Using Crowdsourcing & Big Data to Understand Agriculture in Sub-Saharan Africa Dee Luo
Mapping Africa Project Current Issue: Inaccurate and unreliable representation of agricultural land Key issues: Food security, predicting expansion Solution? Mapping initiative using Amazon Mechanical Turk To get data for all of Sub-Saharan Africa, crowdsourcing expensive
Alternatives? Machine learning; Classification algorithms Merging fields of remote sensing and computer vision Random Forest Algorithm Schroff, F., Criminisi, A. and Zisserman, A.: Object Class Segmentation using Random Forests, Proceedings of the British Machine Vision Conference (2008)
Implementation Feature-based classification : field/nonfield RGB Edge detection Texture gradients Different ways of calculating thresholds Mean values Symmetric patches Absolute points Channel combinations Tokarczyk, P., Wegner J. D., Walk, S., Schindler, K.: Features, Color Spaces, and Boosting: New Insights on Semantic Classification of Remote Sensing Images
ImageHand Labeled Ground Truth
Image
Additional Work Acquiring and analysis of LANDSAT data Multi-spectral images, combinations of spectral bands R: filtering by loud cover, growing seasons, etc. Hand-digitization of field data set QGIS: spatial analysis
Current Results/Future Work Current accuracy at approx. 70% About 60% correctly labeled fields About 80% correctly labeled nonfields Stronger accuracy with large fields, much weaker on smaller residential fields Future improvements: Better Imagery – very high resolution (< 1m) Parameter optimization More features
Summary