Crop type mapping is crucial for improving crop yield predictions, informing agricultural policy decisions, and addressing the impacts of climate change. Despite achieving state-of-the-art performance, current transformer-based architectures that use Satellite Image Time Series (SITS) require costly labelled data and struggle to generalize to new regions. To address this limitation, this paper presents a method that leverages multimodal information from field images and satellite data to address the scarcity of reference data in crop type mapping. The proposed approach consists of three main steps: (1) multimodal data preparation, (2) multimodalpseudo-labelling, and (3) SITS-based crop type mapping. The first step processes both data sources for facilitating their integration, while the novelty lies in the second step, where additionalpseudo-labels are generated by classifying crop parcels with both SITS and field image availability, leveraging multimodal information to produce reliable label predictions. Finally, the third step produces a crop type map for the entire study area using the SITS and the enriched training set. Experimental results at the European level, using the 2018 Land use and land cover survey (LUCAS) field images paired with Sentinel-2 SITS, demonstrate the effectiveness of integrating satellite top-view data with ground-level field images. Thepseudolabelled data achieves performance comparable to models trained on the complete labelled dataset, while substantially reducing the need for extensive labelled data collection.

Monitoring food security and nutrition remains a complex challenge, extremely relevant in underdeveloped countries. In this context, we aim to exploit Earth Observation (EO)-derived geospatial data to extract meaningful indicators that can support the monitoring of food production and consumption. In greater detail, we used EO-derived geospatial data to model the spatial configuration and features of surrounding areas of the markets, while volunteered geographic OpenStreetMap (OSM) data were used to map the physical accessibility to the markets and nearby facilities. The extracted variables were correlated with production sites to automatically classify different market properties, specifically market infrastructure and characterization. Preliminary results obtained in Rwanda reveal a correlation between the analyzed geospatial features and market characteristics, offering valuable insights for improving food security strategies.

This paper presents a novel data-driven approach for estimating cereal food flows at the regional level. To this end, a Random Forest machine learning regression model is trained with features extracted from Land-Use Land-Cover (LULC) maps, spatially explicit socio-economic indicators, geospatial and Earth Observation (EO) data. To identify the driving factors influencing cereal food flows, the significance of the considered feature set is estimated using the permutation of out-of-bag predictor importance. From the experiments conducted at the county level in the Continental United States (CONUS), it turned out that the population and the total area cultivated as corn from both origin and destination regions are the most influential features. Moreover, the results obtained on the test set using EO and geospatial data outperformed those obtained using spatially explicit socio-economic indicators and are comparable with the ones achieved using LULC information.

The increasing availability of high-resolution, open-access satellite data facilitates the production of global Land Cover (LC) maps, an essential source of information for managing and monitoring natural and human-induced processes. However, the accuracy of the obtained LC maps can be affected by the discrepancy between the spatial resolution of the satellite images and the extent of the LC present in the scene. Indeed, several pixels may be misclassified because of their mixed spectral signatures, i.e., more than two LC classes are present in the pixel. To solve this problem, this paper proposes an approach that explores the possibility of using simple but effective unmixing approaches to enhance the classification accuracy of the mixed spectral pixels. The results showed that several pixels, including buildings and grassland LC, are typically classified as cropland. By unmixing their spectral content, it is possible to extract the most prevalent class within the area of each pixel to update the classification map, thus sharply increasing the map accuracy. These promising preliminary results indicate the potential for broader applicability and efficiency in global LC mapping.

Deep Learning (DL) models are extremely effective for crop-type mapping. However, they generalize poorly when there is a temporal shift between the Satellite Image Time Series (SITS) acquired in the source domain (where the model is trained) and the target domain (never seen by the network). To address this challenge, this paper proposes an Explainable Artificial Intelligence (xAI) approach that leverages the interpretability of the inner workings of transformer encoders to automatically capture and mitigate the temporal shift between SITS acquired in different regions. The Positional Encoding (PE) output computed on the source SITS is used as a proxy to quantify the temporal shift with respect to the PE output obtained on the target SITS. This condition allows us to re-align the latter to the representation that the model natively adopts to discriminate crop types through a Dynamic Time Warping (DTW) approach. Compared to the baseline architecture, the proposed method increases the Overall Accuracy (OA) up to 8% on the TimeMatch benchmark dataset.

The continuous stream of high spatial resolution satellite data offers the opportunity to regularly produce land cover (LC) maps. To this end, Transformer deep learning (DL) models have recently proven their effectiveness in accurately classifying long time series (TS) of satellite images. The continual generation of regularly updated LC maps can be used to analyze dynamic phenomena and extract multi-temporal information. However, several challenges need to be addressed. Our paper aims to study how the performance of a Transformer model changes when classifying TS of satellite images acquired in years later than those in the training set. In particular, the behavior of the attention in the Transformer model is analyzed to determine when the information provided by the initial training set needs to be updated to keep generating accurate LC products. Preliminary results show that: (i) the selection of the positional encoding strategy used in the Transformer has a significant impact on the classification accuracy obtained with multi-year TS, and (ii) the most affected classes are the seasonal ones.

Earth Observation (EO) data processing faces challenges due to large volumes, multiple sources, and diverse formats. To address this issue, this paper presents a scalable and parallelizable workflow using Apache Airflow, capable of integrating Machine Learning (ML) and Deep Learning (DL) models with Modular Supercomputing Architecture (MSA) systems. To test the workflow, we considered the production of large-scale Land-Cover (LC) maps as a case study. The workflow manager, Airflow, offers scalability, extensibility, and programmable task definition in Python. It allows us to execute different steps of the workflow in different High-Performance Computing (HPC) systems. The workflow is demonstrated on the Dynamical Exascale Entry Platform (DEEP) and Jülich Research on Exascale Cluster Architectures (JURECA) hosted at the Jülich Supercomputing Centre (JSC), a platform that incorporates heterogeneous JSC systems.

Agricultural field polygons within smallholder farming systems are essential to facilitate the collection of geo-spatial data useful for farmers, managers, and policymakers. However, the limited availability of training labels poses a challenge in developing supervised methods to accurately delineate field boundaries using Earth Observation (EO) data. This letter introduces an open data set for training and benchmarking machine learning methods to delineate agricultural field boundaries in polygon format. The large-scale data set consists of 439,001 field polygons divided into 62 tiles of approximately 5×5 km distributed across Vietnam and Cambodia, covering a range of fields and diverse landscape types. The field polygons have been meticulously digitized from satellite images, following a rigorous multi-step quality control process and topological consistency checks. Multi-temporal composites of Sentinel-2 (S2) images are provided to ensure cloud-free data. We conducted an experimental analysis testing a state-of-the-art Deep Learning (DL) workflow based on fully convolutional networks, contour closing, and polygonization. We anticipate that this large-scale data set will enable researchers to further enhance the delineation of agricultural fields in smallholder farms and to support the achievement of the Sustainable Development Goals (SDG). The data set can be downloaded from https://doi.org/10.17026/dans-xy6-ngg6.

Land cover maps generated by the classification of remote sensing data allow for monitoring Earth processes and the dynamics of objects and phenomena. For accurate land cover variability quantification in environmental monitoring, maps need to be spatiotemporally consistent, continually updated, and indicate permanent changes. However, producing frequent and spatiotemporally consistent land cover maps is challenging because it involves balancing the need for temporal consistency with the risk of missing real changes. In this work, we propose a scalable and semi-automatic method for generating annual land cover maps with labels that are consistently applied from one year to the next. It uses a Transformer deep learning model as a classifier, which is trained on satellite time series of images using High Performance Computing (HPC). The trained model can generate stable maps by shifting the prediction window along the temporal direction. The effectiveness of the proposed approach is tested qualitatively and quantitatively on a multi-annual Sentinel-2 dataset acquired over a three-year period in a study area located in the southern Italian Alps.

To perform specific environmental analyses with high accuracy and spatial resolution, typically dedicated Earth Observation (EO) data are acquired via aircraft or drones. Although valuable, these data can be: (i) limited and sparse in time and space due to their acquisition cost, and (ii) asynchronous to field data collection. To consistently ingest asynchronous EO data and field surveys, this paper generates a spatio-temporal framework by exploiting the ability of Sentinel-1 satellites to provide frequent EO data with global coverage. Experiments, conducted in Indonesia to estimate changes in forest Above-Ground Biomass (AGB) between 2017 and 2019, demonstrate the ability of the spatio-temporal framework to integrate Light Detection and Ranging (LIDAR) data acquired in 2020. The method achieved a$\mathrm{R}^{2}$of 0.76 and a RMSE of 21.24 compared to 0.50 and 0.57 and 28.65 and 23.93 for the standard bi-temporal approach (using field data and Sentinel-1 data) and the bi-temporal approach including the LIDAR data without any adaptation, respectively.

This paper presents an approach that aims to produce a Time-Series (TS) of consistent Land-Cover (LC) maps, typically needed to perform environmental monitoring. First, it creates an annual training set for each TS to be classified, leveraging on publicly available thematic products. These annual training sets are then used to generate a set of preliminary LC maps that allow for the identification of the unchanged areas, i.e., the stable temporal component. Such areas can be used to define an informative and reliable multi-year training set, by selecting samples belonging to the different years for all the classes. The multi-year training set is finally employed to train a unique multi-year Long Short Term Mem-ory (LSTM) model, which enhances the consistency of the annual LC maps. The preliminary results carried out on three TSs of Sentinel 2 images acquired in Italy in 2018,2019 and 2020 demonstrates the capability of the method to improve the consistency of the annual LC maps. The agreement of the obtained maps is$\approx 78{\%}$, compared to the$\approx 74{\%}$achieved by the LSTM models trained separately.

This letter presents a novel approach to tree-top detection in heterogeneous forest structures characterized by mixed species using high-density light detection and ranging (LiDAR) data. Although literature techniques can achieve accurate results in even-size and even-age homogeneous forests, they detect several false tree tops in forests characterized by variable crown dimensions. To solve this problem, the proposed method 1) identifies a preliminary set of candidate tree tops (CTPs) used to build a triangulated network; 2) performs an edge-based local forest analysis to identify groups of CTPs having the highest probability of belonging to the same crown; and 3) removes false tree tops according to a local directed graph analysis. To address large-scale forest analysis, the method exploits the Delaunay triangulation that efficiently defines a network topology made up only by relevant edges, thus sharply reducing the edge-based analyses to be performed. Given the triangulated network properties, the computational effort of the local analysis is not affected by the network size. The method has been tested in a mixed multi-layer multi-age forest located in the southern Italian Alps. The results obtained demonstrate that this computationally scalable algorithm outperforms standard tree-top detection methods increasing the overall detection accuracy up to 15.3%.

This letter proposes a novel method based on Deep Learning (DL) to forest species classification in airborne Light Detection and Ranging (LiDAR) data. Differently from the state-of- the-art approaches, the proposed method: (1) does not assume any prior knowledge either on the forest to be classified or on the sensor used to acquire the LiDAR data, and (2) can be applied to heterogeneous forest characterized by mixed species. First, the 3D point cloud of each individual tree is decomposed into 8 angular sectors to generate a multi-slices representation of the vertical structure of the tree. This representation models the foliage, the stem and the branches of the tree crown as well as depicts the internal and external crown properties. Then, a Multi-View CNN (MVCNN) DL automatically extracts features used to discriminate the different tree species. This network is pre-trained on the massive ImageNet database, thus guaranteeing fast convergence with a relatively small number of ground reference data. Experiments were carried out on high density airborne LiDAR data collected over a multi-layer multi-age forest characterized by four conifers and three broadleaf species. The proposed method outperformed the state-of-the-art approaches increasing the Overall Accuracy (OA) up to 16% and 18.9% compared to a DL and a shallow tree species classification methods, respectively. When applied to coniferous or broadlaef forests, the proposed method showed an increase of OA 10.1% and 15.9% (for conifers), and 9.5% and 21.6% (for broadleafs) compared to the DL and shallow methods, respectively.

Supervised classification of remote sensing data requires a large number of high-quality annotated samples. At the operational level, the definition of a large training set by photograph interpretation is costly and time-consuming. The manual annotation activity is typically supported by high-resolution satellite data. Therefore, when working at country or continental scale, it is necessary to efficiently access large archives of remotely sensed data. To address these issues, this letter presents an interactive strategy implemented in a cloud-computing platform for defining effective training sets with significantly reduced human effort. This is achieved by combining active learning (AL) and self-paced learning (SPL) techniques. First, an initial training set is used to classify the pool of unlabeled samples. Then, the method progressively adds high-confidence samples, selected through an SPL strategy, and low-confidence samples selected considering an AL strategy. While the high-confidence sample labels are self-paced, the low-confidence ones are manually assigned. The cloud-computing platform allows the: 1) definition of a complete training set in a fast and efficient way and 2) access to a multipetabyte catalog of satellite imagery. Experiments carried out on the Google Earth Engine (GEE) Platform demonstrate the effectiveness of the proposed strategy compared to the standard manual annotation.

The accurate classification of remote sensing (RS) data at large scale is typically hampered by the availability of training data representative of the whole study area. To solve this problem, we propose a method that aims to enlarge existing training sets leveraging publicly available thematic products. First, the available thematic product of the target domain ($D_{T}$) (RS data geographically distant from the training samples) is processed to extract few labeled target samples. These labeled target samples are jointly used with the annotated samples of the source domain ($D_{S}$) (RS data where training set is available) to find a mapping space where the data are aligned. This common latent space allows us to enlarge the training set in an unsupervised (no annotated samples from the$D_{T}$are required) but reliable way. The results obtained in Amazon using the Copernicus Global Land Service - Land cover (CGLS-LC) map demonstrate the effectiveness of the method. The enlarged training set achieves an Overall Accuracy (OA) of 87% compared to 80% obtained with the initial training set.

Crown features derived from high-density airborne laser scanning (ALS) data have proven to be effective for forest species classification at the individual tree level. Most of the general state-of-the-art (SoA) techniques rely on coarse-level crown features extracted from ALS data and under-utilize both the spatial and the spectral information available in the point clouds, Moreover, they are designed on the expected properties of the specific analyzed forest. We present a novel species classification approach, based on quantization of the entire 3-D tree crown into smaller elementary crown volumes (ECVs) that effectively captures the spatial distribution of filled (i.e., stem, branch, and foliage) and empty volumes of crowns. In the first step, a data-driven process dynamically tests and compares three quantization strategies to tailor the definition of the ECV to the forest type (e.g., conifer and deciduous forest). In the second step, for each ECV, a histogram vector is made up of features representing the light detection and ranging (LiDAR) point distribution and intensity to model the internal and the external local crown characteristics. Then, tree histogram feature vectors are obtained by stacking all the ECV histogram feature vectors. Finally, classification is performed by a support vector machine (SVM) classifier using the histogram intersection kernel. All experiments were performed on three high-density (50-200 points/m2) ALS data sets of deciduous, conifer, and mixed (i.e., both deciduous and conifer) trees. The higher classification accuracy of the proposed method over the SoA one proves its ability to better capture the crown characteristics of individual trees, including species-specific traits.

Supervised classification algorithms require a sufficiently large set of representative training samples to generate accurate land-cover maps. Collecting reference data is difficult, expensive, and unfeasible at the large scale. To solve this problem, this article introduces a novel approach that aims to extract reliable labeled data from existing thematic products. Although these products represent a potentially useful information source, their use is not straightforward. They are not completely reliable since they may present classification errors. They are typically aggregated at polygon level, where polygons do not necessarily correspond to homogeneous areas. Finally, usually, there is a semantic gap between map legends and remote sensing (RS) data. In this context, we propose an approach that aims to: 1) perform a domain understanding to detect the discrepancies between the thematic map domain and the RS data domain; 2) use RS data contemporary to the map to decompose the thematic product from the semantic and spatial viewpoints; and 3) extract a database of informative and reliable training samples. The database of weak labeled units is used for training an ensemble of classifiers on recent data whose results are then combined in a majority voting rule. Two sets of experimental results obtained on MS images by extracting training samples from a crop type map and the 2018 Corine Land Cover (CLC) map, respectively, confirm the effectiveness of the proposed approach.

Copernicus is the European programme for monitoring the Earth.It consists of a set of systems that collect data from satellites and in-situ sensors, process this data and provide users with reliable and up-to-date information on a range of environmental and security issues.The data and information processed and disseminated puts Copernicus at the forefront of the big data paradigm, giving rise to all relevant challenges, the so-called 5 Vs: volume, velocity, variety, veracity and value.In this short paper, we discuss the challenges of extracting information and knowledge from huge archives of Copernicus data.We propose to achieve this by scale-out distributed deep learning techniques that run on very big clusters offering virtual machines and GPUs.We also discuss the challenges of achieving scalability in the management of the extreme volumes of information and knowledge extracted from Copernicus data.The envisioned scientific and technical work will be carried out in the context of the H2020 project ExtremeEarth which starts in January 2019.

Individual tree detection in Light Detection and Ranging (LiDAR) data has been widely investigated in the literature. However, most of the methods work well on conifers but lead to poor accuracy in broad-leaved forest. The detection of deciduous trees is a complex task due to: (i) multiple local maxima present in the same canopy, and (ii) the tree-top (TP) can be in a different location from the canopy center. This paper presents an automatic technique which exploits high density LiDAR data to refine the detection of deciduous trees. First, the candidate tree-tops (CTPs) are detected using the standard level set method (LSM). Then, the Delaunay triangulation is used to generate a network topology which connects neighboring CTPs. For each pair of connected CTPs, geometrical features are extracted to automatically determine if the CTPs pair belongs to the same tree or to different canopies. The groups of CTPs identified as belonging to the same tree crown are merged into one TP. Preliminary numerical results show that the proposed method halves the commission errors of the initial TP detection by increasing the overall detection accuracy of 8.2%.

The accuracy in classification of remote sensing (RS) images using deep learning architectures is affected by the lack of large sets of training samples. Although a significant effort is currently devoted to generate databases of annotated satellite images, these datasets may not be large enough to accurately model at global level different types of land-cover surfaces. To solve such a problem, this paper presents an unsupervised approach which aims to exploit the RS image that has to be classified and publicly available thematic products to generate a training database of weak samples representative of the considered study area. First, we harmonize the thematic map and the RS image. Then, samples having the highest probability to be correctly associated to their labels are extracted from the map by exploiting the information provided by the RS image to be classified. Finally, the weak labeled samples are used to train a convolutional neural network (CNN). Experimental results obtained training a CNN on Sentinel 2 images with weak labels extracted from the 2018 corine land cover (CLC) map demonstrate the effectiveness of the proposed method.

The repetitive acquisition of airborne light detection and ranging (LiDAR) data for forest surveys is rapidly increasing, thus making possible the forest dynamic analysis. Moreover, the availability of multitemporal data enables the possibility to improve the forest attribute estimates performed at single date, especially when one LiDAR acquisition has a lower pulse density with respect to the other. This letter presents a novel approach that exploits the bitemporal data information to: 1) improve the tree detection at both dates and 2) identify forest changes at single tree level. This is done by using a novel compound approach to the detection of trees in bitemporal data based on the Bayes rule for minimum error. Significant geometric features are extracted for each candidate tree-top and are used to estimate statistical terms employed in the compound approach. The multitemporal information is considered by estimating (in an iterative way) the probabilities of transition, which takes into account the temporal dependence between the LiDAR acquisitions. The proposed approach is evaluated on multitemporal LiDAR data acquired in a coniferous forest located in the Southern Italian Alps. Experimental results confirm the effectiveness of the compound detection that increases the overall accuracy (OA) up to 8.6% with respect to the single-date detection.

Diameter at breast height (DBH) is one of the most important tree parameter for forest inventory. In this paper, we present a novel method for the adaptive and the accurate DBH estimation of trees characterized by small and large stems. The method automatically discriminates among different tree growth models by means of a data-driven technique based on a clustering procedure. First, the method detects young trees belonging to the lowest forest layer by simply considering the vertical structure of the forest. Then, different clusters of mature trees that are expected to share the same growth-model are identified by analyzing the environmental factors that can affect the stem expansion (e.g., topography and forest density). For each detected growth-model cluster, a tailored regression analysis is performed to obtain accurate DBH estimation results. Experiments have been carried out in an homogeneous coniferous forest located in the Alpine mountainous scenario characterized by a complex topography and a wide range of soil fertility. The method was tested on two data sets characterized by different light detection and ranging (LiDAR) point densities and different forest properties. The results obtained demonstrate the effectiveness of having multiple regression models adapted to the different growth models.

This paper aims to provide a compact superresolution formulation specific for multispectral (MS) multiresolution optical data, i.e., images characterized by different scales across different spectral bands. The proposed method, named multiresolution sharpening approach (MuSA), relies on the solution of an optimization problem tailored to the properties of those images. The superresolution problem is formulated as the minimization of an objective function containing a data-fitting term that models the blurs and downsamplings of the different bands and a patch-based regularizer that promotes image self-similarity guided by the geometric details provided by the high-resolution bands. By exploiting the approximately low-rank property of the MS data, the ill-posedness of the inverse problem in hand is strongly reduced, thus sharply improving its conditioning. The state-of-the-art color block-matching and 3D filtering (C-BM3D) image denoiser is used as a patch-based regularizer by leveraging the “plug-and-play” framework: the denoiser is plugged into the iterations of the alternating direction method of multipliers. The main novelties of the proposed method are: 1) the introduction of an observation model tailored to the specific properties of (MS) multiresolution images and 2) the exploitation of the high-spatial-resolution bands to guide the grouping step in the color block-matching and 3D filtering (C-BM3D) denoiser, which constitutes a form of regularization learned from the high-resolution channels. The results obtained on the real and synthetic Sentinel 2 data sets give an evidence of the effectiveness of the proposed approach.

This paper presents an unsupervised approach that extracts reliable labeled units from outdated maps to update them using time series (TS) of recent multispectral (MS) images. The method assumes that: 1) the source of the map is unknown and may be different from remote sensing data; 2) no ground truth is available; 3) the map is provided at polygon level, where the polygon label represents the dominant class; and 4) the map legend can be converted into a set of classes discriminable with the TS of images (i.e., no land-use classes that require manual analysis are considered). First, the outdated map is adapted to the spatial and spectral properties of the MS images. Then, the method identifies reliable labeled units in an unsupervised way by a two-step procedure: 1) a clustering analysis performed at polygon level to detect samples correctly associated to their labels and 2) a consistency analysis to discard polygons far from the distribution of the related land-cover class (i.e., having high probability of being mislabeled). Finally, the map is updated by classifying the recent TS of MS image with an ensemble of classifiers trained using only the reference data derived from the map. The experimental results obtained updating the 2012 Corine Land Cover (CLC) and the GlobLand30 in Trentino Alto Adige (Italy) achieved 93.2% and 93.3% overall accuracy (OA) on the validation data set. The method increased the OA up to 18% and 11.5% with respect to the reference methods on the 2012 CLC and the GlobLand30, respectively.

The increasingly availability of Light Detection and Ranging (LiDAR) data acquired at different times can be used to analyze the forest dynamics at individual tree level. This often requires to deal with LiDAR point clouds having significantly different point densities. To address this issue, this paper presents a method for the fusion of multitemporal Li-DAR data which aims at using the information provided by high density LiDAR data (higher than 10 pts/m2) to improve the single tree parameter estimation of low density data (up to 5 pts/m2) acquired over the same forest at different times. The method first accurately characterizes the crown shapes on the high density data. Then, it uses the obtained estimates to drive the tree parameter estimation on the low density LiDAR data. The method has been tested on a multitemporal dataset acquired in coniferous forests located in the Italian Alps. Experimental results confirmed the effectiveness of the method.

Combining existing thematic vector products and recently acquired satellite images to generate regular updated maps is extremely interesting at operational level. However, employing these maps is not straightforward. They are typically provided at polygon level, where the polygon labels do not necessarily correspond to spectrally homogeneous areas. Moreover, usually there is a semantic gap between the map legend and the set of natural classes discriminable in multispectral images. To overcome these issues, this paper presents a method that first performs a domain understanding to detect the discrepancies between the vector map domain and the multispectral (MS) image domain. Then, it accomplishes a domain modeling which uses a MS image contemporary to the map to extract a set of reliable and informative samples from the map. Finally, the method carries out Domain Adaptation (DA) using a recent MS image to update the map. Experimental results obtained updating a crop thematic map in Czech Republic confirm the effectiveness of the method.

Light Detection and Ranging (LiDAR) data have been widely used to characterize the 3-D structure of the forest. However, their use in a multitemporal framework has been quite limited due to the relevant challenges introduced by the comparison of pairs of point clouds. Because of the irregular sampling of the laser scanner and the complex structure of forest areas, it is not possible to perform a point-to-point comparison between the two data. To overcome these challenges, a novel hierarchical approach to the detection of 3-D changes in forest areas is proposed. The method first detects the large changes (e.g., cut trees) by comparing the Canopy Height Models derived from the two LiDAR data. Then, according to an object-based change detection approach, it identifies the single-tree changes by monitoring both the treetop and the crown volume growth. The proposed approach can compare LiDAR data with significantly different pulse densities, thus allowing the use of many data available in real applications. Experimental results pointed out that the method can accurately detect large changes, exhibiting a low rate of false and missed alarms. Moreover, it can detect changes in terms of single-tree growth, which are consistent with the expected growth rates of the considered areas.

This paper presents a sensor-driven hierarchical domain adaptation method that aims at transferring the knowledge from a source domain (RS image where reference data are available) to a different but related target domain (RS image where no labeled reference data are available) for solving a classification problem. Due to the different acquisition conditions, a difference in the source and target distributions of the features representing the same class is generally expected. To solve this problem, the proposed method takes advantage from the availability of multisensor data to hierarchically detect features subspaces where for some classes data manifolds are partially (or completely) aligned. These feature subspaces are associated with invariant physical properties of classes measured by the sensors in the scene, i.e., measures having almost the same behavior in both domains. The detection of these invariant feature subspaces allows us to infer labels of the target samples that result more aligned to the source data for the considered subset of classes. Then, the labeled target samples are analyzed in the full feature space to classify the remaining target samples of the same classes. Finally, for those classes for which none of the sensors can measure invariant features, we perform the adaptation via a standard active learning technique. Experimental results obtained on two real multisensor data sets confirm the effectiveness of the proposed method.

In this paper, we focus the attention on the superresolution of multispectral (MS) multiresolution images (e.g., Sentinel 2, Aster, MODIS). By taking advantage of the high spatial resolution bands, we minimize an objective function containing a quadratic data fitting term, an edge preserving regularizer, and a patch-based plug-and play prior promoting self-similar images. To cope with the ill-posedness of the problem we i) exploit the fact that the images are approximately low-rank, and ii) propose a hierarchical method which sharpens in the first place the medium resolution bands and then the coarse resolution ones. The optimization is solved with the alternating direction method of multipliers (ADMM), yielding a fast, flexible, and effective solver, named Superresolution MUltiband multireSolution Hierarchical approach (SMUSH). Quantitative and qualitative results obtained on simulated and real Sentinel 2 (S2) images show the SMUSH effectiveness.

This paper presents an approach to the update of land-cover maps by classifying Remote Sensing (RS) images in an unsupervised way. The proposed method assumes that: i) an old thematic map is available; ii) no ground truth data are available; iii) the source used to generate the available thematic map is unknown. To classify the most recent RS image available on the considered area, the method automatically extracts from the considered land-cover map a “pseudo” training set. First, a preprocessing phase adapts the map to the properties of the RS data. Then, we perform an automatic “pseudo” training set identification to select the most reliable samples from the existing thematic map. Finally, a consistency check is defined to determine whether the inconsistencies between the updated and the original maps are due to real changes on the ground or classification errors. Experimental results obtained by updating the 2012 Corine Land Cover Map (CLC) in Trentino, Italy, using Sentinel 2 (S2) images confirm the effectiveness of the proposed method.

Tree crown structural parameters are key inputs to studies spanning forest fire propagation, invasive species dynamics, avian habitat provision, and so on, but these parameters consistently are difficult to measure. While airborne laser scanning (ALS) provides uniform data and a consistent nadir perspective necessary for crown segmentation, the data characteristics of terrestrial laser scanning (TLS) make such crown segmentation efforts much more challenging. We present a data fusion approach to extract crown structure from TLS, by exploiting the complementary perspective of ALS. Multiple TLS point clouds are automatically registered to a single ALS point cloud by maximizing the normalized cross correlation between the global ALS canopy height model (CHM) and each of the local TLS CHMs through parameter optimization of a planar Euclidean transform. Per-tree canopy segmentation boundaries, which are reliably obtained from ALS, can then be adapted onto the more irregular TLS data. This is repeated for each TLS scan; the combined segmentation results from each registered TLS scan and the ALS data are fused into a single per-tree point cloud, from which canopy-level structural parameters readily can be extracted.

Light Detection And Ranging (LiDAR) data have proven to be very effective in the estimation of parameters for forestry applications. However, little research has been done regarding the multitemporal analysis of these data. In this paper we propose a novel hierarchical change detection approach that first performs the detection of major changes (e.g., harvested trees) and then focuses on the detection of minor changes (e.g., single tree growth), using multitemporal LiDAR data having different point densities. Splitting the change detection problem allows us to analyze the different types of changes with different techniques. In particular, the detection of minor changes is carried out directly on the point clouds in order to exploit all the informative content of the LiDAR data. The approach has been tested on a dataset acquired in 2010 and 2014 on a complex forest area located in the Southern Italian Alps. The experimental results confirm the effectiveness of the proposed approach.

In this paper we present a growth-model based approach to the accurate estimation of stem diameter at single tree level by using high-density LiDAR data. First, we detect classes of trees characterized by different growth conditions by means of a data-driven inference process. To this end, all the environmental factors that can affect the growth of the tree (i.e., forest density and topography) are modeled and analyzed. Second, for each detected growth-model class a tailored regression function is trained to adapt the model on the considered class. The crown structure, the topography and the forest density are considered to accurately retrieve the stem diameter. Experiments carried out in mountainous scenario characterized by complex morphology and a wide range of soil fertility demonstrate the effectiveness of the proposed method.

Small-footprint high-density LiDAR data provide information on both the dominant and the subdominant layers of the forest. However, tree detection is usually carried out in the Canopy Height Model (CHM) image domain, where not all the dominant trees are distinguishable and the understory vegetation is not visible. To address these issues, we propose a novel method that integrates the analysis of the CHM with that of the point cloud space (PCS) to 1) improve the accuracy in the detection and delineation of the dominant trees and 2) identify and delineate the subdominant trees. By means of a derivative analysis of the horizontal profile of the forest, the method detects the missed crowns and delineates the crown boundaries directly in the PCS. Then, for each segmented crown, the vertical profile is analyzed to identify the presence of subcanopies and extract them. The proposed method does not require any prior knowledge on the stand properties (e.g., crown size and forest density). Experimental results obtained on two LiDAR data sets characterized by different laser point density show that the proposed method always improved the detection rate compared to other state-of-the-art techniques. It correctly detected 97% and 92% of the dominant trees measured in situ in high- and low-density LiDAR data, respectively. Moreover, it automatically identified 77% of the subdominant trees manually extracted by an expert operator in the high-density LiDAR data.

Modern forest inventory is based on the accurate and precise characterization of the 3D structure of the forest. Although LiDAR (Light Detection and Ranging) is an effective tool to estimate forest parameters, when acquired from single view point it is not able to represent accurately the entire scene. To solve this problem, in this paper we present a method that integrates the terrestrial and airborne LiDAR data. The proposed method first performs an automatic co-registration of the data sources based on the spatial pattern of the structure of the stand plot. Second, it integrates the LiDAR point clouds to accurately represent the structure of the crown. The resulting fused LiDAR point cloud can be used for an accurate estimation of the crown parameters, thus making it possible a more comprehensive representation of the 3D structure of the forest stand. Experimental results carried out in a oakland savanna in Fresno (California) confirm the effectiveness of the proposed method.

In this paper we present a hierarchical approach to the segmentation of high-density LiDAR data which aims to automatically detect and delineate the single tree crowns of both the dominant and the dominated layers of the forest. First, we detect the dominant tree crowns by using both the image derived from the LiDAR data and the LiDAR point cloud. Hence, the detected crowns are delineated directly in the LiDAR point cloud by means of a radial angular analysis. Second, the dominated crowns are detected by analyzing the vertical profile of the dominant trees. Finally, we extract the dominated trees, thus reconstructing the structure of the forest. Experiments carried out in a forest area located in the Southern Italian Alps by using very high density LiDAR data (up to 50 points/m2) point out the effectiveness of the proposed approach.

Light detection and ranging (LiDAR) technology has been extensively used for estimating forest attributes. Although high-spatial-density LiDAR data can be used to accurately derive attributes at single tree level, low-density LiDAR data are usually acquired for reducing the cost. However, a low density strongly affects the estimation accuracy due to the underestimation of the tree top and the possible loss of crowns that are not hit by any LiDAR point. In this paper, we propose a 3-D model-based approach to the estimation of the tree top height based on the fusion between low-density LiDAR data and high-resolution optical images. In the proposed approach, the integration of the two remotely sensed data sources is first exploited to accurately detect and delineate the single tree crowns. Then, the LiDAR vertical measures are associated to those crowns hit by at least one LiDAR point and used together with the radius of the crown and the tree apex location derived from the optical image for reconstructing the tree top height by a properly defined parametric model. For the remaining crowns detected only in the optical image, we reconstruct the tree top height by proposing a k-nearest neighbor trees technique that estimates the height of the missed trees as the average of the k reconstructed height values of the trees having most similar crown properties. The proposed technique has been tested on a coniferous forest located in the Italian Alps. The experimental results confirmed the effectiveness of the proposed method.

In this paper, a sensor-driven domain adaptation method is proposed for the classification of remote sensing images. The method aims at classifying an image where ground truth is not available exploiting the reference data acquired on a different but related image. This is done by taking advantage from a sensor-driven strategy that exploits the invariance of the measurements of some sensors on some classes for adaptation. This invariant property allows us to infer labels on a subset of unlabeled samples of the image that should be classified, thus introducing constrains on the adaptation process. The proposed method is based on two main steps: i) adaptation based on a sensor-driven label inference method for a subset of classes characterized by spatial invariant behaviour; and ii) adaptation based on machine learning for the remaining classes. The proposed method has been validated on 2 different datasets, where LiDAR data, hyperspectral images and high resolution optical images have been considered.

Light detection and ranging (LiDAR) is one of the most efficient remote sensing technologies for the estimation of forest parameters. However, when acquired with a low laser sampling density, LiDAR data fail in providing accurate tree height measures. In order to address this issue, in this paper we propose a novel technique for the reconstruction of tree-top height based on the joint use of low-density LiDAR data and high resolution optical images. The proposed method is based on the following steps: i) detection of all the tree crowns present in the scene by fusing the two remotely sensed data sources; ii) reconstruction of the tree-top height for those crown hit by at least one LiDAR point; iii) estimation of the tree-top height for those crowns without LiDAR points. The proposed technique has been tested on a coniferous forest located in the Italian Alps. The experimental results points out the effectiveness of the proposed method.