Automated Mapping of Land Cover Type within International Heterogenous Landscapes Using Sentinel-2 Imagery with...

Area of Science:

• Earth Observation
• Machine Learning
• Remote Sensing

Background:

• Existing land cover classification methods often require extensive time-series satellite imagery and complex models.
• A gap exists in automated training data generation for land cover mapping using sparse temporal data.

Purpose of the Study:

• To develop and evaluate a framework for automated land cover classification using shallow machine learning and low-density time series imagery.
• To generate accurate training data for land cover mapping from limited Sentinel-2 data.

Main Methods:

• Utilized two dates of Sentinel-2 imagery (winter and non-winter) across seven international sites.
• Employed spectral, textural, and distance decision functions combined with ancillary data to create binary masks for training data generation.
• Applied a random forest classifier with stepwise threshold adjustments and evaluated global and regional adaptive thresholds, including temporal corrections using NDVI composites.

Main Results:

• Regional adaptive thresholds improved land cover classification accuracy across nine-class (73.1%), six-class (82.8%), and five-class (85.1%) schemes compared to global thresholds.
• Temporally corrected models further enhanced accuracy, particularly for the five-class scheme (85.1%).
• The developed five- and six-class models demonstrated accuracies comparable to a manually labeled deep learning model (Esri).

Conclusions:

• A near-global framework for automated training data generation and land cover classification using shallow machine learning with low-density time series imagery is feasible.
• The methodology provides accurate land cover maps without the need for a full annual time series, offering a computationally efficient alternative.
• The results highlight the potential of shallow machine learning approaches for robust land cover mapping in diverse regions.