Selective logging represents a primary driver of forest degradation, being early germs of the deforestation process. It negatively impacts the remaining forests, leading to biodiversity losses, and catalyzing in the long term the climate changes. It is a process related to spatial-temporal changes in the forest areas, and consequently can be monitored by means of satellite images. In this work, we evaluated the potential of image time series of both Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical data to detect selective logging using several variations of the Long ShortTerm Memory (LSTM) network. Particularly, we exploited complex-valued SAR images that contain both intensity and phase information to capture small changes caused by selective logging. The experiments using optical data achieved the highest accuracy of 98.21%, while those using SAR data reached a maximum accuracy of 72.68%. The results demonstrated the effectiveness of optical images and the potential of complex-valued SAR data for selective logging detection.

Cloud and cloud shadows are a main source of concern when using dense time series of optical remote sensing images. Machine learning has the potential to effortlessly overcome this barrier using Long Short-Term Memory (LSTM), which is a deep learning algorithm created to analyze time series and has parts dedicated to suppress irrelevant information. In this context, we evaluated the ability of models with LSTM layers to create LULC maps using either cloudy or gap-filled Landsat-8/OLI time series for Pantanal. Five different LSTM models were trained with tenfold cross validation using samples gathered by the authors. Our results indicate that simple models are more accurate with filled time series, but this difference in accuracy was not present in more complex models. We also present a LULC map created for the entire Pantanal.

The Brazilian Palm Swamps (Veredas) are a vegetation physiognomy of the Cerrado biome. It has a critical importance for biodiversity and also for groundwater sources conservation. With the irrigated agriculture intensification, it's been significantly impacted. Mapping this physiognomy is important to delimit this vegetation type to provide subsides for public policy and monitoring programs. Pixel-based methods do not succeed, since the spatial context is important for this physiognomy. Object-based methods are a great potential on this sense. Deep Learning methods, particularly the convolutional neural networks (CNN), are increasing considerably as a solution for these challenges. We applied both methods in two regions of the Cerrado and evaluated the model transferability. The results are promising, with training model overall accuracies higher than 90% for both methods. The CNN performed better when transferred a different region. We discussed some advantages and limitations, and pointed out to improvements that can still be done.

By 2030, center pivots should assume the position of Brazil's main irrigation system. Along with this growth, there will be an additional water demand of 30 thousand liters/second each year, which makes sustainability increasingly important. Therefore, the objective of this work is to develop a methodology to determine the idleness and seasonality of pivots' use in the irrigated areas of Paracatu, Minas Gerais. NDVI time series were composed by the integration of Landsat-8 and Sentinel-2 images, between 2016 and 2019. Series extraction, processing and analysis were performed using the Python programming language and vector manipulation using software QGIS. Less than 2% of the total amount of pivots presented some inactivity fraction each year and the number of total pivots increased by 15.3%, evidencing the great use of irrigated areas in the municipality of Paracatu and the need for monitoring.

Automating the systematic monitoring of deforestation in the Brazilian biomes has become imperative. In this sense, a promising research field lies upon the exploitation of orbital imaging based on Synthetic Aperture Radar (SAR) sensors, since this technology is less affected by cloud cover, allowing systematic data acquisitions. In addition, the growing availability of with no charge SAR data products enables investigations on the use of time series extracted from this category of instruments, paving the way for more sophisticated temporal analyzes. This work presents the results of a SAR time series classification model designed to identify clearcut deforestation patterns in time, through an Artificial Intelligence approach known as Recurrent Neural Networks. The classification was performed using 5216 samples of Sentinel-1 time series within the Amazon basin, reaching an overall accuracy of 96.74%.

We address the task of mapping deforested areas in the Brazilian Amazon. Accurate maps are an important tool for informing effective deforestation containment policies. The main existing approaches to this task are largely manual, requiring significant effort by trained experts. To reduce this effort, we propose a fully automatic approach based on spatio-temporal deep convolutional neural networks. We introduce several domain-specific components, including approaches for: image preprocessing; handling image noise, such as clouds and shadow; and constructing the training data set. We show that our preprocessing protocol reduces the impact of noise in the training data set. Furthermore, we propose two spatio-temporal variations of the U-Net architecture, which make it possible to incorporate both spatial and temporal contexts. Using a large, real-world data set, we show that our method outperforms a traditional U-Net architecture, thus achieving approximately 95% accuracy.

The Brazilian Savannah, or Cerrado, has gained vital importance in the discussions about sustainable land development after the conversion of half of its natural vegetation. For the last two decades, most of the agricultural expansion in Brazil has occurred in this biome. This is related to technological improvements in agriculture as well as to environmental compliance policies that have effectively reduced soybean expansion in the Brazilian Amazon biome. Therefore, remotely sensed imagery, pattern recognition and image processing techniques have been employed to analyze and monitor the land dynamics over Cerrado. In this work, we present a brief review on Land Use and Land Cover mapping (LULC) in the Cerrado biome from an application perspective: natural vegetation, pastureland, agriculture, and deforestation. In this review we selected some studies whose results could contribute to the development of more detailed and accurate LULC maps for the Cerrado biome.

Classifying dense satellite image time series has become a necessity, especially with the recent efforts to create analysis ready data cubes. Approaches developed to perform this task are usually pixel-based. Even though these approaches can achieve good results, they do not take advantage of the intrinsic spatial correlation of geographic data nor do they consider spatial heterogeneity along with the time series. Region-based classification is a suitable solution to incorporate contextual information for dense satellite image time series classification. In this article, we introduce a new segmentation method based on a superpixel approach. This method creates multitemporal superpixels, which are meaningful regions in space and time. To evaluate the performance of the proposed method, tests were performed on two data sets using a total of 23 ground-truth references. Experimental results showed that the method performed well, achieving a good boundary agreement and obtaining high scores on the three metrics used for evaluation.

Mapping agriculture with high accuracy is important to generate reliable information about crop production. Pixel-based methods still present problems with noise and usually require post-processing approaches to reach satisfactory results. Object-based Image Analysis (OBIA) enable the detection of homogeneous objects in remote sensing images based on spectral similarity. However, traditional OBIA does not consider the multi-temporal characteristics of land cover or land use, such as agriculture. The objective of this study is to evaluate a phenological object-based approach with dense Landsat image time series for mapping agriculture in different level of detail in the Brazilian Cerrado. We derived pixel-wise EVI fitted time series with 8-day temporal resolution and applied multi-resolution segmentation using all image bands to incorporate the influence of space and time. Then we generated phenological metrics and applied OBIA of agricultural lands in Brazil using a hierarchical classification scheme. The overall accuracies for each hierarchical level were around 90%, and the spatial consistency of the generated maps is promising.

The Brazilian Savannah, also known as Cerrado Biome, is a hotspot for the Brazilian biodiversity and is also important for this country water supply. One of the most active Brazilian agricultural frontiers, the region has a history of primary vegetation suppression. Accurately map this phenomenon is an important step to inform and enable government conservation programs. In this work, we used a Long Short-Term Memory network to generate a deforestation map for the Cerrado. The PRODES deforestation inventory was used as ground truth during training and evaluation. We used as inputs a dense Landsat 8 time series composed by 6 spectral bands and 3 vegetation indices, as well as the SRTM terrain slope. The methodology was tested on an area comprising about 31,450 km2, achieving approximately 98.5% global accuracy.

Pasture and croplands comprise two different types of land use, which are very common in Brazil. Mapping these areas using remote sensing techniques is a challenge when using a single date image due to their similarity in spectral response. Time series might aid in discrimination of these areas once it explores the temporal behavior of surface patterns. In this work we explore time series obtained from remote sensing images to separate pasturelands from croplands in Brazilian Cerrado biome, using metrics derived from a data cube. We used Landsat 8 imagery as data source to compose a time series of six bands from OLI sensor (2 to 7) for the year of 2018. Random Forest algorithm was elected to execute the classification obtaining global accuracy of 80% and 0.58 of Kappa.

Producing reliable land use and land cover maps to support the deployment and operation of public policies is a necessity, especially when environmental management and economic development are considered. To increase the accuracy of these maps, satellite image time-series have been used, as they allow the understanding of land cover dynamics through the time. This paper presents the stmetrics, a python package that provides the extraction of state-of-the-art time-series features. These features can be used for remote sensing time-series image classification and analysis. stmetrics aims to be an easy-to-use package. The package is available under the GNU GPL software license, and the full source code is available for download at: github.com/andersonreisoares/stmetrics.

Landslides are a natural, gravity driven phenomena which can cause great economic and human losses. To prevent them, Land Use and Land Cover (LULC) maps are essential to identify areas of high susceptibility and to detect landslide scars. This paper presents results of a classification of a landslide susceptible area, using Random Forest algorithm and time series. The time series dataset is composed by the Normalized Difference Vegetation Index (NDVI) values and 16 metrics derived from the time series. The best performance was achieved using 14 metrics plus the NDVI values, with overall accuracy of 93.23% and kappa equals to 0.8937. The metrics revealed a great capability for landslides detection.

In this research, we compared two different sets of land surface phenological metrics (phenometrics) derived from dense satellite image time series to classify agricultural land in the Cerrado biome. We derived phenometrics from a dense Enhanced Vegetation Index (EVI) data cube with an 8-day temporal resolution and subjected them to classification using the Random Forest (RF) algorithm. We used a hierarchical classification with four levels, from land cover to crop rotation classes. We then evaluated the classification results comparing the use of phenometrics extracted using TIMESAT software [1], those obtained by polar representation, proposed by Körting et al. (2013) and the combination of both. We concluded that the accuracies of semi-perennial and winter crop classes increase substantially when using TIMESAT metrics combined with Polar features, and the misclassifications between single crops with non commercial crops are reduced.

Deep Learning (DL) methods are currently the state-of-the-art in Machine Learning and Pattern Recognition. In recent years, DL has been successfully applied to Remote Sensing (RS) image processing for several tasks, from pre-processing to classification. This paper presents DeepGeo, a toolbox that provides state-of-the-art DL algorithms for RS image classification and analysis. DeepGeo focuses on providing easy-to-use and extensible methods, making it easier to those RS analysts without strong programming skills. It is distributed as free and open source package and is available at https: //github.com/rvmaretto/deepgeo.

An accurate mapping of Brazilian Savanna (Cerrado) is still a difficult task due to the high spatial variability and spectral similarity between its vegetation types, called physiognomies. This work proposes a methodology based on the hierarchy of physiognomies, GEOBIA techniques with Super-pixel and a very high spatial resolution image (WorldView-2) to classify the Cerrado physiognomies in an area of preserved vegetation. Seven classes were distinguished: Gallery Forest, Wooded Savanna, Typical Savanna, Shrub Savanna, Shrub Grassland, Open Grassland and Rocky Grassland. The texture features were essential for the classification and the hierarchical approach obtained higher accuracies than the non-hierarchical approach. Moreover, GEOBIA and Superpixel were essential to represent the context that characterizes each physiognomy.

The availability of a large amount of remote sensing data made Earth Observation increasingly accessible and detailed. High temporal and spatial resolution sensors are responsible for making available data sets of time series in unprecedented proportions. Within this context, the use of efficient segmentation algorithms of remote sensing imagery represents an important role in this scenario, because they provide homogeneous regions in space-time and hence simplify the data set. In addition, the spatio-temporal segmentation can bring a new way of interpreting data by means of analyzing contiguous regions in time. This letter describes a method for image segmentation applied to time series of the Earth Observation data. We adapted the traditional region growing method to detect homogeneous regions in space and time. Study cases were conducted by considering the dynamic time warping algorithm as the homogeneity criterion to grow regions. Tests on high temporal resolution image sequences from Moderate Resolution Imaging Spectroradiometer and Landsat-8 Operational Land Imager vegetation indices and comparisons with other distance measurements provided satisfactory outcomes.

This paper aims to use unique features of hyperspectral data on an automatic process for outlining individual tree crowns (ITCs) in a tropical forest area, with special focus on semi-deciduous species. In order to enhance biophysical and biochemical properties of canopy species, a set of vegetation indices were computed. These indices served as input for a region growing segmentation algorithm that takes into account mutual similarity of pixels and spectral separability between neighbor segments. Segmentation output was evaluated on the basis of a score computed with the proportion of the area of the segments located within manually delineated ITCs. Results show that the segmentation approach is able to automatically delineate up to 70% of the control ITCs.

Although hyperspectral images acquired by on-board satellites provide information from a wide range of wavelengths in the spectrum, the obtained information is usually highly correlated. This paper proposes a novel framework to reduce the computation cost for large amounts of data based on the efficiency of the optimum-path forest (OPF) classifier and the power of metaheuristic algorithms to solve combinatorial optimizations. Simulations on two public data sets have shown that the proposed framework can indeed improve the effectiveness of the OPF and considerably reduce data storage costs.

In this paper we addressed the problem of dimensionality reduction in hyperspectral imagery classification by combining OPF classifier together with three recent evolutionary-based optimization algorithms: PSO, HS and GSA. We conducted experiments with two public datasets (Indian Pines and Salinas), which demonstrated that OPF combined with HS and GSA have obtained promising results, being the former the fastest approach. In regard to Indian Pines dataset, HS and GSA have achieved close classification rates, but HS has selected 46.25% less bands, which means a faster feature extraction step. For future works, we intend to provide a more detailed convergence analysis for PSO, HS and GSA, and also to introduce novel evolutionary-based band selection techniques and also to apply these methodologies for hyperspectral image classification in forest and agriculture applications.

Current generation of satellite imaging sensors include multispectral or even hyperspectral devices. The resulting multiple images that are acquired require new processing and analysis techniques. Image classification processing demands can be very high requiring feature/attribute selection in order to employ a minimum number of bands while keeping good classification accuracy. This work shows the use of the Rough Sets theory for multi-band image classification. This theory has a good and simple mathematical formalism and does not requires further informations such as the pertinence degree or the probability distribution in the classification process. The case study was performed with a 7-band Landsat 5 image showing the suitability of the feature selection approach and its potential to be employed in multi or hyperspectral image classification.

Image segmentation covers techniques for splitting one image into its components as homogeneous regions. This letter presents a resegmentation approach applied to urban images. Resegmentation represents the set of adjustments from a previous segmentation in which the elements are small regions with a high degree of spectral similarity (a condition known as oversegmentation). The focus of this letter is the house roofs, which are assumed to have a rectangular shape. These regions are merged according to an objective function, which, in the technique presented here, maximizes the rectangularity. With oversegmentation, we create a graph known as a region adjacency graph (RAG) that relates border elements. The main contribution of this letter is a technique, which works with the RAG, to maximize the objective function in a relaxationlike approach that splits and merges oversegmented regions until they form a meaningful object. The results showed that the method was able to detect rectangles according to user-defined parameters, such as the maximum level of the graph depth and the minimum degree of rectangularity for objects of interest.

This work studies the impact of restoration process on radiometry derived from remote sensing images. For this purpose, a set of six Landsat TM bands was selected for the evaluation procedure. Experimental results indicated that the spectral characterization and surface reflectance values were not compromised by the restoration process.

Often images are slightly blurred due to the blurring effect of sensors (optical diffractions, detectors size, eletronic filters). As a consequence, the effective resolution is, in general, worse than the nominal resolution that corresponds to the detector projection on the ground that does not take into account sensor imperfections. Thus restoration techniques aims at deblurring the image and this way improving its spatial resolution up to a certain level. In this paper different restoration methods are applied and evaluated to restore CBERS-2B images: Wiener filter, Richardson-Lucy, Modified Inverse Filter and a Row Action Projection filter. In the experiments images of HR CBERS-2B covering inland water regions are restored. Results showed that Richardson-Lucy approach outperformed the others.

Bayesian Network (BN) techniques can be used to represent the causal relationships among random variables on probabilistic models. Only few studies have applied these techniques to remote sensing and other spatial data integrated in geographic information systems. The objective of the present work was to map soybean plantation using minimum of EVI (M), range of EVI (R) and terrain slope (L) as input variables in the BN. Soybean plantations were evaluated in the state of Rio Grande do Sul, Brazil during the 2000/01 crop year. The probability function was discretized with five different numbers of intervals. Results were improved with the increase of the number of intervals. Best soybean mapping result presented sensitivity, specificity and overall accuracy indices equal to 77.62, 77.56 and 77.58%, respectively, indicating that the method is promising and has potential to be improved with the use of additional input variables.

The first China-Brazil Earth Resources Satellite (CBERS-J) launched in 1999 has been developed by China and Brazil. It carries on-board a multisensor payload with different spatial resolutions called: wide field imager (WFI), high resolution CCD camera (CCD) and infrared multispectral scanner (IR-MSS). The performance of these sensors can be evaluated through the point spread function (PSF) that enables an objective assessment of the spatial resolution. This work describes an approach to estimate on-orbit CBERS-I CCD spatial resolution using an image of a black square target on the Gobi desert (China), The results show that the spatial resolution in across-track direction is not complied to the design specification for all bands whereas the spatial resolution in along-track direction is conform to the specification for all bands, except the band 4.

Describes a system for automatic and semi-automatic registration/mosaic of remote sensing images. Information provided by the user can be used to speed up processing or to avoid mismatched control points. A statistical procedure is used to characterize good and bad registrations. Based on this "good fit-bad fit" statistical test the user can stop, modify the parameters, or continue the processing. Several tests have been performed by registering optical, radar, multi-sensor, high-resolution images and video sequences. We have included very difficult image registration examples in order to show the strengths and limits of our system. An online registration system demo containing several examples can be executed using a Web browser.

Registering radar imagery is not an easy task since the images may be heavily contaminated with noise. Strong speckle noise can produce artifacts that mimic good control points and may produce low precision or even wrong registration. This article presents an automatic registration method that tries to overcome these problems. The method was originally developed for optical images and adapted for radar images.