Crop Map Review

Review and score a crop-type or land-cover classification against ground truth. Load your field parcels as GeoJSON, drape them over satellite imagery, then colour them by predicted class, by ground truth, or by AGREEMENT so every misclassified field lights up red on the map. Press Score and the plate computes the confusion matrix, overall accuracy, Cohen's kappa and per-class precision / recall / F1 / IoU from your own table — no spreadsheet, and no hand-typed matrix. Click a field to inspect its attributes, capture the current view as an area of interest, sweep a confidence threshold to see which detections survive, and plot per-parcel NDVI over the season to check whether two crops are separable at all. Built for the accuracy-assessment half of a deep-learning mapping project, where the model is trained elsewhere and the question is how well it actually did. LIMITS, STATED RATHER THAN DISCOVERED: no model training or inference happens here — that stays in PyTorch, TensorFlow or ArcGIS Pro, and this reviews the output. No raster support: imagery is XYZ basemap tiles, so GeoTIFF and Cloud-Optimized GeoTIFF are out, as are WMS and WMTS. No band math — NDVI arrives as a parcel attribute, which is exactly what a Google Earth Engine or ArcGIS zonal-statistics export produces. GeoJSON must be WGS84 lon/lat (EPSG:4326); projected input is refused with a message rather than silently plotted in the wrong hemisphere. Hand-labelling is not included: this plate scores labels you already have, it does not help you draw new ones. Area of interest capture records the current map extent, it is not a freehand digitising tool. There is deliberately no snapshot button, because map tiles cannot be captured and it would hand you a PNG with a hole in it.

Supported file formats: CSV, TSV