WBCrop: A georeferenced dataset of 18 crop types across West Bengal, India

Dataset Description Developing robust agricultural monitoring systems for smallholder farming landscapes is fundamentally challenged by the scarcity of geographically diverse, field-level training data. To address this gap, we introduce WBCrop, an open-access, manually curated, and georeferenced dataset comprising 45,616 ground-truth samples across 18 distinct crop types in West Bengal, India. The dataset was systematically collected over a full agricultural year (July 2023 to June 2024) to encompass the complete crop phenological cycles of the primary Kharif, Rabi, and Zaid seasons, along with perennial crops. Data acquisition integrates extensive in-situ field surveys (9,243 points) with a scalable remote annotation pipeline utilizing Google Street View and high-resolution Google satellite imagery (36,373 points). Particular emphasis was placed on maintaining class balance to mitigate long-tail distribution biases commonly observed in real-world agricultural datasets. This dataset provides a robust foundational resource for advancing supervised learning, foundation model fine-tuning, domain adaptation studies, and large-scale agricultural analytics in smallholder ecosystems. Technical Overview File Name: `wbcrop_points.gpkg` Total Records: 45,616 point Geometry Type: Point Coordinate Reference System (CRS): EPSG:4326 (WGS 84) Temporal Coverage: Sowing/harvest cycles spanning July 2023 to June 2024 (Collection dates range from 2023-01-10 to 2024-08-21). Geographic Coverage: 21 administrative districts across West Bengal, India. Variable Details Attribute Data Type Description id Integer Unique identifier for each sample point. crop String Crop type label. One of 18 classes (see Fig. 4). district String Administrative district name within West Bengal (21 districts). state String State name; constant value ‘west_bengal’ for all records. sowing String Approximate sowing period (e.g., jul_2023, feb_mar_2024). Value is NA if unavailable. harvest String Approximate harvest period (e.g., nov_2023, mar_apr_2024). Value is NA if unavailable. collection_date String Date of field data collection (YYYY-MM-DD). Ranges from 2023-01-10 to 2024-08-21. 90 unique dates. season String Cropping season: kharif, rabi, zaid, or perennial. pheno_stage String Crop phenological stage at the time of collection (e.g., vegetative, reproductive, maturity). latitude Float Latitude of the sample location in decimal degrees (WGS 84). longitude Float Longitude of the sample location in decimal degrees (WGS 84). geometry Point Point geometry (EPSG:4326) Usage Notes Spatial Projections: The dataset is provided in unprojected geographic coordinates (EPSG:4326). Users should ensure that any spatial analysis or integration with other geospatial datasets accounts for this CRS, reprojecting as necessary. Temporal Considerations: The `collection_date` field records either the timestamp embedded in the Google Street View imagery or the actual field survey date for each sample. Account for this when analyzing crop phenology or seasonality. Code Availability: All custom code, scripts, and notebooks utilized for data pre-processing and technical validation are open-source and hosted on GitHub at: https://github.com/geonextgis/WBCrop.git