Evaluating Self-Supervised Redundancy-Reduction Representations for Intra-Site and Inter-Site Tropical Tree Species Classification from Airborne Hyperspectral Imagery : Dataset

Hyperspectral Spectral-Pixel Datasets for Tropical Tree Species Classification This repository accompanies the paper "Evaluating Self-Supervised Redundancy-Reduction Representations for Intra-Site and Inter-Site Tropical Tree Species Classification from Airborne Hyperspectral Imagery" and provides the spectral-pixel data used to train and evaluate the models. All data are stored as Python .pickle files. Each spectrum has a spectral dimension of 343 bands (Ndim = 343). Two datasets are provided: a supervised dataset for classification and a self-supervised dataset of spectral pairs. 1. Supervised dataset : GroundTruth/ Labeled spectral pixels for 14 tropical tree species, each spectrum being a vector of length 343. Two classification regimes are provided: Intra-site: a single study site (Paracou), with the train and test partitions split across distinct tree crowns of the same area.Inter-site: training crowns from Paracou and test crowns from Nouragues. Each .pickle stores a dictionary with the following keys: Key Type Shape / contentx tensor Nsample × Ndim : the spectral pixels (Ndim = 343)y tensor Nsample : integer label assigned to each spectral pixelspecies list maps each integer label in y to its species name2. Self-supervised dataset : SSL_Datasets/ Pairs of spectral acquisitions built with several pairing strategies. Each strategy has its own subfolder, and every subfolder is split into several subsets ("shards") to allow lighter and more dynamic loading. Each shard is a .pickle storing a dictionary with: Key Type Shape / contentx tensor Nsample × 2·Ndim : the two spectra of each pair, stacked along the spectral axisDim_Spectra int spectral dimension of a single spectrum (343), used to split x into its two halves To recover the two views of a pair from x: import pickle with open(shard_path, "rb") as f: data = pickle.load(f) d = data["Dim_Spectra"] # 343x1 = data["x"][:, :d] # first viewx2 = data["x"][:, d:] # second (positive) view Each pairing strategy contains 10 shards of 100,000 samples, for a total of 1,000,000 spectral pairs per strategy. Citation If you use these data, please cite the accompanying paper and this archive %%%%%%%%%%%%%%%%## Scope of the released data The **supervised dataset** (`GroundTruth/`) is released in full: it is exactly the data used for the classification results reported in the paper. The **self-supervised dataset** (`SSL_Datasets/`) is a **subsample** of the pairs used for pretraining, released to keep the archive at a size suitable for distribution. Each strategy is provided as 10 shards of 100,000 pairs (1,000,000 pairs per strategy), whereas the full training sets ranged from 1.2 to 12.7 million pairs depending on the strategy (see Table II of the paper). The subsample preserves the construction logic of each strategy and is intended to allow the pipeline to be reproduced and inspected; retraining on it will not reproduce the reported accuracies exactly.