Authors: Lyndon Estes*, Clark University, Kelly Caylor, Earth Resources Institute and Department of Geography, University of California Santa Barbara, Stephanie Debat, Descartes Labs, Ron Eastman, Clark Labs and Graduate School of Geography, Clark University, David R Thompson, Jet Propulsion Lab
Topics: Remote Sensing, Land Use and Land Cover Change, Coupled Human and Natural Systems
Keywords: machine learning, computer vision, active learning, land cover, agriculture, sub-Saharan Africa
Session Type: Paper
Start / End Time: 8:00 AM / 9:40 AM
Room: Grand Ballroom A, Astor, 2nd Floor
Presentation File: No File Uploaded
During the next few decades, agriculture in Sub-Saharan Africa (SSA) will undergo a large-scale expansion to meet the region’s rapidly growing food demands. This change will have a significant impact on the trajectory of global change, but understanding that impact is difficult because mapping trends in the region’s smallholder-dominated croplands (which is critical to understanding agriculture-environment interactions) is a major Earth Observation challenge. To address this challenge, we present a new method for mapping land cover based on active learning. Active learning engages a large number of human mappers via a crowdsourcing platform to iteratively train and test a computer vision/machine learning algorithm. After initial training on a random sample, the algorithm directs the human mappers to collect new sites in the areas of highest classification uncertainty, whereupon it retrains and re-evaluates performance. This process iterates until accuracy gains saturate. Tests of the crowdsourcing platform demonstrate an ability to rapidly collect high quality training data, while the classifier shows a strong ability to distinguish cropland from non-cropland across agricultural systems, ranging from complex smallholder landscapes to large-scale, irrigated croplands. When these two components are combined in the active learning framework, high accuracy (True Skill Statistic > 0.65) classifications are produced that require substantially less training data (62% fewer sites) than an alternative approach using a purely randomized approach to sites selection. This method is now being scaled up on cloud computing infrastructure, and will be used to generate a next-generation cropland map for sub-Saharan Africa.