Supplementary Materials for "Semantic Segmentation of Coffee Fields Under Spatial Generalization: An Ablation Study with PlanetScope and Sentinel-2"

This record contains supplementary materials for the BRACIS 2026 paper “Semantic Segmentation of Coffee Fields Under Spatial Generalization: An Ablation Study with PlanetScope and Sentinel-2”. The materials support the experimental analyses reported in the paper, which evaluate coffee field semantic segmentation under municipality-level spatial generalization using PlanetScope and Sentinel-2 imagery. The experiments compare different segmentation architectures, loss functions, patch sizes, vegetation-index configurations, and ensemble inference with test-time augmentation. The record is organized into three ZIP files: code.zip: source code used for single-model inference, ensemble with test-time augmentation evaluation, and statistical analysis. dataset.zip: redistributable Sentinel-2 image patches, associated rasterized masks, and split metadata. results.zip: trained model weights, experimental result tables, statistical-analysis outputs, and prediction outputs associated with the Sentinel-2 and PlanetScope experiments. The Sentinel-2 patches and masks are provided as GeoTIFF (.tif) files. The Sentinel-2 image patches are distributed as scaled digital numbers (DN). To obtain surface reflectance values, users should apply the same scale factor described in the associated paper and Sentinel-2 product documentation. The rasterized masks are binary label masks, where pixel value 1 represents coffee and pixel value 0 represents background. The baseline Sentinel-2 inputs include visible and near-infrared bands, while vegetation-index configurations include an additional channel corresponding to EVI, GNDVI, or NDRE, depending on the experimental phase. The trained weights are stored together with their corresponding result folders to preserve the association between each checkpoint, sensor, experimental phase, input configuration, model architecture, loss function or vegetation-index setting, prediction outputs, and reported metrics. PlanetScope imagery is not redistributed in this record due to licensing restrictions. PlanetScope-related files included here are limited to trained weights, tabular results, prediction outputs, and other derived experimental artifacts associated with the reported experiments. A GitHub repository with the source code, an example pipeline notebook, execution instructions, and additional code-level documentation.