Program


Friday 6th November Friday 13th November Friday 20th November Friday 27th November
Friday 4th December

Back to program

A picture's worth...Improving the use of geographical characteristics in surveys

Friday 20th November, 10:00 - 11:30

Advances in big data and the impact on grid based sampling

Ms Justine Allpress (RTI International) - Presenting Author
Mr James Cajka (RTI International)
Dr Stephanie Eckman (RTI International)
Dr Charles Lau (RTI International)

Geosampling is sampling methodology which selects representative samples in locations where frames do not exist or are of poor quality. The method takes advantage of gridded population estimates which are models of human populations distributed across the world’s surface. RTI International (RTI) developed this approach and has used it in many low- and middle-income countries.

We have enhanced the Geosampling approach over the years in response to drastic improvements in data resolution and machine learning. The flexible structure of Geosampling has allowed it to take advantage of rapid advances in data quality, quantity, resolution, and availability. The first Geosampling studies used population estimates in 1km x 1km grid cells. The most recent rounds have used 100m x 100m cells. We have also incorporated computer vision models. We extracted aerial imagery for each component cell and used a machine learning model to identify building outlines. A sample of buildings were selected and plotted on GPS-enabled maps, which allowed field staff to navigate to the buildings in sequence.

Looking to the future, we are interested in the worldwide building footprint datasets which are becoming available. Once these data are available at affordable cost and attributed with residential classification and the number of occupants, the development of sampling frames will take another step forward in efficiency and accuracy. With each new project, we can push this methodology forward. This presentation will be of interest to researchers who want to select high-quality samples in areas where good census data are not available.


Street Sense: Learning from Google Street View

Dr Gaurav Sood (Convoy) - Presenting Author
Miss Kimberly Ortleb (University of Texas, Austin)
Mr Suriyan Laohaprapanon (Appeler)

How good are the public services and the public infrastructure? Does their quality vary by income? These are vital questions—they shed light on how well the government is doing its job, the consequences of disparities in local funding, etc. But there is little good data on many of these questions. We fill this gap by describing a scalable method of getting data on one crucial piece of public infrastructure: roads. We assess the quality of roads and sidewalks by exploiting data from Google Street View.

To efficiently learn about the condition of the streets, sidewalks, and such, from Google Street View data, we devise a new workflow. We start by downloading data on the kinds of roads we are interested from the Open Street Map (OSM). We then chunk the roads into half a kilometer segments, and then randomly sample from the segments. We provide an opensource Python package that implements this workflow. We then take the starting latitude and longitude of the sampled segments and query the Google Street View API. We download the images for those locations and code the images using Amazon’s Mechanical Turk.

We apply this method to assess the quality of roads in Bangkok, Jakarta, Lagos, and Wayne County, Michigan. Jakarta’s roads have nearly four times the potholes than roads of any other city. Surprisingly, the proportion of road segments with potholes in Bangkok, Lagos, and Wayne is about the same, between .06 and .07. Using the data, we also estimate the relation between the condition of the roads and local income in Wayne, MI. We find that roads in more affluent census tracts have somewhat fewer potholes.


Remote sensing in support of agricultural surveys: use of Sentinel images and UAV-acquired ground truth for crop mapping in Rwanda

Dr Dorota Temple (RTI International) - Presenting Author
Mr Jason Polly (RTI International)
Dr Meghan Hegarty-Craver (RTI)
Ms Maggie O'Neil (RTI)
Mr Noel Ujeneza (Agri-Consultant, Kigali, Rwanda)
Mr James Rineer (RTI International)
Mr Daniel Lapidus (RTI International)
Dr Robert Beach (RTI International )

There is an increasing interest in augmenting agricultural surveys with information from other sources. Recent years brought the launch of Sentinel satellites carrying sensors with the ground resolution of 10 m, high enough to image individual fields. This new capability, aided by the establishment of open-access infrastructure for processing of the high-resolution images and the recent progress in artificial intelligence, makes satellite-based crop analytics increasingly relevant to agricultural information systems. This is particularly important in low resource environments where frequent and extensive field surveys are not practical. In this presentation, we will discuss the use of freely available satellite images to provide information about agricultural production in low- and medium-income countries, and we will comment on research challenges and opportunities. We will highlight the near-term potential of the methodology in the context of Rwanda, a country in sub-Saharan Africa, whose government has recognized early the value of information technology in its strategic planning for food security and sustainability.
Specifically, we will discuss a machine-learning (ML) model we have developed for mapping of key crops in Rwanda. To construct ground-truth datasets for training of ML algorithms, we collected high-resolution imagery using unmanned aerial vehicles (UAVs) in a series of flights conducted by Charis Unmanned Aerial Solutions (Kigali, Rwanda). The flights covered approximately 500 ha in six locations selected to represent different agroecological zones. The UAVs were equipped with high-resolution RGB cameras and included a built-in data-link connection to a base station so that the images could be georeferenced to within a 10-cm accuracy. The images had the ground resolution of 3 cm, enabling a human analyst to label the type of crops in multiple locations captured by the UAV. We created the ground-truth dataset by registering labeled locations in the image grid of Sentinel satellites and by obtaining signal values in selected optical and radar bands of Sentinel-1 and Sentinel-2 sensors, using the open-access Google Earth Engine data-processing infrastructure.
We trained a Random Forest (RF) algorithm to classify Sentinel image pixels into one of seven classes: maize, beans, bananas, cassava (key crops in Rwanda), other vegetation, forest, and a non-vegetative category that included bare ground, roads and buildings. We characterized performance of the model by classifying a set-aside test ground-truth data. We obtained the overall accuracy of approximately 80%, with the accuracy of maize and bananas approaching 85%. We will present country-level crop maps generated using the developed RF model and will discuss challenges posed by the smallholder nature of the Rwandan agriculture and by the frequently cloudy weather during the growing season that may interfere with remote sensing. We will also discuss future research directions in the context of the increasing relevance of satellite-based crop analytics to agricultural information systems in resource-limited environments.
Acknowledgment: This work was supported in part by the RTI Grand Challenge Initiative in Food Security and Sustainability.


Complimenting agricultural surveys in Rwanda: Using deep learning to classify crops from drone imagery

Mr Rob Chew (RTI International) - Presenting Author
Mr Jay Rineer (RTI International)
Dr Robert Beach (RTI International)
Ms Maggie O'Neil (RTI International)
Mr Noel Ujeneza (Independent Agri-Consultant)
Mr Daniel Lapidus (RTI International)
Mr Thomas Miano (RTI International)
Dr Meghan Hegarty-Craver (RTI International)
Dr Dorota Temple (RTI International)

Although Rwanda has experienced improvements in agricultural productivity in recent years, one fifth of its population remains food insecure. The projected impacts of population growth and climate change will exacerbate the ongoing challenge of further improving food security. While district-level surveys conducted by Rwandan Ministry of Agriculture and Animal Resources (MINAGRI) provide a wealth of information about Rwandan agriculture, results are not available in time to inform decisions during the growing season. More timely information could aid in yield prediction and assist in improved resource allocation to regions where the growing season may be delayed or where crops are not growing as expected. Unfortunately, acquiring timely and accurate crop estimates is particularly challenging in smallholder farming systems like Rwanda due to the abundance of small plots, intense intercropping, and high diversity of crop types.

In this study, we trained a deep convolutional neural network to remotely identify strategic crop types in Rwandan agriculture (bananas, maize, and legumes). This model was trained on high-resolution RGB images collected from multiple drone flights across six sites in Rwanda. To reduce the burden associated with field data collection, a remote web-based GIS system was used to label crops for the training and test data sets. A local Rwandan agricultural expert performed the initial labeling of crops and supervised a team of three independent labelers remotely. These findings suggest that though certain staple crops such as maize and bananas can be classified at this scale with high accuracy, crops involved in heavy intercropping (legumes) can still be difficult to consistently identify.

We discuss potential use cases and implications for these findings, with recommendations for future research. In particular, we explore the use of drone crop predictions as training data for developing country-wide models using multi-spectral satellite imagery, the use of drone crop classifications in sampled plots to aid agricultural surveys, and opportunities for remote sensing to provide timely updates on key crop progress between publication of seasonal agricultural surveys.


Land take analysis : A neural network applied to the shape of cadastral parcels

Ms Stéphanie Himpens (National Institute of Statistics and Economic Studies) - Presenting Author
Ms Mathilde Poulhes (National Institute of Statistics and Economic Studies)
Mr François Sémécurbe (National Institute of Statistics and Economic Studies)

Limiting land take has become a major political concern since soils provide a very wide range of vital ecosystem functions. Soil sealing, i.e. the covering of the ground by an impermeable material, is one of the main causes of soil degradation: it often affects fertile agricultural land and puts biodiversity at risk.
Land consumption has been widely documented at the national level and some local analysis showed that land take is more or less spatially concentrated on French territory, from constructions located nearby already built-up areas that contribute to expand urban zones to isolated constructions in rural areas corresponding to unplanned land take (Albizzati et al., 2017). But to our knowledge, no determinant analysis of these phenomena has been conducted.
In this project we try to link the shape of the cadastral parcels to the characteristics of land take (more or less concentrated). Based on the methodology of Louf et al. (2014) and Moosavi (2017), we apply different neural networks to images of land register in order to distinguish between different shapes of cadastral parcels. Supervised networks are tested and their capacity to differentiate scattered settlements from grouped settlements (bocage vs open field) is assessed.
The results enable us to identify zones where different settlements coexist and to evaluate the influence of the type of settlement on the localisation of new construction (we exploit building permits between 2000 and 2014).