Agriculture - National 2

The high demand for pastoral land is the main driver of deforestation in the Brazilian Amazon today and causes substantial losses of a unique biodiversity and severe emissions of greenhouse gases (GHG). Sustainable intensification of existing livestock production systems is seen as a potential way to reduce the demand for new land and thereby the Brazilian GHG emissions. Intensification is mainly achieved by shifting from traditional and extensive, mostly fire based management practices towards advanced tillage systems and integrated crop- and grazing rotations. Deeper knowledge on pasture land use intensity is hence needed to evaluate how land management policies, such as REDD+ (Reducing Emissions from Deforestation and Forest Degradation) or private sector driven zero-deforestation agreements, influence pasture land use and related GHG emissions.

Remote sensing based studies on spatio-temporal patterns of agricultural intensification in the Amazon are rare and mostly focus on expansion or dynamics of cropland agriculture – not the least because describing pasture dynamics based on remote sensing data is challenging. Timing of grazing, recuperation periods and management practices are largely driven by individual decisions. Existing state-of-the-art approaches in ecosystem disturbance detection are hence less suited to discriminate the short-termed processes in pasture environments.

To better address this knowledge gap, we developed an approach to map fire and tillage events on pasture areas. Our study area is the region of Novo Progresso, Pará, Brazil, a hot-spot of deforestation where most of the deforested land is converted to grazing land. We use all Landsat observations from the USGS Landsat data archive with standard terrain correction (Level 1T) that have been recorded from the study area (footprint 227/065) between July 2012 and September 2015. All scenes are transformed to surface reflectance and cloud masked with Fmask. The dense time series and a comprehensive set of field records from 2014 and 2015 are used to label land cover changes in exemplary reference areas. Our algorithm selects at the pixel-level the multi-temporal sequences of unclouded Landsat observations around a management event, such as ploughing or burning. . It then extracts features like spectral indices, spatial statistics, and temporal metrics to account for e.g. seasonal variations. The final clear observation feature stack then describes the temporal feature trajectory before and after a management event.

Based on our reference data we train a global random forest classifier to distinguish forests, pastures, burned pastures and tilled pastures. First results showed accuracies of around 86% for fire and 67% for tillage events according to our reference data. The maps of burned and tilled pastures highlight the management differences between small and large-holder communities in the study region. Based on these findings, we conclude on future opportunities related to the synergies of Landsat and upcoming Sentinel-2 data to characterize land-use processes in the Amazon.

The Sahelian pastoral zone is completely dependent on its unimodal rainy season for the creation of biomass. Transhumant pastoralism, which is one of the key economic activities of the region, is particularly vulnerable to volatile rainfall, due to a reliance on rain-fed pasture and water points. Hence, knowledge about a delay in the start of rainy season is essential to counteract food insecurity and to implement mitigation strategies.

Late starts in the rainy season can often indicate shorter growing periods and low biomass production or water availability. ESA’s Global Monitoring of Food Security (GMFS) service element has the intention to support the ESA TIGER initiative with Earth Observation based information products for agriculture monitoring and food security. As one of the project users Action Contre la Faim (ACF) has been using soil water index (SWI) estimates derived from the Advanced SCATterometer (ASCAT) as well as SWI-based Start of Season (SoS) information to quickly identify zones of delayed rainfall and abnormal aridity in West Africa. These datasets form a key input to the annual analyses conducted by ACF during the rainy season to provide an outlook on the dry period and identify areas with poor biomass production.

The SWI-based SoS is understood as the point in time, when the onset of the wet season can be seen in the SWI time series. Hence, it is indicating suitable conditions for agricultural activities like soil preparation and sowing, but it is also an excellent indicator for pastoral activities. As a result, SWI and SoS data have proven to be essential in analyzing pastoral conditions across the Sahelian zone as well as potential movements of herds. In fact, by observing the relationship between delayed season start and sub-par biomass production, SoS and SWI datasets are crucial to determine pastoral movements and to predict potential areas of crisis. First results reveal that pastoral zones classified as delayed starts by SoS data or abnormal SWI aridity are often subject to an abnormally early pastoral exodus. Although our preliminary study is only based on a few datasets in Western Africa, we are confident that there exists a strong connection between a delayed SoS, sub-par biomass and deviations in transhumance routes. Our preliminary results are promising, but further in-depth research is needed to confirm the reliability of such an approach. Ongoing research should also steer efforts towards the development of a Sentinel-1 based processing chain for agricultural monitoring that will allow the application on the local level and could potentially be integrated into existing early warning remote sensing frameworks within Africa.

Global agricultural expansion has slowed while crop production continues to keep pace with food, fuel, and fiber demands. With less prime land available for expansion, farmers have transformed existing farmland in some areas by opting for intensification over expansion. The benefits of technological advancements are obvious, with improved crop yields and more sustainable soil management among the key variables. However, large-scale, intensive farming can also have negative impacts (e.g., water quality, biodiversity). Research in recent years has drawn attention to agricultural expansion in tropical forests, and rightly so. However, many important details are lost when scrutinizing only agricultural expansion into non-agricultural vegetation. That is, traditional monitoring approaches do not consider structural changes within land cover regions. We present methodology to identify individual cropland field parcels from temporal Landsat imagery across South America. Our goals are to improve estimates of cropland area and simultaneously construct a dataset that best depicts the scale of farming across the region. Remote sensing offers a twofold advantage in this endeavor by: providing higher resolution capabilities over administratively collected data, and avoiding sample biases that transcend national borders. Our approach consists of (1) temporal Landsat, pixel-based composites, (2) spectral transformations, (3) multi-scale and -spectral image edge extraction, (4) multi-scale contrast limited adaptive histogram equalization (CLAHE), (5) multi-scale adaptive thresholding, and (6) image morphological ‘cleaning’ to extract individual land cover parcels from multi-temporal Landsat imagery. The CLAHE normalization is a localized adjustment that avoids the use of global values. Similarly, the adaptive threshold step is a localized method to produce binary edge estimates, which avoids over- and under-segmentation based on a global value. The object detection is fully automated, but the methodology is semi-automatic because we estimate a Landsat-scale cropland land cover map separately. Objects are considered cropland objects if the cropland pixel count exceeds the majority of the object area. We tested the procedure across an area of South America that contains roughly 82% of all cropland on the continent. The validation was conducted to:  (1) assess the cropland land cover map, using a spatially constrained approach in order to identify regional areas of weakness; and (2) assess the land cover parcel extraction using object based metrics. We validated the algorithm against approximately 5,500 manually delineated cropland field parcels.            

Our approach offers a new technique of analyzing agricultural changes across a broad geographic scale. The assessment reveals the cropland estimates are an improvement over large-scale cropland estimates for the region, and the low standard error of the per-parcel estimates highlight the applicability of the methods over a large region. By using multi-temporal Landsat imagery with a semi-automatic field extraction approach, we can monitor within-agricultural changes at a high degree of accuracy and advance our understanding of regional agricultural expansion and intensification dynamics across South America.

Policies and institutional changes during the Soviet and post-soviet period have induced widespread land cover and land use changes in the former Soviet Union, particularly including substantial conversions between cropland and grassland. We estimate spatial and temporal patterns of land use change in steppe region of Northern Kazakhstan from 1984 to 2015. However, separating croplands and grasslands is challenging due to spectral similarities of cropland and grassland, the subtle land cover and use changes within and between years, and the extreme climatic conditions with recurring droughts that lead to large fluctuations of biomass growth and frequent occurrence of large fire events. Our analysis focuses on cropland abandonment and recultivation within three Landsat footprints in Kostanay region for which we use all available Landsat TM, ETM+, and OLI scenes to create pixel based composites for three representative days of the year. These days were chosen based on wheat phenology, the predominant crop in the study area, to optimally separate croplands from grasslands. The first selected day represents the time when the croplands have already been plowed, but not yet vegetated, which is between end of May to early June. The second day represents the time when the crops show the maximum photosynthetic activity, captured with the Normalized Difference Vegetation Index (NDVI), while natural vegetation starts to become photosynthetically inactive due to water stress. This time approaches around mid-July. The third day is between middle to end of October, when wheat has already been harvested. Yearly training data was collected from randomly chosen points. Fields were considered abandoned when their phenology corresponds to cropland in one year and to grassland in at least two consequent years. Accordingly, the fields plowed after at least two years of abandonment were considered recultivated. The dense Landsat time series were classified using the Random Forests algorithm to map the abandonment and recultivation in the study area. Results were validated with available reference data as well as reference data collected from visual image interpretation and during fieldwork. The resulting land-use change map reveals the dynamic changes from cropland to grassland that occurred especially in the years following the breakdown of the Soviet Union as well as the rebound of cultivation since about 2000. Our approach demonstrates the ability of high-density Landsat time series to recognize with high precision the cropland dynamics in this semi-arid steppe region.

The commissioning of Sentinel-1 allows for a novel approach on land cover monitoring through the large-scale availability of remote sensing acquisitions with consistent temporal resolution. In combination with high-resolution Radarsat-2 as well as Landsat-8 data a land cover change monitoring technique is constructed, based on a specifically designed Markov chain model. The classifier makes distinction between sugarcane, annual crops, forestry, pasture, urban and water. The model was applied on São Paulo state, Brazil, which hosts the largest sugarcane acreage in the world and the produced bio-ethanol provides for about half of the energy consumed by the automobile transportation sector. The resulting regional land cover maps over time are useful for research groups studying socio-economic and environmental impacts with expanding bio-energy crops.

In this study, special attention is given to how the radar data can complement optical data in the classification of land covers and what its effect is on classification accuracies. More than one thousand fields in the São Paulo region have been inspected, spread over two growing seasons in 2015, which were used to train and validate the classifier. Radar profiles of the annual crops, sugarcane, pastures and forests show a clear potential for enhanced classification with respect to only optical-based classification which suffers from cloud coverage-induced temporal gaps. The land cover maps of 2005 and 2015 were created, whereby the 2005 map is based on an optical-based classifier and the 2015 map based on a radar and optical integration-based classifier. The radar data consists of Sentinel-1 Interferometric Wide and Extra Wide mode scenes (both HH+HV) and 20 Radarsat-2 Fine Quad and 16 Standard Dual Pol (HH+HV) scenes provided by ESA grants.

For the application on the whole São Paulo state, roughly the size of the United Kingdom, a Wide Area Processor (WAP) was developed for the acquiring, pre-processing and stacking of the radar and optical remote sensing data over large regions. A schematic overview of the big data collecting and processing algorithms will be given along with our practices of feeding the data to the classifier in an effective manner.