National-scale cropland mapping based on spectral-temporal features and outdated land cover information.

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Agricultural Monitoring

Background:

• Supervised learning for satellite-derived thematic maps is constrained by insufficient ground truth data.
• Frequent updates for large-area applications like cropland mapping are particularly challenging.
• Existing methods struggle with generating up-to-date, spatially consistent maps.

Purpose of the Study:

• To develop an automated method for producing spatially consistent cropland maps at the national scale.
• To overcome the limitations of insufficient ground truth data in supervised learning.
• To enable routine, operational crop monitoring through reliable mapping.

Main Methods:

• An unsupervised approach utilizing spectral-temporal features and outdated land cover information.
• Extraction of reliable calibration pixels based on existing labels and spectral signatures.
• Normalization of time series, derivation of spectral-temporal features, and country stratification for spatial consistency.
• Application of a weighted majority filter for speckle removal based on classification confidence.

Main Results:

• Achieved an overall accuracy of 92% for the cropland map in South Africa.
• Demonstrated large spatial variations in accuracy, with intensive grain-growing areas better characterized than smallholder systems.
• Identified key features, including vegetation minimum and short-wave infrared, as consistently important across different strata.

Conclusions:

• The proposed method shows significant potential for the routine production of consistent cropland maps over large areas.
• The approach effectively leverages outdated land cover data and spectral-temporal features for automated mapping.
• Future work may focus on improving characterization of diverse farming systems for more comprehensive monitoring.