agroimpacts / activemapper

Code and data for the paper "High resolution, annual maps of field boundaries for smallholder-dominated croplands at national scales"

activemapper

Repository for the manuscript High resolution, annual maps of field boundaries for smallholder-dominated croplands at national scales.

This branch contains version 4 of the manuscript, which was accepted for publication in Frontiers in Artificial Intelligence.

Citation

Estes, L.D., Ye, S., Song, L., Luo, B., Eastman, J.R., Meng, Z., Zhang, Q., McRitchie, D., Debats, S.R., Muhando, J., Amukoa, A.H., Kaloo, B.W., Makuru, J., Mbatia, B.K., Muasa, I.M., Mucha, J., Mugami, A.M., Mugami, J.M., Muinde, F.W., Mwawaza, F.M., Ochieng, J., Oduol, C.J., Oduor, P., Wanjiku, T., Wanyoike, J.G., Avery, R. & Caylor, K. (2021) High resolution, annual maps of the characteristics of smallholder-dominated croplands at national scales. Frontiers in Artificial Intelligence 10.3389/frai.2021.744863

activemapper's Issues

Re-run learner with new labels

• First create new consensus labels @ldemaz
• Then retrain on labeller15, testing against current high quality validation labels, iterations 1-3 (already redid with initial train)
• Then 1, 2, 3 next. Evaluate from there
• Make predictions is warranted

Please review these edits to the segmentation methods

activemapper/inst/manuscript/manuscript.Rmd

Lines 363 to 373 in f002fc0

	Upon completion of the active learning process, we deployed a five-step algorithm to create a segmented map of field boundaries. In the first step, we identified edge features within the imagery. To do this, we applied the meanshift algorithm [@YizongChengMeanshiftmode1995a] to each dry-season composite tile, and then passed a Sobel filter over the mean-shifted green, red, and near-infrared bands, and the corresponding map of predicted cropland probabilities. We then summed the four resulting edge images to produce a combined edge image.
	In the second step, we used a compact watershed algorithm [@neubertCompactWatershedPreemptive2014] to segment the weighted edge image, specifying a high number of segments (6,400) segments per tile, so that the mean segment size (<0.5 ha) was finer than the expected mean field size (1-2 ha).
	In the third step, we hierarchically merged the resulting polygons. We first constructed a region adjacency graph for each tile, with each node representing all image pixels within each polygon. The edge between two adjacent regions (polygons) was calculated as the difference between the means of the normalized colors of all bands. We then merged the most similar pairs of adjacent nodes until there were no edges remaining below the predetermined threshold of 0.05.
	In the fourth step, we overlaid the merged polygons with the cropland probability images, and polygons in which the mean probability was greater than 0.5 were retained as crop fields.
	In the fifth and final step, we refined the crop field polygons, by removing holes and smoothing boundaries using the Visvalingam algorithm [@visvalingamLineGeneralisationRepeated1993]. We then merged neighboring polygons that overlapped along tile boundaries.
	The resulting map represents dry season crop field boundaries, as we did not segment growing season images. We made this choice because labels were primarily drawn on dry season composites, when boundaries were typically more visible.

Apply new segmentation algorithm on activelearning results

• Test against validation sites and current prediction maps to see it improves size class distributions
• Reprocess AOIs as needed if improvement, either with or without new predictions