Advances of Remote Sensing in Land Cover and Land Use Mapping

1. Introduction

Understanding and monitoring land use and land cover (LULC) dynamics have become increasingly critical in the face of mounting socio-economic, political, and environmental challenges, including climate change, urbanization, and resource degradation [1,2]. Effective spatial planning and sustainable development strategies necessitate the integration of up-to-date information on land use and land cover changes, which are driven by complex interactions among climatic variability, demographic shifts, and economic transformations [3,4]. Addressing these challenges requires robust methods for observing, mapping, monitoring, analyzing, and modeling land surface processes. Remote sensing has emerged as a fundamental tool in this domain, benefiting from recent advances in artificial intelligence and machine learning, the expansion of Earth observation platforms, and the increasing availability of high-resolution satellite and airborne data, often supplemented with ancillary geospatial datasets [5,6]. These developments have significantly enhanced the capacity to map, characterize, and interpret land cover and land use changes at various spatial and temporal scales for a variety of applications.

The aim of this Special Issue is to present recent advances in the methodologies for mapping land cover and land use using remote sensing technologies, with a particular focus on enhancing methodological consistency, comparability, accuracy, and transparency in the assessment, monitoring, and prediction of land use and land cover changes. The contributions emphasize the integration of spatial analysis, machine learning, and remote sensing techniques to improve the detection and understanding of spatial and temporal land dynamics. Recent progress in the spatial modeling of territorial patterns and their evolution through remote sensing has also raised important conceptual questions concerning the interpretation of environmental and territorial processes, including the representation of multi-scalar dynamics and the diverse temporalities associated with land cover and land use changes.

Altogether, this Special Issue features 13 papers, including 12 research articles and 1 literature review. The esteemed first authors contributing to this Special Issue represent seven countries, namely China (5 publications), the United States (2), Ethiopia (2), Brazil, Egypt, Germany, and South Africa. In addition, their co-authors are affiliated with institutions in seven other countries beyond those of the first authors, reflecting the international and collaborative nature of the research. More detailed information on the individual articles published in this Special Issue is provided below, listed alphabetically by the first author’s name.

2. An Overview of Published Articles

We begin with a comprehensive review paper by Ajibola and Cabral analyzing semantic segmentation models in land cover mapping (Contribution 1). They conducted a systematic literature review and bibliometric analysis of semantic segmentation models used in land cover mapping from 2017 to 2023. By analyzing 106 articles, they identified dominant research themes (e.g., urban areas, agriculture, forests), key geographic focuses (notably China and the USA), and widely used data sources like Sentinel-2 and Landsat. Most models rely on encoder–decoder and CNN-based architectures due to their strong performance, while transformer-based models remain underutilized because of their high computational demands. The study also highlights major research gaps, providing a solid foundation for future advancements in deep learning-based land cover mapping.

Admas et al. (Contribution 2) evaluated the combined and individual impacts of LULC and climate change on runoff and sediment flows in Ethiopia’s Megech watershed using the GeoWEPP model and CA-ANN-based land use/land cover projections. Their results indicate that cropland and settlements are expanding while forests and rangelands decline, and their climate projections (RCP4.5 and RCP8.5) suggest increased temperatures and shifts in rainfall patterns. These changes are projected to significantly elevate soil loss and sediment yield—up to 41.01%, especially under RCP8.5—posing a threat to reservoir capacity. The study emphasizes the urgent need for climate adaptation and land management strategies to mitigate these impacts.

Bhungeni et al. (Contribution 3) evaluated the effectiveness of four machine learning algorithms—Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), and Naive Bayes (NB)—for classifying LULC in the uMngeni Catchment, South Africa, using Landsat 8 imagery. Among the tested models, RF demonstrated the highest accuracy (OA = 97.02%, Kappa = 0.96), outperforming SVM and ANN, while NB showed the lowest performance, likely due to its sensitivity to limited training data. The findings underscore RF’s suitability for high-dimensional, remote sensing-based LULC mapping in complex watershed environments.

De Barros de Sousa et al. (Contribution 4) demonstrated that rainfall in a semi-arid Brazilian basin exhibited strong spatial dependence and increased with elevation, as revealed through geostatistical analysis of 55 years of CHIRPS and station data. The Gaussian semi variogram model effectively captured this spatial structure, and the remote sensing indices indicated that rainfall variability significantly influenced land use and vegetation dynamics. The findings highlight geostatistics as a robust tool for understanding precipitation patterns and guiding land and water resource management in semi-arid environments.

Khan and Chen (Contribution 5) examined the impact of LULC change on environmental conditions in Khyber Pakhtunkhwa, Pakistan, using Landsat imagery from 2000 to 2020, along with future projections up to 2100. Urban areas increased by 138%, significantly contributing to rising land surface temperatures, which in turn negatively affected key environmental indices such as NDVI, MNDWI, and NDMI, indicating declines in vegetation, water resources, and soil moisture. Future scenarios project continued urban expansion, vegetation loss, and intensifying environmental stress, underscoring the urgent need for sustainable land use planning, green infrastructure, and climate adaptation measures to mitigate ecological degradation and heat-related risks.

The Region-Specific Model Adaptation (RSMA) method was introduced by Li et al. (Contribution 6) to improve training data extraction for long-term, large-scale land cover mapping in Alaska—a region with limited historical datasets and complex land cover variability. RSMA combines data from NLCD, LANDFIRE EVT, and NWI with machine learning to generate high-quality training samples, accounting for regional spectral variations while preserving generalizable land cover traits. Validation using vegetation plot data showed that RSMA improved classification accuracy by approximately 30% over models using only NLCD-derived training data, demonstrating its effectiveness for large-scale, data-sparse regions.

Liu et al. (Contribution 7) introduced an Explainability Feature Bands Adaptive Selection Network (EFBASN) approach for hyperspectral image classification, addressing challenges such as poor interpretability, high dimensionality, and classification accuracy. The method identified and visualized salient spectral bands that contributed most to classification, enhancing both performance and interpretability of deep learning models. On benchmark datasets like Pavia University, EFBASN achieved state-of-the-art accuracy (97.68%), outperforming existing methods. Importantly, it assigned significantly higher attention weights to physically meaningful bands (e.g., 672 nm for iron oxide), offering scientifically grounded insights into spectral feature relevance.

The study by Lucas et al. (Contribution 8) evaluated past and predicted future LULC changes in Ethiopia’s Upper Omo–Gibe River basin from 1991 to 2067 using satellite imagery and a hybrid Cellular Automata–Artificial Neural Network (CA-ANN) model implemented through QGIS’s MOLUSCE plugin. Classification was performed using the random forest algorithm, achieving high accuracy (kappa = 0.82). The results showed that agricultural land and shrubland are projected to expand significantly between 2022 and 2037, while forest and grassland decline. From 2037 to 2067, urban expansion and continued agricultural growth are expected to further reduce forest cover. These trends highlight the need for targeted land management strategies to address environmental concerns stemming from ongoing LULC changes.

Selmy et al. (Contribution 9) applied a CA–Markov hybrid model, along with Landsat imagery and GIS techniques, to detect, analyze, and predict LULC changes in the arid Sohag Governorate of Egypt from 1984 to 2050. The analysis revealed significant urban expansion and growth in cultivated lands—primarily at the expense of desert lands—driven by urbanization and land reclamation. The model demonstrated high predictive accuracy (Kappa: 0.84–0.93), indicating its strong utility for forecasting LULC trends and informing sustainable land management and policy decisions in arid regions.

Shastry et al. (Contribution 10) demonstrated that machine learning (ML) could significantly improve the accuracy of bedrock outcrop mapping in the Sierra Nevada Mountains compared to the existing 30 m resolution USGS NLCD, which underrepresents barren land. By training a model on high-resolution 0.6 m NAIP imagery and labeled data from 83 km² of the region, the ML approach achieved 90% overall accuracy, 83% precision, and 78% recall in detecting bedrock. The results showed that NLCD captured only about 40% of bedrock mapped by the ML model, suggesting that ML-based methods provide enhanced detail and reliability for environmental and geological applications like erosion modeling and land–atmosphere interaction studies.

Sun et al. (Contribution 11) introduced a novel, high-quality 1 km monthly NDVI time series for China, reconstructed using a Random Forest (RF) model to overcome contamination issues (e.g., clouds, snow) in MODIS NDVI data. By integrating climate, geographic, and auxiliary datasets, the RF model accurately predicted NDVI values, achieving superior performance (MAE = 0.024, RMSE = 0.034, R² = 0.974) compared to existing methods like STSG and LSTM. The resulting product showed improved spatial–temporal continuity and robustness in complex environmental conditions, offering a valuable tool for monitoring vegetation dynamics in China.

Sun et al. (Contribution 12) presented a hybrid method combining Otsu thresholding and Random Forest classification to produce high-resolution (10 m) LULC maps using Sentinel-2 imagery for Wensu County, focusing on seasonal changes, particularly in water and snow/ice dynamics. The approach achieved strong classification accuracy—93.77% for vegetation and 85.50% for water/snow/ice—highlighting the significance of terrain features in improving model performance. Seasonal analysis revealed notable expansion of snow/ice in winter, covering large areas of bare land and grassland, with most changes occurring in areas with moderate winter temperatures (>−8 °C). The method offers a robust framework for monitoring environmental changes and supporting resource and agricultural management.

Zhang et al. (Contribution 13) introduced a High-Performance Automated Large-Area Land-Cover Mapping Framework (HALF), designed to address the inefficiencies and computational demands of large-scale, multi-temporal land cover mapping. By integrating Docker, workflow technologies, and high-performance computing, HALF streamlined model deployment, automated sample generation, and significantly accelerated image processing and classification—achieving a tenfold increase in speed during sample–image matching and the rapid mosaicking of classification outputs. The framework’s modular, scalable design demonstrates strong potential for producing high-resolution land cover maps efficiently across extensive regions, advancing environmental monitoring and sustainable land management.

3. Conclusions

This Special Issue presents a broad array of methodological approaches and applications in land use and land cover mapping, monitoring, and analysis using remote sensing. A central theme across the contributions is the integration of machine learning and deep learning techniques—particularly Random Forest, Convolutional Neural Networks, and novel explainable AI architectures—which have notably enhanced classification accuracy, scalability, and interpretability. These methods are especially effective in handling high-dimensional and heterogeneous remote sensing datasets, highlighting their increasing significance in operational LULC monitoring. Notably, traditional supervised classification methods appeared in only one of the studies.

Several contributions extend beyond mapping current and historical LULC patterns, focusing instead on long-term monitoring and scenario-based projections of future changes. These studies often employ hybrid models such as Cellular Automata–Artificial Neural Networks and CA-Markov, typically using remote sensing products as primary inputs. Such approaches enable detailed assessments of the environmental impacts of land transformations, including vegetation change, soil erosion, and urban expansion, frequently in the context of climate change scenarios (e.g., RCP pathways).

A distinguishing feature of the contributions is their geographical breadth, with case studies spanning diverse ecological and socio-economic contexts—from the semi-arid regions of Brazil and Egypt to the mountainous landscapes of Alaska and the Sierra Nevada. These region-specific analyses highlight the importance of context-sensitive model adaptation, and the development of methodologies suited to data-scarce and environmentally complex settings.

Most studies utilized imagery from Landsat sensors (TM, ETM+, and OLI), with additional cases drawing on Sentinel-2 and MODIS-derived products. Thematic databases derived from remote sensing also played a key role in supporting these analyses. Advances in data fusion—particularly the integration of satellite imagery with climatic and ancillary datasets—were shown to contribute to the production of more detailed and accurate LULC classifications.

Finally, many of the studies exhibited a clear alignment with policy-relevant issues, including urbanization, climate adaptation, and resource management. This highlights the practical utility of remote sensing-based LULC monitoring not only for advancing scientific understanding but also for informing evidence-based decision-making in environmental and land governance.

In summary, the contributions to this Special Issue reflect a rapidly evolving field characterized by increasing technical sophistication and growing applied relevance. Nevertheless, the contributions also highlight key directions for future research in the application of remote sensing to LULC mapping. Future progress will depend on expanding access to high-quality training data, enhancing model generalizability and interpretability, and fostering deeper integration of LULC science with broader climate, ecological, and socio-economic systems. The increasing availability and diversity of Earth observation data present new opportunities for near-real-time LULC monitoring, emphasizing the need for robust pipelines capable of automated change detection and alert generation. While there remains significant scope for advancing explainable AI techniques that provide transparent insights into classification and prediction processes, future work should also focus on adapting interpretable AI frameworks to address the spatiotemporal complexities of LULC dynamics. In parallel, research should prioritize the development of scalable approaches for multimodal data assimilation, integrating optical, radar, thermal, and ancillary socio-environmental datasets to improve predictive performance. The creation, thorough documentation, and open dissemination of large, labeled datasets would further support the development and validation of generalizable AI models. Additionally, extending research on transfer learning and domain adaptation to LULC applications—particularly in data-scarce regions or environmentally heterogeneous isolated ecosystems—remains a critical priority. Finally, a closer coupling of LULC modeling efforts with socio-economic forecasting and Earth system models could enable more comprehensive scenario analyses under varying climate trajectories, thereby enhancing both scientific insight and the utility of remote sensing for evidence-based environmental planning and policy-making.