Large area cropland extent mapping with Landsat data and a generalized classifier

Accurate and up-to-date cropland maps play an important role in the study of food security. Traditional mapping of croplands using medium resolution (10–100 m) remote sensing imagery involving a “one-time, one-place” approach requires significant computing and labor resources. Although high mapping accuracies can be achieved using this approach, it is tedious and expensive to collect reference information to train the classifiers at each location and to apply over large areas, such as a continent. Moreover, large area cropland mapping presents additional challenges including a wide range of agricultural management practices, climatic conditions, and crop types. To overcome these challenges, here we report on a generalized image classifier to map cropland extent, which builds a classification model using training data from one location and time period, applied to other times and locations without the need for additional training data. The study was demonstrated across eight agro-ecological zones (AEZs) in Europe, the Middle East and North Africa using Landsat data acquired between 2009 and 2011. To reduce between-scene variability associated with image availability and cloud cover, input data were reduced to salient temporal statistics derived from enhanced vegetation index (EVI) combined with topographic variables. The generalized classifier was then tested across three levels of generalization: 1. individual - where training data were extracted from and applied to the same Landsat footprint; 2. AEZ where training data were extracted from a set of Landsat footprints within an AEZ and applied to any other Landsat footprint in the same AEZ; and 3. regional where training data were extracted from a set of Landsat footprints in the whole study area and applied to any other Landsat footprint inside the study area. Results showed that the generalized classifier is successful in identifying and mapping croplands with comparable success across all three levels of generalization with minimal cost: average loss in accuracy (as measured by overall accuracy) from the individual level (average overall accuracy of 80 ± 5%) to regional level (average overall accuracy of 74 ± 10%) is between 2 and 10% depending on the location. Results also show that generalization is not sensitive to the choice of the classification algorithm – the Linear Discriminant Analysis (LDA) model performs equally well compared to many popular machine learning algorithms found in the literature. This work suggests the generalization/signature extension framework has a great potential for rapid identification and mapping of croplands with reasonable accuracies over large areas using only easily computed vegetation indices with very little user input and ground information requirement.