cover and Land Use (LCLU) classification task aiming to categorize remotely sensed images based on their semantic content. Recently, Deep Neural Networks (DNNs) have been widely used for different applications in the field of remote sensing and they have profound impacts; however, improvement of the generalizability and robustness of the DNNs needs to be progressed further to achieve higher accuracy for a variety of sensing geometries and categories. We address this problem by deploying three different Deep Neural Network Ensemble (DNNE) methods and creating a comparative analysis for the LCLU classification task. DNNE enables improvement of the performance of DNNs by ensuring the diversity of the models that are combined. Thus, enhances the generalizability of the models and produces more robust and generalizable outcomes for LCLU classification tasks. The experimental results on NWPU-RESISC45 and AID datasets demonstrate that utilizing the aggregated information from multiple DNNs leads to an increase in classification performance, achieves state-of-the-art, and promotes researchers to make use of DNNE.

Reliable metrics to monitor human impacts on biodiversity are essential for informing conservation policy. As insects are indicators of global change, whose declines profoundly affect ecosystems, insect diversity may predict biodiversity status. Here we present an unbiased and straightforward biodiversity status metric based on insect diversity (richness) and landscape naturalness. Insect diversity was estimated using spatially explicit earth observation data and insect species assemblages across microhabitats in two agro-ecological zones in Africa. Landscape naturalness was estimated using various human impact factors. Biodiversity status values differed considerably (p < 0.05) between protected and non-protected areas, while protected areas, regardless of agro-ecology, shared similar biodiversity status values. The metric is consistent when using richness from different indicator taxa (i.e., stingless bees, butterflies, dragonflies) and independent data for landscape naturalness. Our biodiversity status metric is applicable to data-scarce environments and practical for conservation actions and reporting the status of biodiversity targets.

The advancement in deep learning (DL) techniques has led to a notable increase in the number and size of annotated datasets in a variety of domains, with remote sensing (RS) being no exception [1] . Also, an increase in Earth observation (EO) missions and the easy access to globally available and free geodata have opened up new research opportunities. Although numerous RS datasets have been published in the past years [2] , [3] , [4] , [5] , [6] , most of them addressed tasks concerning man-made environments, such as building footprint extraction and road network classification, leaving the environmental and ecology-related subareas of RS underrepresented. Nevertheless, environmental protection has always been an important topic in the RS community, with RS being a useful tool to support conservation policies and strategies combating challenges such as deforestation and loss of biodiversity [7] , [8] , [9] . Thus, in this article, to meet the pressing need to better understand the nature we are living in, we introduce a novel task of wilderness mapping and advertise the MapInWild dataset [10] —a multimodal large-scale benchmark dataset designed for the task of wilderness mapping from space.

Advancements in Earth observation (EO) have led to an increase in the volume of and easier access to multimodal geospatial data, making environmental monitoring and analysis more accessible. However, understanding the influence of each input modality on decision-making within deep learning models remains an open challenge. This letter proposes a deep occlusion framework to enhance the interpretability of a multimodal model for land naturalness assessment, using a supervised pixelwise regression task for naturalness mapping with the input modalities Sentinel-2 and Sentinel-1 imagery, land cover maps, and nighttime lights intensity data. The proposed framework systematically occludes individual input modalities to create modality-level influence scores. Influence scores are attributed to input modalities by measuring the distance between the embedding of the nonoccluded input and the embedding of the input with a single modality occluded, revealing how each modality influences predictions and clarifying their contributions (and, thus, importance) in the model’s decision-making process. The results provide further insights into how input modalities influence the model’s decision-making at both the sample level, enabling regional case studies, and the dataset level, allowing for data pruning and improving training and inference times. The code is available at https://github.com/burakekim/embedding_occlusion .