Ieda Del'Arco Sanches

Self-Guided Segmentation and Classification of Multi-Temporal Landsat 8 Images for Crop Type Mapping in Southeastern Brazil

Only well-chosen segmentation parameters ensure optimum results of object-based image analysis (OBIA). Manually defining suitable parameter sets can be a time-consuming approach, not necessarily leading to optimum results; the subjectivity of the manual approach is also obvious. For this reason, in supervised segmentation as proposed by Stefanski et al. (2013) one integrates the segmentation and classification tasks. The segmentation is optimized directly with respect to the subsequent classification. In this contribution, we build on this work and developed a fully autonomous workflow for supervised object-based classification, combining image segmentation and random forest (RF) classification. Starting from a fixed set of randomly selected and manually interpreted training samples, suitable segmentation parameters are automatically identified. A sub-tropical study site located in Sao Paulo State (Brazil) was used to evaluate the proposed approach. Two multi-temporal Landsat 8 image mosaics were used as input (from August 2013 and January 2014) together with training samples from field visits and VHR (RapidEye) photo-interpretation. Using four test sites of 15 × 15 km2 with manually interpreted crops as independent validation samples, we demonstrate that the approach leads to robust classification results. On these samples (pixel wise, n ≈ 1 million) an overall accuracy (OA) of 80% could be reached while classifying five classes: sugarcane, soybean, cassava, peanut and others. We found that the overall accuracy obtained from the four test sites was only marginally lower compared to the out-of-bag OA obtained from the training samples. Amongst the five classes, sugarcane and soybean were classified best, while cassava and peanut were often misclassified due to similarity in the spatio-temporal feature space and high within-class variabilities. Interestingly, misclassified pixels were in most cases correctly identified through the RF classification margin, which is produced as a by-product to the classification map.

Hidden Markov Models for crop recognition in remote sensing image sequences

This work proposes a Hidden Markov Model (HMM) based technique to classify agricultural crops. The method uses HMM to relate the varying spectral response along the crop cycle with plant phenology, for different crop classes, and recognizes different agricultural crops by analyzing their spectral profiles over a sequence of images. The method assigns each image segment to the crop class whose corresponding HMM delivers the highest probability of emitting the observed sequence of spectral values. Experimental analysis was conducted upon a set of 12 co-registered and radiometrically corrected LANDSAT images of region in southeast Brazil, of approximately 124.100ha, acquired between 2002 and 2004. Reference data was provided by visual classification, validated through extensive field work. The HMM-based method achieved 93% average class accuracy in the identification of the correct crop, being, respectively, 10% and 26% superior to multi-date and single-date alternative approaches applied to the same data set.

TOPOGRAPHIC EFFECT ON SPECTRAL VEGETATION INDICES FROM LANDSAT TM DATA: IS TOPOGRAPHIC CORRECTION NECESSARY?

The full potentiality of spectral vegetation indices (VIs) can only be evaluated after removing topographic, atmospheric and soil background effects from radiometric data. Concerning the former effect, the topographic effect was barely investigated in the context of VIs, despite the current availability correction methods and Digital elevation Model (DEM). In this study, we performed topographic correction on Landsat 5 TM spectral bands and evaluated the topographic effect on four VIs: NDVI, RVI, EVI and SAVI. The evaluation was based on analyses of mean and standard deviation of VIs and TM band 4 (near-infrared), and on linear regression analyses between these variables and the cosine of the solar incidence angle on terrain surface (cos i). The results indicated that VIs are less sensitive to topographic effect than the uncorrected spectral band. Among VIs, NDVI and RVI were less sensitive to topographic effect than EVI and SAVI. All VIs showed to be fully independent of topographic effect only after correction. It can be concluded that the topographic correction is required for a consistent reduction of the topographic effect on the VIs from rugged terrain.

Cloud Cover Assessment for Operational Crop Monitoring Systems in Tropical Areas

The potential of optical remote sensing data to identify, map and monitor croplands is well recognized. However, clouds strongly limit the usefulness of optical imagery for these applications. This paper aims at assessing cloud cover conditions over four states in the tropical and sub-tropical Center-South region of Brazil to guide the development of an appropriate agricultural monitoring system based on Landsat-like imagery. Cloudiness was assessed during overlapping four months periods to match the typical length of crop cycles in the study area. The percentage of clear sky occurrence was computed from the 1 km resolution MODIS Cloud Mask product (MOD35) considering 14 years of data between July 2000 and June 2014. Results showed high seasonality of cloud occurrence within the crop year with strong variations across the study area. The maximum seasonality was observed for the two states in the northern part of the study area (i.e., the ones closer to the Equator line), which also presented the lowest averaged values (15%) of clear sky occurrence during the main (summer) cropping period (November to February). In these locations, optical data faces severe constraints for mapping summer crops. On the other hand, relatively favorable conditions were found in the southern part of the study region. In the South, clear sky values of around 45% were found and no significant clear sky seasonality was observed. Results underpin the challenges to implement an operational crop monitoring system based solely on optical remote sensing imagery in tropical and sub-tropical regions, in particular if short-cycle crops have to be monitored during the cloudy summer months. To cope with cloudiness issues, we recommend the use of new systems with higher repetition rates such as Sentinel-2. For local studies, Unmanned Aircraft Vehicles (UAVs) might be used to augment the observing capability. Multi-sensor approaches combining optical and microwave data can be another option. In cases where wall-to-wall maps are not mandatory, statistical sampling approaches might also be a suitable alternative for obtaining useful crop area information.

Detecção de áreas agrícolas em tempo quase real com imagens Modis

The objective of this work was to develop a method to identify and monitor, in near real-time, crop field areas cultivated with temporary summer crops, using Modis orbital images, in the state of Rio Grande do Sul, Brazil. The methodology was called near real-time detection of crop fields (DATQuaR) and uses Modis sensor images of the NDVI and EVI vegetation indices (VIs) from 16-day composites. Four different metrics were used to aggregate the values of VIs per pixel, in the bimonthly periods evaluated: average, maximum, minimum, and median. To generate the images (ImDATQuaR), the aggregated image for the previous period was subtracted from the aggregated image for the monitored period. These images were classified by slicing and compared with the reference classes obtained by the visual interpretation of randomly selected pixels in Landsat images. Each ImDATQuaR image generated two DATQuaR maps: one with a 3x3 pixel window mode filter and another without filtering. The best DATQuaR map is produced using EVI images and filtering - by subtracting the image of minimum value for the previous period from the image of maximum value for the monitored period - and achieves agreement with the reference over 81%.

Mapping croplands, cropping patterns, and crop types using MODIS time-series data

Abstract The importance of mapping regional and global cropland distribution in timely ways has been recognized, but separation of crop types and multiple cropping patterns is challenging due to their spectral similarity. This study developed a new approach to identify crop types (including soy, cotton and maize) and cropping patterns (Soy-Maize, Soy-Cotton, Soy-Pasture, Soy-Fallow, Fallow-Cotton and Single crop) in the state of Mato Grosso, Brazil. The Moderate Resolution Imaging Spectroradiometer (MODIS) normalized difference vegetation index (NDVI) time series data for 2015 and 2016 and field survey data were used in this research. The major steps of this proposed approach include: (1) reconstructing NDVI time series data by removing the cloud-contaminated pixels using the temporal interpolation algorithm, (2) identifying the best periods and developing temporal indices and phenological parameters to distinguish croplands from other land cover types, and (3) developing crop temporal indices to extract cropping patterns using NDVI time-series data and group cropping patterns into crop types. Decision tree classifier was used to map cropping patterns based on these temporal indices. Croplands from Landsat imagery in 2016, cropping pattern samples from field survey in 2016, and the planted area of crop types in 2015 were used for accuracy assessment. Overall accuracies of approximately 90%, 73% and 86%, respectively were obtained for croplands, cropping patterns, and crop types. The adjusted coefficients of determination of total crop, soy, maize, and cotton areas with corresponding statistical areas were 0.94, 0.94, 0.88 and 0.88, respectively. This research indicates that the proposed approach is promising for mapping large-scale croplands, their cropping patterns and crop types.

Crop Type Recognition Based on Hidden Markov Models of Plant Phenology

This work introduces a hidden Markov model (HMM) based technique to classify agricultural crops. The method recognizes different crops by analyzing their spectral profiles over a sequence of satellite images. Different HMMs, one for each of the considered crop classes, are used to relate the varying spectral response along the crop cycles with plant phenology. The method assigns for a given image segment the crop class whose corresponding HMM presents the highest probability of emitting the observed sequence of spectral values. Experiments were conducted upon a sequence of 12 previously classified LANDSAT images. The performance of the proposed multitemporal classification method was compared to that of a monotemporal maximum likelihood classifier, and the results indicated a remarkable superiority of the HMM-based method, which achieved an average of no less than 93% accuracy in the identification of the correct crop, for sequences of data containing a single crop class.

Identificação do modo de colheita da cana-de-açúcar em imagens multitemporais landsat-like

Brazil is the worlds leading producer of sugarcane, however, a critical point in the production of this culture is the practice of burning to facilitate manual harvest, resulting in environmental, social and economic impacts. Aiming to know and minimize these impacts, the use of remote sensing is well suited for monitoring of sugarcane, including the identification of the harvest mode. In this context, this study aimed to map and identify the crop to the sugarcane in the middle region of Triângulo Mineiro and Alto Paranaiba (TM) for a period of seven months (April-October 2015) using Landsat sensor images -like (sensors with characteristics similar to the Landsat series). The steps of this work were: 1) map the sugarcane fields; detect 2) the time of the sugarcane harvest; 3) and the type of harvest (with or without pre-harvest burning). The methodology consisted in the use of multi-temporal images of OLI/ Landast-8, ETM+ / Landsat-7 and AwiFIS / ResourceSat-2 in NIR-SWIR-Red RGB color composition. The adopted preprocessing techniques included geometric correction (AwiFS), gap filling (ET +) and a normalization procedure (Appearance Equalized ApEq). Two maps of sugarcane were generated by automatic classification using J48 and Random Forest (RF) classifiers. A field reference was obtained based on the 2013 sugarcane mapping data from INPE, which have been updated for 2015 by an experienced interpreter. This reference was used to validate the thematic maps generated. Comparing the two automatic classification methods tested, the best cane mapping result was obtained with RF. The analysis of the spectrum-temporal profile of the sugarcane fields in Landsatlike images allowed to identify with efficiency the period and mode of harvest of this culture. The results show that the months of greatest intensification of harvest were September and October, being that 75.13% of the total area cultivated with sugarcane was harvested using the mechanized system and 24.87% adopted the pre-harvest burning.

Use of MSI/Sentinel-2 and airborne LiDAR data for mapping vegetation and studying the relationships with soil attributes in the Brazilian semi-arid region

Abstract The Caatinga is an important ecosystem in the semi-arid region of northeast Brazil and a natural laboratory for the study of plant adaptation to seasonal water stress or prolonged droughts. The soil water availability for plants depends on plant root depth and soil properties. Here, we combined for the first time the remote sensing classification of Caatinga physiognomies with soil information derived from geostatistical analysis to relate vegetation distribution with physico-chemical attributes of soils. We evaluated the potential of multi-temporal data acquired by the MultiSpectral Instrument (MSI)/Sentinel-2 for Random Forest (RF) classification of seven physiognomies. In addition, we analyzed the contribution of airborne LiDAR metrics to improve classification accuracy compared to six vegetation indices (VIs) and 10 reflectance bands from the MSI instrument. Using a detailed soil survey, the spatial distribution of the vegetation physiognomies mapped by RF was associated with the variability of 20 physico-chemical attributes of 75 soil profiles submitted to principal components analysis (PCA) and ordinary kriging. The results showed gains in overall classification accuracy with use of the multi-temporal data over the mono-temporal observations. Gains in classification of arboreous Caatinga were also observed after the insertion of LiDAR metrics in the analysis, especially the percentage of vegetation cover with height greater than 5 m, the terrain elevation and the standard deviation of vegetation height. Overall, the most important metrics for classification were the VIs, especially the Enhanced Vegetation Index (EVI), Normalized Difference Infrared Index (NDII-1), Optimized Soil-Adjusted Vegetation Index (OSAVI) and the Normalized Difference Vegetation Index (NDVI). The most important MSI/Sentinel-2 bands were positioned in the red-edge spectral interval. From PCA, soil attributes responsible for most of the data variance were related to soil fertility, soil depth and rock fragments in the surface horizon. The amounts of gravels and pebbles were factors of physiognomic variability with shrub and sub-shrub Caatinga occurring preferentially over shallow and stony soils. By contrast, arboreous Caatinga occurred over soils with total profile depth greater than 1 m. Finally, areas of sub-shrub Caatinga had greater values of cation exchange capacity (CEC) and water retention at field capacity than areas of arboreous Caatinga. The differences were statistically significant at 95% confidence level, as indicated by Mann-Whitney U tests.

Spatial-temporal conditional random field based model for crop recognition in tropical regions

This work presents a spatio-temporal Conditional Random Field (CRF) based model for crop recognition from multi-temporal remote sensing image sequences. The association potential at each image site is based on the class posterior probabilities computed by a Random Forest (RF) classifier given the features at the corresponding site. A contrast-sensitive Potts model is used as a label smoothing method in the spatial domain, whereas the interactions in the temporal domain are modeled based on expert knowledge about the possible transitions between adjacent epochs. The CRF based model was tested for crop mapping in two subtropical areas based on a sequences of 9 Landsat and 14 Sentinel-1 images from Ipuã, São Paulo and Campo Verde, Mato Grosso, respectively, two municipalities in Brazil. The experiments showed significant improvements of the accumulated F1 score per class against a mono-temporal CRF approach of up to 50% and 75% for a total of 8 and 11 classes using Optical and SAR images respectively.