A High-Resolution Remote Sensing Semantic Segmentation Dataset for Cultivated, Forest, and Garden Land (cfg_v2)

<p>This dataset (data_final_balanced_v2_unique) is a high-resolution remote sensing semantic segmentation benchmark specifically designed for the fine-grained classification of Cultivated Land, Forest Land, and Garden Land. It addresses the challenges of distinguishing subtle differences between various agricultural and natural green covers using automated weak-supervision labels.</p> <p>The dataset contains 2,842 highly curated, spatially unique GeoTIFF image-mask pairs. The input imagery is sourced from high-resolution NAIP aerial data. The corresponding ground-truth masks are synthetically fused from multi-source data: the Esri 10m Annual Land Cover dataset (for cultivated and forest land) and OpenStreetMap crowdsourced vectors (for specialized garden land such as orchards, vineyards, and nurseries).</p> <p>To ensure robust model training and prevent spatial data leakage, the dataset has undergone rigorous spatial deduplication (0 overlapping spatial windows). Furthermore, class balancing was achieved by downsampling majority classes, resulting in a healthy distribution across all categories. The dataset is officially partitioned into training (1,989), validation (426), and testing (427) splits, providing a reliable foundation for geospatial deep learning applications.</p>