Change detection in highly dynamic areas, such as vegetated areas, is necessary to improve the thematic accuracy of national land use/cover map production. Sentinel-1 data have proved its potential for change detection. However, its suitability in mountainous and sparse forested areas is not well known. Here, we propose a straightforward clear-cut detection method, using monthly backscatter and coherence composites in both polarizations. The monthly clear-cut masks are estimated using a multivariate alteration detection (MAD) algorithm, using the previous and following composite, and a threshold value corresponding to the 98th percentile. The method was applied to a test site in northern Portugal, mainly characterized by the presence of dense eucalyptus forest, in a mountainous area. A Sentinel-1 time series from February to October 2018 was considered. Results showed that the monthly clear-cut estimation is not only highly influenced by rainfall events, with F1-score values less than 0.11, for the rainy season, but also by terrain-induced geometric distortions and foliage characteristics of the eucalyptus stands. The highest accuracy metrics were obtained for June, with an F1-score value of 0.45, which was still considered unfavorable. Inaccurate monthly estimations are also affected by clear-cut events that occur over more than one month, with the sum of all masks producing results with F1- scores less than 0.41. When applied to another region, located southwest of the former site and near the coastline, but characterized by smoother slopes and scattered and sparse forested stands, the method retrieved equivalent recall, but increased precision of around 0.78 (false alarms reduction), resulting in higher F1-score values (0.56).
Portugal produced a land cover map for 2018 based on Sentinel-2 data and represents 13 classes, including agriculture, six tree forest species, and shrubland. The map was updated for 2020. The strategy focused on three strata where annual changes occur: S1 (agriculture) due to crop rotation, S2 (forest and shrubland) due to wildfires and clear-cuts, and S3 (fire scars and clear-cuts of previous years) where vegetation regeneration occurs. The methodology included i) change detection, ii) classification, and iii) knowledge-based rules. Stratum S1 was classified with images of the entire 2020 crop year and a training dataset extracted from the national Land Parcel Identification Systems (LPIS) of 2020. The land cover nomenclature was expanded and class agriculture was split in three distinct classes, hence resulting a map with 15 classes in total. Change detection, implemented in stratum S2, analyzed the profile of NDVI since 2018 to find potential loss of vegetation. S2 and S3 were classified through two stages. First, images of the entire 2020 crop year were used and then data of October 2020 (end of crop year) to capture late changes. The training points of the 2018 land cover map were used, but only if not associated with NDVI change. For all the three strata, knowledge-based rules corrected misclassifications and ensured consistency between the maps. A comparison between 2018 and 2020 reveal important land cover dynamics related to vegetation loss and regeneration on ~5% of the country.
This paper presents an annual crop classification exercise considering the entire area of continental Portugal for the 2020 agricultural year. The territory was divided into landscape units, i.e. areas of similar landscape characteristics for independent training and classification. Data from the Portuguese Land Parcel Identification System (LPIS) was used for training. Thirty-one annual crops were identified for classification. Supervised classification was undertaken using Random Forest. A time-series of Sentinel-2 images was gathered and prepared. Automatic processes were applied to auxiliary datasets to improve the training data quality and lower class mislabeling. Automatic random extraction was employed to derive a large amount of sampling units for each annual crop class in each landscape unit. An LPIS dataset of controlled parcels was used for results validation. An overall accuracy of 85% is obtained for the map at national level indicating that the methodology is useful to identify and characterize most of annual crop types in Portugal. Class aggregation of the annual crop types by two types of growing season, autumn/winter and spring/summer, resulted in large improvements in the accuracy of almost all annual crops, and an overall accuracy improvement of 2%. This experiment shows that LPIS dataset can be used for training a supervised classifier based on machine learning with high-resolution remote sensing optical data, to produce a reliable crop map at national level.
The performance of supervised classification depends on the size and quality of the training data. Multiple studies have used reference datasets to extract training data automatically in an efficient way. However, automatic extraction might be inappropriate for some classes. Furthermore, classes can have distinct spectral characteristics across large areas. Thus, dividing the study area into subregions can be beneficial. This study proposes to assess the impact of the introduction of spatial stratification and manually collected training data on classification performance. Two classifications were conducted with the Random Forest classifier and multi-temporal Sentinel-2 data. The classifications’ performance was evaluated by accuracy metrics and visual inspection of the maps. The results indicate that introducing spatial stratification and manual training yielded a higher overall accuracy (66.7%) when compared to the accuracy of a benchmark classification (60.2%) conducted without stratification and with training data collected exclusively by automatic methods. Visual inspection of the maps also revealed some advantages of the novel approach, namely constraining some land cover classes to be present only within specific strata, which avoids commission errors of the class to spread freely across the map. Most of the classification improvements were observed in subregions with specific landscapes and spectral patterns, although these strata represent a small fraction of the study area, which might have contributed to the small increase in accuracy.
Experiments were carried out to investigate the use of Land Use and Coverage Area frame Survey (LUCAS) dataset and Sentinel-2 imagery to produce a land cover map in Portugal through automated supervised classification. LUCAS is a free land cover land use (LCLU) dataset based in Europe, while Sentinel-2 satellites provide also free images with short revisit frequency. The goal was to evaluate if LUCAS dataset from 2018 can be used as a single reference dataset for land cover classification at national level. The Random Forest (RF) algorithm was used. Some processing steps were undertaken to use LUCAS as reference dataset. The original LUCAS LCLU nomenclature was modified into a new nomenclature composed of 12 and 6 level-2 and level-1 map classes, respectively. Filtering was performed on LUCAS metadata, reducing the initial number of LUCAS points over Portugal from 7168 to 4910. Monthly composites of Sentinel-2 images acquired between October 2017 and September 2018 were used. To reduce the imbalance in LUCAS training points, an oversampling technique based on Synthetic Minority Over-Sampling Technique (SMOTE) was used. An independent validation dataset was produced with 600 points. RF shows an overall accuracy (OA) of 57% for level-2 and 72% for level-1 nomenclatures. When using the oversampling technique, the OA accuracy increases by 3% for level2 and 2% for level-1. The preliminary results of this experiment show that LUCAS dataset used in supervised machine learning classification has potential to produce a reliable land cover map at national scale.
In this work we study the problem of mapping soil moisture by means of Synthetic Aperture Radar (SAR) images. A test site has been set in Companhia das Lezirias, close to Lisbon, Portugal. The main advantage of using SAR images is their capability to map soil moisture at a very high spatial resolution. This opens interesting perspectives for agricultural applications, where soil moisture can abruptly change across field boundaries depending on the agricultural practices. The study area is characterized by flat topography, large agricultural areas and sparse vegetation. Five sensors have been deployed in a test area to measure soil moisture with a sampling time of one hour for a period of seven months. In-situ measurements are compared with the results obtained by processing 33 C-band Sentinel-1 images using the SAR interferometry technique. The aim of the study is to analyze the relation between the interferometric phase and time varying soil moisture. The main advantage of SAR interferometry with respect to the use of radar cross-section is that the information about soil moisture can be recovered using a reduced number of in-situ measurements. In particular, we combine three interferograms obtained from three SAR images, acquired over the same area at different times, to derive maps of bi-coherence and phase triplet. This last quantity allows to disentangle the phase contribution due to soil moisture from those related to microwave propagation in atmosphere and terrain displacements. Results are compared to those obtained using the interferometric phase and coherence to emphasize the importance to split the effects due to propagation (e.g. atmosphere) from those related to volume scattering.
In this study, an experiment aimed to integrate Global Navigation Satellite System (GNSS) atmospheric data with meteorological data into a neural network system is performed. Precipitable Water Vapor (PWV) estimates derived from GNSS are combined with surface pressure, surface temperature and relative humidity obtained continuously from ground-based meteorological stations. The work aims to develop a methodology to forecast short-term intense rainfall. Hence, all the data is sampled at one hour interval. A continuous time series of 3 years of GNSS data from one station in Lisbon, Portugal, is processed. Meteorological data from a nearby meteorological station are collected. Remote sensing
data of cloud top from SEVIRI is used, providing collocated data also on an hourly basis. A 3 year time series of hourly accumulated precipitation data are also available for evaluation of the neural network results. In previous studies, it was found that time varying PWV is correlated with rainfall, with a strong increase of PWV peaking just before intense rainfall, and with a strong decrease afterwards. However, a significant amount of false positives was found, meaning that the evolution of PWV does not contain enough information to infer future rain. In this work a multilayer fitting network is used to process the GNSS and meteorological data inputs in order to estimate the target outputs, given by the hourly
precipitation. It is found that the combination of GNSS data and meteorological variables processed by neural network improves the detection of heavy rainfall events and reduces the number of false positives.
Observing the water vapor distribution on the troposphere remains a challenge for the weather forecast. Radiosondes provide precise water vapor profiles of the troposphere, but lack geographical and temporal coverage, while satellite meteorological maps have good spatial resolution but even poorer temporal resolution. GPS has proved its capacity to measure the integrated water vapor in all weather conditions with high temporal sampling frequency. However these measurements lack a vertical water vapor discretization. Reconstruction of the slant path GPS observation to the satellite allows oblique water vapor measurements. Implementation of a 3D grid of voxels along the troposphere over an area where GPS stations are available enables the observation ray tracing. A relation between the water vapor density and the distanced traveled inside the voxels is established, defining GPS tomography. An inverse problem formulation is needed to obtain a water vapor solution. The combination of precipitable water vapor (PWV) maps obtained from MODIS satellite data with the GPS tomography is performed in this work. The MODIS PWV maps can have 1 or 5 km pixel resolution, being obtained 2 times per day in the same location at most. The inclusion of MODIS PWV maps provides an enhanced horizontal resolution for the tomographic solution and benefits the stability of the inversion problem. A 3D tomographic grid was adjusted over a regional area covering Lisbon, Portugal, where a GNSS network of 9 receivers is available. Radiosonde measurements in the area are used to evaluate the 3D water vapor tomography maps.
The electromagnetic signal transmitted by the global navigation and positioning systems (GNSS) suffers a delay which is
mainly caused by the water vapor in the atmosphere. Estimating the delay affecting the signal propagation, it is possible
to estimate the water vapor column on the troposphere above each station. The aim of this study is to characterize the
water vapor field on the troposphere over time by GNSS techniques. It is expected that can also come to assist in the
Nowcasting particularly in the prediction of severe meteorological phenomena. Several events of strong, intense and
short precipitation, observed in the Lisbon region throughout 2012 were analyzed. The choice of these events was based
on the analysis of hourly precipitation given by a meteorological station located on Lisbon center. This region is
monitored by a network of 15 GNSS stations covering about 100 square kilometers. The relationship between the GPS
precipitable water vapor (PWV) and the hourly accumulated precipitation was evaluated over time (1D closest GPSmeteorological
station plots) and spatially (2D maps) interpolated over the GNSS and meteorological stations. It was
verified that there were a high and sudden increment of the GPS PWV prior to severe precipitation events. The PWV
increment starts 6 to 10 hours before the rain and the value has increased between 57% and 75% relatively to the PWV
value observed previously. In this study is shown that GPS data has good potential for forecasting severe rain events and
high moisture flux situations.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.