Agriculture - National 3

Crop area extent estimates and crop type maps provide crucial
information for agricultural monitoring and management. Remote sensing
imagery in general and, in particular, high temporal and high spatial
resolution data as the ones which will be available with recent
systems such as Sentinel-1 [1] and Sentinel-2 [2] constitute a major
asset for this kind of application.

Recent works have shown that multi-temporal optical imagery as the one
that will be provided by the Sentinel-2 satellites is able to provide
accurate crop type mapping over different climates and crop systems
[3]. However, since optical imagery is affected by cloud cover, the
performances of the crop type mapping system can be hindered in some
cases, even with a 5-day revisit cycle.

Added to annual mapping as the one presented in
[3], early crop type detection before the end of the season is needed
for yield forecasting and irrigation management. In this case, image
availability is still more crucial.

Optical data is usually preferred to SAR imagery because of the better
understanding of the link between the observations and vegetation
phenology. However, the all-weather acquisitions provided by SAR and
recent advances in the understanding of the underlying physical
phenomena, make multi-temporal SAR imagery an interesting candidate for
crop type mapping.

SAR imagery for crop type mapping has also been used together with
optical data. For instance, McNairn et al. [4] used 2 SAR
(Envisat-ASAR) images and one optical (SPOT4) to achieve acceptable
accuracies. Zhu et al. [5] proposed a Bayesian formulation for the
fusion of Landsat TM and ERS SAR (2 images of each type). Their
approach was based on building a specific statistical link between the
2 types of data for a particular set of acquisition dates, which is
not general enough to be implemented in operational settings. However,
their results show the complementarity of the 2 types of data. The 2
works cited above used a small number of images. Sentinel-1 and
Sentinel-2, with their short revisit cycle will offer improved
possibilities.

Some works in the literature have already explored the use long time
series, but using a dense time series of one of the sensors (either
optical or SAR) and a few images coming from the other modality.

For instance, Blaes et al. [6] use 15 SAR images (ERS and Radarsat)
and 3 optical ones (Landsat TM). They perform a field level
classification together with photo-interpretation schemes. The show an
improvement of the accuracy thanks to the use of SAR data, with
respect to the accuracy achieved with 3 optical images. As most
approaches in the literature, a date selection is used, which is
incompatible with fully automatic operational systems.

Unlike the literature cited above, in this work we assess the joint
use of dense time series of both SAR and optical imagery in order to
devise a strategy for the operational exploitation of both Sentinel-1
and Sentinel-2 data in the frame of crop type mapping at early stages
of the agricultural season.

The work presented here builds upon the results reported in
[3], where the optical image exploitation workflow was assessed.
Therefore, the focus of the present work is on the integration of
multi-temporal SAR data into the existing processing chain. In order
to do that, the SAR image processing and feature extraction was
investigated. In particular, we studied the impacts of image
pre-processing (geometric resolution, speckle filtering), the
selection of polarimetric channels, the contribution of several image
features (local statistics, texture descriptors) and the influence of
post-processing steps like classification map regularisation.

The data set used for the study was composed by 11 Landsat-8 and 9
Sentinel-1 images acquired between the September 2014 and June 2015 over
an agricultural site in the South-West of France. A reference data set
obtained by several field surveys was used for validation and quality
assessment and contained 8 land cover classes (bare soil, grass,
alfalfa, rapeseed, soybean, sunflower, corn and wheat).

The results show that the integration of Sentinel-1 time series into the
existing system based on Sentinel-2-like optical image time series
improves the performances in terms of class recognition. This
improvement has the most impact in the early stages of the crop season
(December to January in our test site) and after spring. Speckle
filtering and post-classification regularisation have similar impacts.
Haralick texture descriptors and local statistics are the most pertinent
features for the exploitation of Sentinel-1 data for crop
identification.

[1] Ramon Torres, Paul Snoeij, Dirk Geudtner,  David Bibby, Malcolm Davidson, Evert Attema, Pierre Potin, Björn Rommen, Nicolas  Floury, Mike Brown, Ignacio Navas Traver, Patrick Deghaye, Berthyl Duesmann, Betlem  Rosich, Nuno Miranda, Claudio Bruno,  Michelangelo L'Abbate, Renato Croci, Andrea  Pietropaolo, Markus Huchler and Friedhelm Rostan, "GMES Sentinel-1 Mission", Remote Sensing of Environment, 120:9-24 (2012)

[2] M. Drusch, U. Del Bello, S. Carlier, O.  Colin, V. Fernandez, F. Gascon, B. Hoersch, C. Isola, P. Laberinti, P. Martimort,  A. Meygret, F. Spoto, O. Sy, F. Marchese  and P. Bargellini, "Sentinel-2: ESA's Optical High-Resolution Mission  for Gmes Operational Services", Remote Sensing of Environment, 120:25-36 (2012)

[3] Jordi Inglada, Marcela Arias, Benjamin Tardy, Olivier Hagolle, Silvia Valero, David  Morin, Gérard Dedieu, Guadalupe Sepulcre, Sophie Bontemps, Pierre Defourny and Benjamin  Koetz, "Assessment of an Operational System for Crop Type  Map Production Using High Temporal and Spatial  Resolution Satellite Optical Imagery", Remote Sensing, 7:12356-12379 (2015)

[4] Heather McNairn, Catherine Champagne, Jiali  Shang, Delmar Holmstrom and Gordon Reichert, "Integration of Optical and Synthetic Aperture Radar  (SAR) Imagery for Delivering Operational Annual Crop  Inventories", ISPRS Journal of Photogrammetry and Remote Sensing, 64:434-449 (2009)

[5] L. Zhu and R. Tateishi, "Fusion of Multisensor Multitemporal Satellite Data  for Land Cover Mapping", International Journal of Remote Sensing, 27:903-918 (2006)

[6] X. Blaes, L. Vanhalle and P. Defourny, "Efficiency of Crop Identification Based on Optical  and SAR Image Time Series", Remote Sensing of Environment, 96:352-365 (2005)

Object-based classification approaches have several advantages compared to pixel-based approaches. For example, the negative impact of noise can be minimized while at the same time textural features, form and context information can be used for improving classification results.

One difficulty of using object-based image analysis (OBIA) relates to the fine-tuning of segmentation parameters as only well-chosen segmentation parameters ensure optimum segmentation results. Manually defining suitable parameter sets, however, can be a time-consuming approach, not necessarily leading to optimum results; the subjectivity of the manual approach is also obvious. A visually appealing segmentation also does not necessarily lead to a good (subsequent) classification. For this reason, in supervised segmentation 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 idea formulated by Stefanski et al. (2013) 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. The automatisation is achieved by using the overall classification accuracy of the subsequent image classification as a cost function to be minimized.

A sub-tropical study site located in São Paulo State (Brazil) was used to evaluate the proposed approach. Two Landsat-8 image mosaics were used as multi-temporal input (from August 2013 and January 2014) together with training samples from field visits and VHR (RapidEye) photo-interpretation. For validation, four test sites of 15 x 15 km² were prepared with manually interpreted crops as independent validation samples. 

Using this dataset, we demonstrate that the approach leads to robust classification results. On the independent validation samples (~1 million pixel) an overall accuracy (OA) of 80 % could be reached while classifying five classes: sugarcane, soybean, cassava, peanut and others. Amongst the five classes, sugarcane and soybean were classified best, while cassava and peanut were more 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. The produced land cover map not only well represents the land cover distribution in the study area, but the automatically identified object boundaries generally correspond well with real field boundaries. 

One drawback of the approach is that the processing is quite long. For running 1000 segmentations and subsequent classifications (incl. feature extraction etc.) almost seven days on a regular PC were needed. To reduce the processing time, we therefore also investigated if it is possible to optimize the segmentation directly using image-derived statistics such as Moran’s I and explained variance. First pre-liminary test reveal the potential of this approach.

In past works, it has been demonstrated over various agro-ecological zones, crop establishment methods and water management practices in South-East Asia that spaceborne Synthetic Aperture Radar (SAR) time-series combined with yield crop modeling offers an effective alternative to conventional terrestrial methods and to yield estimations modeled at point level and up-scaled at some arbitrary administrative levels. Although seasonal rice information could have been generated on an operational basis providing accurate estimates, it has been recognized, due to the unfavorable SAR data availability, that the proposed SAR-yield model based solution had significant restrictions in terms of spatial coverage (sub-national level) and monitoring capabilities (i.e. limited to one crop season). Moreover, given the nature of the significant spatial-temporal seasonal rice dynamic, the relatively long revisiting time between subsequent acquisitions has been identified – for the accurate detection of key rice growth stages and inference of biophysical parameters – as an additional drawback. Nowadays, the availability over Europe of systematic Sentinel-1A dual polarization 12-days acquisitions in ascending and descending mode opens new frontiers in the processing, analysis and use of hyper-temporal SAR intensity and coherence stacks.

The purpose of this paper is multiple, namely:

To demonstrate the robustness of the multi-temporal s^o rule based rice detection algorithm, developed for X-band (Cosmo-SkyMed, TerraSAR-X) and extended to C-band (Rasarsat-2, RISAT-1 and Sentinel-1) data. It is worth mentioning that this algorithm is applicable only if the rice peak of season is reached.

To detect the cultivated rice area at earliest stage by means of temporal descriptors derived from SAR s^o time series. It is additionally shown, that in complex agricultural regions, the integration of temporal descriptors extracted from multi-temporal Landsat-8 and in future Sentinel-2 data (hence leading to a temporal-spectral descriptors approach) is doubtless of advantage. Furthermore, given the stable Sentinel-1A baseline, temporal descriptors from multi-temporal coherence can additionally obtained and used to characterize the fields status and their evolution during the non-vegetative phase.

To monitor the initial irrigated area, date and duration by combining ascending and descending Sentinel-1, ascending Radarsat-2 and Landsat-8 time series.

To assess the main seasonal rice phenological moments, i.e. Start of Season, Start of the Vegetative Phase, and Start of Senescence. 

To infer the Leaf Area Index during the vegetative phase. It is worth mentioning that the retrieval of this bio-physical parameter is performed by combining the detected rice phenological stages with a water cloud model, hence enabling to adapt it according to the different frequencies (i.e. X- and C-band) and phenology.    

To assess, during the different phenological moments, the fields spatial and temporal variability by integrating very high resolution SAR data (Cosmo-SkyMed and TerraSAR-X StripMap).   

Based on multi-temporal Sentinel-1A, Landsat-8, Cosmo-SkyMed and TerraSAR-X data, results over the three major European rice regions located in Italy, Spain and Greece are shown and evaluated. The scalability of the proposed methodology is already on-going in West-Africa and South-East Asia, where adequate Sentinel-1A – and possibly Sentinel-2A – time series are available.  

Acknowledgements

This work received funding from the ERMES FP7 project (Contract N°: 606983. Starting Date: 01/03/2014. Duration: 36 months). Radarsat and Cosmo-Skymed data were acquired in context of the SOAR/ASI project 2907/5242 and 2908-5233.

Timely information about agricultural crops (e.g. typology, phenology, productivity, health) is crucial for proper agronomic planning and management by farmers and public administrations. Earth Observation (EO) data have proved to be very effective for monitoring crops and their intra-annual cycles, especially through the integration of optical and microwave data. SAR data at different frequencies have been used for retrieving biophysical crop parameters pertinent to the crops, e.g. for the estimation of biomass and growth status.

Here we present an experimental analysis based on the investigation of X-band backscattering derived from multitemporal TerraSAR-X data over three different types of summer crops (rice, maize, soybean). The study area is located in south-eastern portion of Lombardy region, Northern Italy, framed within the Po river Plain and the Ticino river basin, and covers two farms located in Rosasco municipality (45°15’00” N, 8°35’00” E). In situ campaigns have been conducted in 2014 and 2015 along the summer crop season (May-September) on this area in order to measure biophysical parameters related to: i) agronomy (crop type and variety, agro-practices, seeding date, seeding scheme and density), ii) substrate (soil roughness, soil moisture, flooding conditions), iii) crop phenology (BBCH scale stage, context and detail photos), iv) morphometry (plant height, number of leaves, leaves size), and v) biomass and density (biomass, plant water content, LAI). Field observations and measurements of crop parameters were carried out over 20 fields (9 for rice, 9 for maize, 2 for soybean), on 9 dates. Contextually, the acquisition of TerraSAR-X (TSX) Stripmap dual-pol HH/HV scenes was planned and carried out in the context of TSX scientific proposals LAN2412 and LAN2984. TSX images acquired for both HH and HV polarizations were radiometrically calibrated, including multilooking (7R x 4A) and terrain correction, to obtain sigma nought (σ°) maps in both polarizations.

Sensitivity of σ°HH and HV to measured biophysical parameters was assessed using experimental data collected on test fields. Analysis of preliminary results led to some interesting remarks. Specifically, we observed a significant sensitivity of σ°HH to fresh and dry biomass, with different behaviour for the different crops. Rice σ°HH shows a sensible decrement over the whole range of biomass values, while a different trend is observed for maize, with a first steep increment of σ°HH during the initial biomass growth phase, followed by a flat trend for values above 2.0 kg m-2. A reduced sensitivity to biomass changes has been observed for HV polarization case. According to these preliminary analyses, backscatter over soybean fields does not show a significant relationship with the measured parameters.

A radiative transfer model (Paloscia et al., 2014) was then used for simulating multi-temporal backscatter of rice and maize fields and model sensitivity analysis was performed in order to evaluate the contribution of different components of the vegetation-soil system to total backscattering. Eventually the model performance were assessed in comparison to measured TSX backscatter for maize and rice fields in non-drying conditions (plant water content >60%) scoring satisfying results with coefficient of determination (R²) > 0.5 and RMSE~1-3 dB.

Coupling electromagnetic scattering modelling with experimental data can provide an effective tool for predicting the temporal radar response during the growing cycle of crop, depending on crop characteristics (e.g. morphology, water content, density), health status and phenological stage.

Maize Lethal Necrosis (MLN) is a serious disease in maize that significantly reduces yields up to 90% in some areas in Kenya. The disease causes chlorotic mottling of leaves and severe stunting which ultimately leads to plant death. The spread of MLN in the maize growing regions of eastern Africa has intensified since the first outbreak was reported in September 2011 in Kenya. In this study, 30-meter Landsat-8 and 5-meter multi-temporal RapidEye imagery was combined with field-based assessments on MLN infection rates to map three MLN severity levels in Bomet County, Kenya. Two RapidEye level 3A images, acquired during maize planting season and the milk stage respectively, were combined with one cloud free Landsat-8 image (path 169, row 061) acquired during maximum phenology stage. The Landsat was re-sampled to 5 meter pixel resolution using nearest neighbor re-sampling. Thirty spectral indices for each RapidEye time step were computed and included in the mapping model. Machine learning using random forest classification, was used on the fused satellite data sets to create a map separating maize field from all other land cover and land use classes. Subsequently, MLN severity levels (mild, moderate and high) were mapped using the in situ data set from the field as training data in random forest. Integrating RapidEye and Landsat-8 data improved the classification accuracy for the cropped versus non-cropped area map. The random forest algorithm yielded an overall accuracy of 95%, representing high model performance, in mapping the three MLN severity levels. These results indicated the possibility of using time-series of multi-sensor satellite data (with permissible spatial resolution) and machine learning to monitor the spatial distribution of disease infestation rates in small scale and fragmented agro-ecological landscapes.