Search

Given the systematic acquisition of satellite data, it is possible to generate up-to-date land cover (LC) maps, essential for effective agricultural territory management, environmental monitoring, and informed decision-making. Typically, creating a LC map requires collecting high-quality labeled data, a process that is both costly and time-consuming. To mitigate the need to collect large volume of labeled data, we propose a deep learning framework called REFeD (data Reuse with Effective Feature Disentanglement for land cover mapping), which leverages already available out-of-year reference data to enhance the production of up-to-date LC maps. To this end, REFeD integrates remote sensing and reference data from different domains (e.g., historical and recent data) utilizing a disentanglement strategy based on contrastive learning. By separating domain-invariant and domain-specific features, REFeD isolates useful information associated to the downstream LC mapping task and mitigates distribution shifts between domains. Moreover, REFeD incorporates an effective supervision scheme to reinforce feature disentanglement through multiple levels of supervision at different granularities. Experimental evaluation on study areas characterized by diverse landscapes, including Koumbia (West Africa, Burkina Faso) and Centre-Val de Loire (central Europe, France), demonstrates the effectiveness of the proposed approach.

Landform mapping is the initial step of many geomorphological analyses (e.g. assessment of natural hazards and natural resources) and requires vast resources to be applied to wide areas at high-resolution. Among geomorphological objects, we focus on glacial moraine mapping, since it is a task relevant to many fields (e.g. paleoclimate and glacial geomorphology). Here we proposed to exploit the potential of Deep Learning-based approaches to map moraine landforms by exploiting multi-source remote sensing imagery. To this end, we propose the first Deep Learning model to map glacial moraines, namely MorNet. As multi-source remote sensing information, we combine together three different sources: Topographic (Pleiades-derived DSM), Multispectral (Sentinel-2), and SAR (Sentinel-1) data. To cope with such heterogeneous information, the proposed model has a dedicated branch for each input source and, a late fusion mechanism is leveraged to combine them with the aim to provide the final mapping. The performance of the MorNet model is evaluated on several glacier valleys in China in the Himalayan range. This area contains minimally eroded moraines, so they are well-defined and of varied morphology. The behavior of the proposed method is compared to models using individual mono-source models in order to highlight the benefit to simultaneously leverage multi-source information. The use of multi-source data allows MorNet to exploit the complementarity of the three input sources and improve its performance from an f1-score of about 41.6 using a single source to 52.8 using three sources. MorNet provides a first-order moraine map through its ability to identify well-defined moraines. Consequently, MorNet can identify areas likely to contain moraines and intends to be used as a tool by experts to facilitate and support large-scale mapping.

Abstract With real-world network systems typically comprising a large number of interactive components and inherently dynamic, Graph Continual Learning (GCL) has gained increasing popularity in recent years. Furthermore, most applications involve multiple entities and relationships with associated attributes, which has led to widely adopting Heterogeneous Information Networks (HINs) for capturing such rich structural and semantic meaning. In this context, we deal with the problem of learning multi-type node representations in a time evolving graph setting, harnessing the expressive power of Graph Neural Networks (GNNs). To this purpose, we propose a novel framework, named DyHANE—Dynamic Heterogeneous Attributed Network Embedding, which dynamically identifies a representative sample of multi-typed nodes as training set and updates the parameters of a GNN module, enabling the generation of up-to-date representations for all nodes in the network. We show the advantage of employing HINs on a data-incremental classification task. We compare the results obtained by DyHANE on a multi-step, incremental heterogeneous GAT model trained on a sample of changed and unchanged nodes, with the results obtained by either the same model trained from scratch or the same model trained solely on changed nodes. We demonstrate the effectiveness of the proposed approach in facing two major related challenges: (i) to avoid model re-train from scratch if only a subset of the network has been changed and (ii) to mitigate the risk of losing established patterns if the new nodes exhibit unseen properties. To the best of our knowledge, this is the first work that deals with the task of (deep) graph continual learning on HINs.

Accurate crop type mapping using satellite imagery is crucial for food security, yet accurately distinguishing between crops with similar spectral signatures is challenging. This study assessed the performance of Sentinel-2 (S2) time series (spectral bands and vegetation indices), Sentinel-1 (S1) time series (backscattering coefficients and polarimetric parameters), alongside phenological features derived from both S1 and S2 time series (harmonic coefficients and median features), for classifying sunflower, soybean, and maize. Random Forest (RF), Multi-Layer Perceptron (MLP), and XGBoost classifiers were applied across various dataset configurations and train-test splits over two study sites and years in France. Additionally, the InceptionTime classifier, specifically designed for time series data, was tested exclusively with time series datasets to compare its performance against the three general machine learning algorithms (RF, XGBoost, and MLP). The results showed that XGBoost outperformed RF and MLP in classifying the three crops. The optimal dataset for mapping all three crops combined S1 backscattering coefficients with S2 vegetation indices, with comparable results between phenological features and time series data (mean F1 scores of 89.9% for sunflower, 76.6% for soybean, and 91.1% for maize). However, when using individual satellite sensors, S1 phenological features and time series outperformed S2 for sunflower, while S2 was superior for soybean and maize. Both phenological features and time series data produced close mean F1 scores across spatial, temporal, and spatiotemporal transfer scenarios, though median features dataset was the best choice for spatiotemporal transfer. Polarimetric S1 data did not yield effective results. The InceptionTime classifier further improved classification accuracy over XGBoost for all crops, with the degree of improvement varying by crop and dataset (the highest mean F1 scores of 90.6% for sunflower, 86.0% for soybean, and 93.5% for maize).

Positive and unlabelled (PU) learning for multi-variate time series classification refers to build a binary classification model when only a small set of positive and a large set of unlabelled samples are accessible at training stage. Different from binary semi-supervised scenario in which the training set contains labelled samples from both positive and negative classes, in the PU learning setting, only positive samples are labelled due to cost-restriction or issues related to defining what belongs to the negative class. With the objective to deal with this challenging task, here, we propose a new deep learning framework, referred as DMTS-PUL. Our method has two different steps: firstly, it selects a set of reliable negative samples from the set of unlabelled data and, successively, it iteratively enriches the training data by selecting pseudo-labels to train a binary classification model via self-training. Experimental evaluations on several benchmarks have highlighted the quality of DMTS-PUL w.r.t. competing approaches and the obtained findings have pointed out the suitability of our proposal when only small amounts of positive labelled samples are available.

Sensitivity of lidar metrics to scan angle can affect the robustness of area-based approach (ABA) models, and modelling the interplay of scan geometry and terrain properties can be complex. The study hypothesizes that neural networks can manage the interplay of lidar acquisition parameters, terrain properties, and vegetation characteristics to improve ABA models. The study area is in Massif des Bauges Natural Regional Park, eastern France, comprising 291 field plots in a mountainous environment with broadleaf, coniferous, and mixed forest types. Field plots were scanned with a high overlap from multiple flight lines and the corresponding point clouds were considered independently to expand the standard ABA dataset (291 observations) to create a dataset containing 1095 independent observations. Computation of lidar, terrain, and scan angle metrics for each point cloud associated each observation in the expanded dataset with the scan information in addition to the lidar and terrain information. A multilayer perceptron (MLP) was used to model basal area and total volume to compare the predictions resulting from standard and expanded ABA datasets. With expanded datasets containing lidar, terrain, and scan information, the R2 for the median predictions per plot were higher (R2 of 0.83 and 0.85 for BA and Vtot) than predictions with standard datasets (R2 of 0.66(BA) and 0.71(Vtot)) containing only lidar metrics. It also outperformed an MLP model for a dataset with lidar and terrain information [R2 of 0.77(BA and Vtot)]. The MLP performed better than Random forest regression, which could not sufficiently exploit additional terrain and scan information.

This study aims to enhance rapeseed field detection accuracy using Sentinel-1 (S1) time series data and addressing challenges in collecting ground samples. The proposed solutions include model transfer between years without retraining and secondly, developing models with limited training samples. The research evaluates the performance of Random Forest (RF) and three deep learning (DL) algorithms: Long Short-Term Memory Fully Convolutional Network (LSTM-FCN), InceptionTime, and Multi-layer Perceptron (MLP). All four algorithms are evaluated initially with abundant ground samples and later with smaller sample sizes (100, 300, 500 and 1000 samples). Model transferability is tested across years. The impact of S1 image count on transfer accuracy is examined. Additionally, the effect of the phenological shift in the rapeseed growth cycle of 15 and 30 days between the training and test years was also investigated. The findings demonstrate strong model performance when training and testing occur in the same year (F1-score up to 95%). Within sample sizes of 300 to 1000, RF and InceptionTime stand out with high accuracy (F1-score>90%). When employing different years for training and testing with abundant sample sizes, all four algorithms correctly classified rapeseed (F1-score between 85.5% and 92.7%). In cases of a reduced number of images, the performance of InceptionTime and LSTM-FCN decreased (16% decrease in the F1-score), while RF and MLP maintain their performance. Notably, RF outperforms DL algorithms with an F1-score of 89.1%. In the context of a phenological shift, only InceptionTime and LSTM-FCN demonstrated strong performance (F1-score between 87.7% and 92.6%).

Nowadays, satellite image time series (SITS) are commonly employed to derive land-cover maps (LCM) to support decision makers in a variety of land management applications. In the most general workflow, the production of LCM strongly relies on available GT data to train supervised machine learning models. Unfortunately, these data are not always available due to time-consuming and costly field campaigns. In this scenario, the possibility to transfer a model learnt on a particular year (source domain) to a successive period of time (target domain), over the same study area, can save time and money. Such a kind of model transfer is challenging due to different acquisition conditions affecting each time period, thus resulting in possible distribution shifts between source and target domains. In the general field of machine learning, unsupervised domain adaptation (UDA) approaches are well suited to cope with the learning of models under distribution shifts between source and target domains. While widely explored in the general computer vision field, they are still underinvestigated for SITS-based land-cover mapping, especially for the temporal transfer scenario. With the aim to cope with this scenario in the context of SITS-based land-cover mapping, here we propose spatially aligned domain-adversarial neural network, a framework that combines both adversarial learning and self-training to transfer a classification model from a time period (year) to a successive one on a specific study area. Experimental assessment on a study area located in Burkina Faso characterized by challenging operational constraints demonstrates the significance of our proposal. The obtained results have shown that our proposal outperforms all the UDA competing methods by 7 to 12 points of F1-score across three different transfer tasks.

Up-to-date canopy height model (CHM) estimates are of key importance for forest resource monitoring and disturbance analysis. In this article, we present a study on the potential of deep learning (DL) for the regression of forest height from TanDEM-X bistatic interferometric synthetic aperture radar (InSAR) data. We propose a novel fully convolutional neural network framework, trained in a supervised manner using reference CHM measurements derived from the LiDAR LVIS airborne sensor from NASA. The reference measurements were acquired during the joint NASA–ESA 2016 AfriSAR campaign over five sites in Gabon, Africa, characterized by the presence of different kinds of vegetation, spanning from tropical primary forests to mangroves. Together with the DL architecture and training strategy, we present a series of experiments to assess the impact of different input features on the network estimation accuracy (in particular of bistatic InSAR-related ones). When tested on all the considered sites, the proposed DL model achieves an overall performance of 1.46-m mean error, 4.2-m mean absolute error, and 15.06% mean absolute percentage error. Furthermore, we perform a spatial transfer analysis aimed at deriving preliminary insights on the generalization capability of the network when trained and tested on datasets acquired over different locations, combining different kinds of tropical vegetation. The obtained results are promising and already in line with state-of-the-art methods based on both physical-based modeling and data-driven approaches, with the remarkable advantage of requiring only one single TanDEM-X acquisition at inference time.

Crop yield forecasting during an ongoing season is crucial to ensure food security and commodity markets. For this reason, here, a scalable approach to forecast corn yields at the field-level using machine learning and satellite imagery from Sentinel-2 and Landsat missions is proposed. The model, evaluated on 1319 corn fields in the U.S. Corn Belt from 2017 to 2022, integrates biophysical parameters from Sentinel-2, Land Surface Temperature (LST) from Landsat, and agroclimatic data from ERA5 reanalysis dataset. Resampling the time series over thermal time significantly enhances predictive performance. The addition of LST to our model further improves in-season yield forecasting, through its capacity to detect early drought, which is not immediately visible to optical sensors such as the Sentinel-2. Finally, we propose a new two-stage machine learning strategy to mitigate early season partially available data. It consists in extending the current time series on the basis of complete historical data and adapting the model inference according to the crop progress.

This study investigates the potential of Sentinel-1 (S1) multi-temporal data for the early-season mapping of the rapeseed crop. Additionally, we explore the effectiveness of limiting the portion of a considered time series to map rapeseed fields. To this end, we conducted a quantitative analysis to assess several temporal windows (periods) spanning different phases of the rapeseed phenological cycle in the following two scenarios relating to the availability or constraints of providing ground samples for different years: (i) involving the same year for both training and the test, assuming the availability of ground samples for each year; and (ii) evaluating the temporal transferability of the classifier, considering the constraints of ground sampling. We employed two different classification methods that are renowned for their high performance in land cover mapping: the widely adopted random forest (RF) approach and a deep learning-based convolutional neural network, specifically the InceptionTime algorithm. To assess the classification outcomes, four evaluation metrics (recall, precision, F1 score, and Kappa) were employed. Using S1 time series data covering the entire rapeseed growth cycle, the tested algorithms achieved F1 scores close to 95% on same-year training and testing, and 92.0% when different years were used, both algorithms demonstrated robust performance. For early rapeseed detection within a two-month window post-sowing, RF and InceptionTime achieved F1 scores of 67.5% and 77.2%, respectively, and 79.8% and 88.9% when extended to six months. However, in the context of temporal transferability, both classifiers exhibited mean F1 scores below 50%. Notably, a 5-month time series, covering key growth stages such as stem elongation, inflorescence emergence, and fruit development, yielded a mean F1 score close to 95% for both algorithms when trained and tested in the same year. In the temporal transferability scenario, RF and InceptionTime achieved mean F1 scores of 92.0% and 90.0%, respectively, using a 5-month time series. Our findings underscore the importance of a concise S1 time series for effective rapeseed mapping, offering advantages in data storage and processing time. Overall, the study establishes the robustness of RF and InceptionTime in rapeseed detection scenarios, providing valuable insights for agricultural applications.

Standard supervised classification methods make the assumption that the training data is fully annotated thus requiring an a-priory labelling process which is both costly and time-consuming. To relax this requirement, many different flavors of weakly supervised learning have been proposed. Among weakly supervised learning strategies, Positive Unlabelled learning (PUL) is gaining attention from the research community due to the wide spectrum of applications it can fit. However, the majority of research studies related to PUL only consider binary classification tasks while real-world applications commonly involve multiple categories. To deal with this limitation, Multi-Positive Unlabelled learning (MPUL) has been recently introduced to learn from examples labelled with multiple positive labels and a single unknown negative label. Up to today, only a limited number of research works were proposed to cope with this more general setting. In this paper, we propose a new MPUL framework based on deep learning strategies. Our framework, named ProtoMPUL (Prototype based Multi-Positive and Unlabelled Learning), combines metric learning and clustering strategies to model the set of positive classes as well as to characterize the unknown negative one. Experimental evaluations on real-world benchmarks considering recent MPUL competitors demonstrates that the proposed framework achieves state-of-the-art performances, thus supporting the validity of the proposed approach.

Forest ecosystems play a fundamental role in natural balances and climate mechanisms through their contribution to global carbon storage. Their sustainable management and conservation is crucial in the current context of global warming and biodiversity conservation. To tackle such challenges, earth observation data have been identified as a valuable source of information. While earth observation data constitute an unprecedented opportunity to monitor forest ecosystems, its effective exploitation still poses serious challenges since multimodal information needs to be combined to describe complex natural phenomena. To deal with this particular issue in the context of structure and biophysical variables estimation for forest characterization, we propose a new deep learning-based fusion strategy to combine together high density three-dimensional (3-D) point clouds acquired by airborne laser scanning with high-resolution optical imagery. In order to manage and fully exploit the available multimodal information, we implement a two-branch late fusion deep learning architecture taking advantage of the specificity of each modality. On the one hand, a 2-D CNN branch is devoted to the analysis of Sentinel-2 time series data, and on the other hand, a multilayer perceptron branch is dedicated to the processing of LiDAR-derived information. The performance of our framework is evaluated on two forest variables of interest: total volume and basal area at stand level. The obtained results underline that the availability of multimodal remote sensing data is not a direct synonym of performance improvements but, the way in which they are combined together is of paramount importance.

Satellite image time series (SITS) collected by modern Earth Observation (EO) systems represent a valuable source of information that supports several tasks related to the monitoring of the Earth surface dynamics over large areas. A main challenge is then to design methods able to leverage the complementarity between the temporal dynamics and the spatial patterns that characterize these data structures. Focusing on land cover classification (or mapping) tasks, the majority of approaches dealing with SITS data only considers the temporal dimension, while the integration of the spatial context is frequently neglected. In this work, we propose an attentive spatial temporal graph convolutional neural network that exploits both spatial and temporal dimensions in SITS. Despite the fact that this neural network model is well suited to deal with spatio-temporal information, this is the first work that considers it for the analysis of SITS data. Experiments are conducted on two study areas characterized by different land cover landscapes and real-world operational constraints (i.e., limited labeled data due to acquisition costs). The results show that our model consistently outperforms all the competing methods obtaining a performance gain, in terms of F-Measure, of at least 5 points with respect to the best competing approaches on both benchmarks.

Extensive research studies have been conducted in recent years to exploit the complementarity among multisensor (or multimodal) remote sensing data for prominent applications such as land cover mapping. In order to make a step further with respect to previous studies, which investigate multitemporal SAR and optical data or multitemporal/multiscale optical combinations, here, we propose a deep learning framework that simultaneously integrates all these input sources, specifically multitemporal SAR/optical data and fine-scale optical information at their native temporal and spatial resolutions. Our proposal relies on a patch-based multibranch convolutional neural network (CNN) that exploits different per-source encoders to deal with the specificity of the input signals. In addition, we introduce a new self-distillation strategy to boost the per-source analyses and exploit the interplay among the different input sources. This new strategy leverages the final prediction of the multisource framework to guide the learning of the per-source CNN encoders supporting the network to learn from itself. Experiments are carried out on two real-world benchmarks, namely, the Reunion island (a French overseas department) and the Dordogne study site (a southwest department in France), where the annotated reference data were collected under operational constraints (sparsely annotated ground-truth data). Obtained results providing an overall classification accuracy of about 94% (respectively, 88%) on the Reunion island (respectively, the Dordogne) study site highlight the effectiveness of our framework based on CNNs and self-distillation to combine heterogeneous multisensor remote sensing data and confirm the benefit of multimodal analysis for downstream tasks such as land cover mapping.

Nowadays, huge volume of satellite images, via the different Earth Observation missions, are constantly acquired and they constitute a valuable source of information for the analysis of spatiotemporal phenomena. However, it can be challenging to obtain reference data associated to such images to deal with land use or land cover changes as often the nature of the phenomena under study is not known a priori. With the aim to deal with satellite image analysis, considering a real-world scenario, where reference data cannot be available, in this article, we present a novel end-to-end unsupervised approach for change detection and clustering for satellite image time series (SITS). In the proposed framework, we first create bitemporal change masks for every couple of consecutive images using neural network autoencoders (AEs). Then, we associate the extracted changes to different spatial objects. The objects sharing the same geographical location are combined in spatiotemporal evolution graphs that are finally clustered accordingly to the type of change process with gated recurrent unit (GRU) AE-based model. The proposed approach was assessed on two real-world SITS data supplying promising results.

The unprecedented possibility to acquire high resolution Satellite Image Time Series (SITS) data is opening new opportunities to monitor the different aspects of the Earth Surface but, at the same time, it is raising up new challenges in term of suitable methods to analyze and exploit such huge amount of rich image data. One of the main tasks associated to SITS data analysis is related to land cover mapping. Due to operational constraints, the collected label information is often limited in volume and obtained at coarse granularity level carrying out inexact and weak knowledge that can affect the whole process. To cope with such issues, in the context of object-based SITS land cover mapping, we propose a new deep learning framework, named TASSEL (aTtentive weAkly Supervised Satellite image time sEries cLassifier), to deal with the weak supervision provided by the coarse granularity labels. Our framework exploits the multifaceted information conveyed by the object-based representation considering object components instead of aggregated object statistics. Furthermore, our framework also produces an additional outcome that supports the model interpretability. Quantitative and qualitative experimental evaluations are carried out on two real-world scenarios. Results indicate that not only TASSEL outperforms the competing approaches in terms of predictive performances, but it also produces valuable extra information that can be practically exploited to interpret model decisions.

Applications in which researchers aim to extract a single land type from remotely sensed data are quite common in practical scenarios: extract the urban footprint to make connections with socio-economic factors; map the forest extent to subsequently retrieve biophysical variables and detect a particular crop type to successively calibrate and deploy yield prediction models. In this scenario, the (positive) targeted class is well defined, while the negative class is difficult to describe. This one-class classification setting is also referred to as positive unlabelled learning (PUL) in the general field of machine learning. To deal with this challenging setting, when satellite image time series data are available, we propose a new framework named positive and unlabelled learning of satellite image time series (PUL-SITS). PUL-SITS involves two different stages: In the first one, a recurrent neural network autoencoder is trained to reconstruct only positive samples with the aim to higight reliable negative ones. In the second stage, both labelled and unlabelled samples are exploited in a semi-supervised manner to build the final binary classification model. To assess the quality of our approach, experiments were carried out on a real-world benchmark, namely Haute-Garonne, located in the southwest area of France. From this study site, we considered two different scenarios: a first one in which the process has the objective to map Cereals/Oilseeds cover versus the rest of the land cover classes and a second one in which the class of interest is the Forest land cover. The evaluation was carried out by comparing the proposed approach with recent competitors to deal with the considered positive and unlabelled learning scenarios.

European satellite missions Sentinel-1 (S1) and Sentinel-2 (S2) provide at high spatial resolution and high revisit time, respectively, radar and optical images that support a wide range of Earth surface monitoring tasks, such as Land Use/Land Cover mapping. A long-standing challenge in the remote sensing community is about how to efficiently exploit multiple sources of information and leverage their complementarity, in order to obtain the most out of radar and optical data. In this work, we propose to deal with land cover mapping in an object-based image analysis (OBIA) setting via a deep learning framework designed to leverage the multi-source complementarity provided by radar and optical satellite image time series (SITS). The proposed architecture is based on an extension of Recurrent Neural Network (RNN) enriched via a modified attention mechanism capable to fit the specificity of SITS data. Our framework also integrates a pretraining strategy that allows to exploit specific domain knowledge, shaped as hierarchy over the set of land cover classes, to guide the model training. Thorough experimental evaluations, involving several competitive approaches were conducted on two study sites, namely the Reunion island and a part of the Senegalese groundnut basin. Classification results, 79% of global accuracy on the Reunion island and 90% on the Senegalese site, respectively, have demonstrated the suitability of the proposal.

Mapping irrigated plots is essential for better water resource management. Today, the free and open access Sentinel-1 (S1) and Sentinel-2 (S2) data with high revisit time offers a powerful tool for irrigation mapping at plot scale. Up to date, few studies have used S1 and S2 data to provide approaches for mapping irrigated plots. This study proposes a method to map irrigated plots using S1 SAR (synthetic aperture radar) time series. First, a dense temporal series of S1 backscattering coefficients were obtained at plot scale in VV (Vertical-Vertical) and VH (Vertical-Horizontal) polarizations over a study site located in Catalonia, Spain. In order to remove the ambiguity between rainfall and irrigation events, the S1 signal obtained at plot scale was used conjointly to S1 signal obtained at a grid scale (10 km × 10 km). Later, two mathematical transformations, including the principal component analysis (PCA) and the wavelet transformation (WT), were applied to the several SAR temporal series obtained in both VV and VH polarization. Irrigated areas were then classified using the principal component (PC) dimensions and the WT coefficients in two different random forest (RF) classifiers. Another classification approach using one dimensional convolutional neural network (CNN) was also performed on the obtained S1 temporal series. The results derived from the RF classifiers with S1 data show high overall accuracy using the PC values (90.7%) and the WT coefficients (89.1%). By applying the CNN approach on SAR data, a significant overall accuracy of 94.1% was obtained. The potential of optical images to map irrigated areas by the mean of a normalized differential vegetation index (NDVI) temporal series was also tested in this study in both the RF and the CNN approaches. The overall accuracy obtained using the NDVI in RF classifier reached 89.5% while that in the CNN reached 91.6%. The combined use of optical and radar data slightly enhanced the classification in the RF classifier but did not significantly change the accuracy obtained in the CNN approach using S1 data.

The use of Very High Spatial Resolution (VHSR) imagery in remote sensing applications is nowadays a current practice whenever fine-scale monitoring of the earth’s surface is concerned. VHSR Land Cover classification, in particular, is currently a well-established tool to support decisions in several domains, including urban monitoring, agriculture, biodiversity, and environmental assessment. Additionally, land cover classification can be employed to annotate VHSR imagery with the aim of retrieving spatial statistics or areas with similar land cover. Modern VHSR sensors provide data at multiple spatial and spectral resolutions, most commonly as a couple of a higher-resolution single-band panchromatic (PAN) and a coarser multispectral (MS) imagery. In the typical land cover classification workflow, the multi-resolution input is preprocessed to generate a single multispectral image at the highest resolution available by means of a pan-sharpening process. Recently, deep learning approaches have shown the advantages of avoiding data preprocessing by letting machine learning algorithms automatically transform input data to best fit the classification task. Following this rationale, we here propose a new deep learning architecture to jointly use PAN and MS imagery for a direct classification without any prior image sharpening or resampling process. Our method, namely M u l t i R e s o L C C , consists of a two-branch end-to-end network which extracts features from each source at their native resolution and lately combine them to perform land cover classification at the PAN resolution. Experiments are carried out on two real-world scenarios over large areas with contrasted land cover characteristics. The experimental results underline the quality of our method while the characteristics of the proposed scenarios underline the applicability and the generality of our strategy in operational settings.