Extracting smallholder cropped area in Tigray, Ethiopia with wall-to-wall sub-meter WorldView and moderate resolution Landsat 8 imagery

Highlights

• First very high resolution wall-to-wall smallholder crop area for Tigray, Ethiopia

• Semi-automated segmentation of over 40 TB of WorldView-1 and WorldView-2 images

• Timely big data processing on NASA Advanced Data Analytics Platform (ADAPT)

• Results show 46% of area is smallholder croplands where field sizes are ≤ 1 ha.

Abstract

Very high resolution (VHR) satellite data is experiencing rapid annual growth, producing petabytes of remotely sensed data per year. The WorldView constellation, operated by DigitalGlobe, images over 1.2 billion km² annually at < 2 m spatial resolution. Due to computation, data cost, and methodological concerns, VHR satellite data has mainly been used to produce needed geospatial information for site-specific phenomena. This project produced a VHR spatiotemporally explicit wall-to-wall cropland area map for the rainfed residential cropland mosaic of the Tigray Region, Ethiopia, which is comprised mostly of smallholder farms. Moderate resolution satellite data do not have adequate spatial resolution to capture the total area occupied by smallholder farms, i.e., farms with agricultural fields of ≤ 45 × 45 m in dimension. In order to accurately map smallholder cropped area over a large region, hundreds of VHR images spanning two or more years are needed. Sub-meter WorldView-1 and WorldView-2 segmentation results were combined with median phenology amplitude from Landsat 8 data to map cropped area. Over 2700 VHR WorldView-1, -2 data were obtained from the U.S. National Geospatial-Intelligence Agency (NGA) via the NextView license agreement and were processed from raw imagery to produce a smallholder crop map in ~ 1 week using a semi-automated method with the large computing capacity of the Advanced Data Analytics Platform. We estimated cropped area in Tigray to be 46% with a commission error of 5% ± 10% and omission error of 15% ± 12%. This methodology is extensible to other regions with similar vegetation texture and can easily be expanded to run on much larger regions.

Introduction

Satellite data is a significant source of big data currently undergoing explosive growth (Ma et al., 2015), with geospatial data growing by as much as 20% annually (Lee and Kang, 2015). The constellation of very high resolution (VHR) 0.31–2 m sensors on satellites operated by DigitalGlobe – with cameras on WorldView-1 (WV1), GeoEye-1, WorldView-2 (WV2), WorldView-3 (WV3), and WorldView-4 (WV4) – between them can revisit the same location on the globe in the same day, resulting in ~ 1.233,700,000 km² of imagery annually (Digital Globe, 2016). Individual WV1 images are ~ 37,000 × 37,000 pixels, compared to ~ 8000 × 8000 pixels for Landsat 8 Operational Land Imager (OLI) images. The U.S. National Aeronautical and Space Administration's (NASA's) Goddard Space Flight Center (GSFC) currently stores and provides access to over 2.0 PB of VHR data to NASA-funded investigators via an online portal called Commercial Archive Data for NASA investigators (CAD4NASA - http://cad4nasa.gsfc.nasa.gov/). This is primarily a subset of the U.S. National Geospatial-Intelligence Agency (NGA) Commercial Archive Data holdings. These data are also available to U.S. government-funded investigators as well as non-governmental organizations (NGOs), state and local governments, intergovernmental agencies, universities, and even foreign governments when the use is in support of U.S. interests from online portals from the United States Geological Survey and NGA. CAD4NASA includes data from GeoEye-1, IKONOS-2, Quickbird-2, WorldView-1, -2, -3 (Neigh et al., 2013). With the volume of VHR big data growing at ~ 2.0 PB annually with the launch of WorldView-4 (Babcock, 2013), there is an opportunity to use these data for large regional-scale analyses. Applying object-based image segmentation and other machine learning methodologies to these data can resolve what has been overlooked and/or missed with moderate resolution Earth observation data. For example, the 30 m U.S. Department of Agriculture Cropland Data Layer (Boryan et al., 2011) classifies all cropped field in the contiguous U.S., from large to small, at an accuracy necessary for yield monitoring and commodity pricing. However, 30 m cropland mapping has not produced overall accuracies required for food security applications (Gong et al., 2013, Yu et al., 2013, Lobell et al., 2015), presenting a clear need for VHR cropland mapping in parts of the world that have smallholder subsistence agriculture (Fritz et al., 2015, See et al., 2015, Husak and Grace, 2016).

VHR satellite data has been used to produce needed geospatial information on site-specific phenomena and event-specific needs. Commonly, VHR data has been used to create validation/verification points on a scene by scene basis to produce uncertainty assessments for moderate and coarse level products (Chen et al., 2015, Waldner et al., 2015, GOFC-GOLD, 2015). Localized agricultural applications of VHR data have included estimating crop yields (Yang et al., 2013, Yang et al., 2009), monitoring agricultural field conditions (Muller and van Niekerk, 2016), identifying agricultural fields in a heterogeneous landscape (Debats et al., 2016), classifying crop types (Chellasamy et al., 2015), identifying tillage patterns (Chehata et al., 2013) and creating probabilistic frameworks to map vegetation, like tree cover delineation (Basu et al., 2015).

This project focuses on characterizing cropland area, both fallow and active, comprised entirely of smallholder farms (i.e., agricultural fields of ≤ 45 × 45 m or 0.5 acres) that are difficult to resolve with moderate to coarse resolution imagery in a rainfed residential cropland mosaic (See et al., 2015, Ellis and Ramankutty, 2008). In this study we define cropped area as cultivated crop areas and pasture. Agricultural areas in Ethiopia are diverse and multi-use by necessity (Belay et al., 2015). We assume that cropped area can then be extracted from remote sensing imagery as homogeneous texture objects that exist in the landscape at ~ 10 × 10 to 30 × 30 m scale with vegetation growth and senescence that can be distinguished from natural heterogeneous objects. Rainfed agriculture accounts for 90–95% of croplands in Sub-Saharan Africa (SSA) and is characterized by small, unshifting fields (Sheahan and Barrett, 2017, IWMI, 2016, World Bank Group, 2016). Field boundaries can be easily distinguishable in VHR data in areas with established croplands, though less clear when considering the impact of fallow fields and transition to pasture and savanna. However, processing large volumes of VHR scenes for a region is computationally intensive, and has been previously a cost prohibitive task, so it typically has not been done. Standard costs for DigitalGlobe commercial archived imagery are >$10 per square km. Without government-licensed access, this study would have been cost prohibitive and no direct cost access to DigitalGlobe data via the NextView license agreement made this study possible (Neigh et al., 2013). With these challenges in mind, this project completed a near wall-to-wall segmentation mapping of cropped area using VHR data. The VHR cropped area can be used to target financial instruments like crop index insurance for farmers in SSA, thus providing monetary and insurance support to improve food security.