Contrastive Pretrain and Supervised Fine-Tune for Land-Cover Classification in Data-Scarce Countries

Abstract

Deep neural networks are widely used to extract patterns and features from satellite imagery in Earth Observation, but their reliance on large labeled datasets limits deployment in data-scarce regions where ground surveys are expensive or infeasible. This study introduces a novel operational pipeline that couples SimCLR contrastive pretraining on unlabeled Sentinel-2 tiles with lightweight supervised fine-tuning to deliver tile-level predictions that are aggregated into district-level land-cover statistics, explicitly targeting administrative reporting rather than pixel-wise mapping. The key methodological contribution is this tile-to-district design—together with domain-aware, conservative augmentations and RGB-only pretraining—that aims to learn transferable representations under seasonality and atmospheric variability while requiring only a small expert-labeled set. The approach is evaluated in challenging contexts (with a pilot case study in Lesotho)and can generate updated district summaries within hours once trained, supporting rapid deployment where labels and compute are constrained. Benchmarking against ESA WorldCover and Google Dynamic World shows strong coherence of aggregated class distributions using cosine similarity, complemented by standard classification metrics on labeled tiles, indicating that reliable administrative-level indicators can be produced even without per-pixel agreement. These results underscore the practical relevance of the method for national statistical offices, humanitarian programs,and monitoring initiatives that need scalable, affordable land-cover statistics while minimizing dependence on extensive field campaigns.

Received: 1 October 2025 | Revised: 3 March 2026 | Accepted: 23 March 2026

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

Data sharing is not applicable to this article as no new data were created or analyzed in this study.

Author Contribution Statement

Gianfausto Bottini: Conceptualization, Methodology, Soft ware, Validation, Formal analysis, Investigation, Data curation, Writing – original draft, Writing – review &editing, Visualization. Francesco Nicola Tubiello: Resources, Writing – review &editing, Supervision, Project administration. Pengyu Hao: Writing – review &editing.