Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Weak Labeling for Cropland Mapping in Africa

Gilles Quentin Hacheme, Akram Zaytar, Girmaw Abebe Tadesse, Caleb Robinson, Rahul Dodhia, Juan M. Lavista Ferres, Stephen Wood

TL;DR

This work tackles the scarcity of high-resolution cropland maps in Africa by integrating weak labels from global cropland maps with sparse human annotations through unsupervised object clustering. A region-specific workflow segments high-resolution imagery, intersects resulting objects with weak labels, and produces enhanced cropland/non-cropland samples to train a semantic segmentation model (U-Net with ResNet-50). Empirical results show that using mined negative labels alongside half the available human labels improves the cropland F$_1$ score from $0.53$ to $0.84$ (and non-cropland from around $0.96$ to $0.99$), whereas raw weak labels can hurt performance. The approach demonstrates a scalable path toward accurate, region-tailored cropland maps in Africa, reducing labeling burdens while enabling large-area mapping.

Abstract

Cropland mapping can play a vital role in addressing environmental, agricultural, and food security challenges. However, in the context of Africa, practical applications are often hindered by the limited availability of high-resolution cropland maps. Such maps typically require extensive human labeling, thereby creating a scalability bottleneck. To address this, we propose an approach that utilizes unsupervised object clustering to refine existing weak labels, such as those obtained from global cropland maps. The refined labels, in conjunction with sparse human annotations, serve as training data for a semantic segmentation network designed to identify cropland areas. We conduct experiments to demonstrate the benefits of the improved weak labels generated by our method. In a scenario where we train our model with only 33 human-annotated labels, the F_1 score for the cropland category increases from 0.53 to 0.84 when we add the mined negative labels.

Weak Labeling for Cropland Mapping in Africa

TL;DR

This work tackles the scarcity of high-resolution cropland maps in Africa by integrating weak labels from global cropland maps with sparse human annotations through unsupervised object clustering. A region-specific workflow segments high-resolution imagery, intersects resulting objects with weak labels, and produces enhanced cropland/non-cropland samples to train a semantic segmentation model (U-Net with ResNet-50). Empirical results show that using mined negative labels alongside half the available human labels improves the cropland F score from to (and non-cropland from around to ), whereas raw weak labels can hurt performance. The approach demonstrates a scalable path toward accurate, region-tailored cropland maps in Africa, reducing labeling burdens while enabling large-area mapping.

Abstract

Cropland mapping can play a vital role in addressing environmental, agricultural, and food security challenges. However, in the context of Africa, practical applications are often hindered by the limited availability of high-resolution cropland maps. Such maps typically require extensive human labeling, thereby creating a scalability bottleneck. To address this, we propose an approach that utilizes unsupervised object clustering to refine existing weak labels, such as those obtained from global cropland maps. The refined labels, in conjunction with sparse human annotations, serve as training data for a semantic segmentation network designed to identify cropland areas. We conduct experiments to demonstrate the benefits of the improved weak labels generated by our method. In a scenario where we train our model with only 33 human-annotated labels, the F_1 score for the cropland category increases from 0.53 to 0.84 when we add the mined negative labels.
Paper Structure (7 sections, 1 figure, 1 table)

This paper contains 7 sections, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Problem statement
  3. Cluster-based refinement of weak cropland labels
  4. Data
  5. Method
  6. Experiments
  7. Conclusion

Figures (1)

  • Figure 1: An overview of our proposed approach. Given satellite imagery (A) and weak cropland labels (C) over a given AOI we first use a K-Means clustering and filtering method to perform unsupervised object segmentation of the imagery (B). We intersect the resulting objects (polygons) with the weak labels to mine stronger positive and negative samples (D). Our experimental results show that adding these mined labels to human labels improves model performance.