import kornia as K import torch from torchgeo.datasets.geo import NonGeoDataset import os from collections.abc import Callable, Sequence from torch import Tensor import numpy as np import rasterio import cv2 from pyproj import Transformer from datetime import date from typing import TypeAlias, ClassVar import pathlib import logging import pdb logging.getLogger("rasterio").setLevel(logging.ERROR) Path: TypeAlias = str | os.PathLike[str] class SenBenchAirQualityS5P(NonGeoDataset): url = None splits = ('train', 'val', 'test') split_fnames = { 'train': 'train.csv', 'val': 'val.csv', 'test': 'test.csv', } def __init__( self, root: Path = 'data', split: str = 'train', modality = 'no2', # or 'o3' mode = 'annual', # or 'seasonal' transforms: Callable[[dict[str, Tensor]], dict[str, Tensor]] | None = None, download: bool = False, ) -> None: self.root = root self.transforms = transforms self.download = download #self.checksum = checksum assert split in ['train', 'val', 'test'] self.modality = modality self.mode = mode if self.mode == 'annual': mode_dir = 's5p_annual' elif self.mode == 'seasonal': mode_dir = 's5p_seasonal' self.img_dir = os.path.join(root, modality, mode_dir) self.label_dir = os.path.join(root, modality, 'label_annual') self.split_csv = os.path.join(self.root, modality, self.split_fnames[split]) with open(self.split_csv, 'r') as f: lines = f.readlines() self.pids = [] for line in lines: self.pids.append(line.strip()) self.reference_date = date(1970, 1, 1) self.patch_area = (16*10/1000)**2 # patchsize 16 pix, gsd 10m def __len__(self): return len(self.pids) def __getitem__(self, index): images, meta_infos = self._load_image(index) #meta_info = np.array([coord[0], coord[1], np.nan, self.patch_area]).astype(np.float32) label = self._load_target(index) if self.mode == 'annual': sample = {'image': images[0], 'groudtruth': label, 'meta': meta_infos[0]} elif self.mode == 'seasonal': sample = {'image': images, 'groudtruth': label, 'meta': meta_infos} #pdb.set_trace() if self.transforms is not None: sample = self.transforms(sample) return sample def _load_image(self, index): pid = self.pids[index] s5p_path = os.path.join(self.img_dir, pid) img_fnames = os.listdir(s5p_path) s5p_paths = [] for img_fname in img_fnames: s5p_paths.append(os.path.join(s5p_path, img_fname)) imgs = [] meta_infos = [] for img_path in s5p_paths: with rasterio.open(img_path) as src: img = src.read(1) img[np.isnan(img)] = 0 img = cv2.resize(img, (56,56), interpolation=cv2.INTER_CUBIC) img = torch.from_numpy(img).float() img = img.unsqueeze(0) # get lon, lat cx,cy = src.xy(src.height // 2, src.width // 2) if src.crs.to_string() != 'EPSG:4326': # convert to lon, lat crs_transformer = Transformer.from_crs(src.crs, 'epsg:4326', always_xy=True) lon, lat = crs_transformer.transform(cx,cy) else: lon, lat = cx, cy # get time img_fname = os.path.basename(img_path) date_str = img_fname.split('_')[0][:10] date_obj = date(int(date_str[:4]), int(date_str[5:7]), int(date_str[8:10])) delta = (date_obj - self.reference_date).days meta_info = np.array([lon, lat, delta, self.patch_area]).astype(np.float32) meta_info = torch.from_numpy(meta_info) imgs.append(img) meta_infos.append(meta_info) if self.mode == 'seasonal': # pad to 4 images if less than 4 while len(imgs) < 4: imgs.append(img) meta_infos.append(meta_info) return imgs, meta_infos def _load_target(self, index): pid = self.pids[index] label_path = os.path.join(self.label_dir, pid+'.tif') with rasterio.open(label_path) as src: label = src.read(1) label = cv2.resize(label, (56,56), interpolation=cv2.INTER_NEAREST) # 0-650 # label contains -inf #pdb.set_trace() label[label<-1e10] = np.nan label[label>1e10] = np.nan label = torch.from_numpy(label.astype('float32')) return label class RegDataAugmentation(torch.nn.Module): def __init__(self, split, size): super().__init__() mean = torch.Tensor([0.0]) std = torch.Tensor([1.0]) self.norm = K.augmentation.Normalize(mean=mean, std=std) if split == "train": self.transform = K.augmentation.AugmentationSequential( K.augmentation.Resize(size=size, align_corners=True), K.augmentation.RandomRotation(degrees=90, p=0.5, align_corners=True), K.augmentation.RandomHorizontalFlip(p=0.5), K.augmentation.RandomVerticalFlip(p=0.5), data_keys=["input", "input"], ) else: self.transform = K.augmentation.AugmentationSequential( K.augmentation.Resize(size=size, align_corners=True), data_keys=["input", "input"], ) @torch.no_grad() def forward(self, batch: dict[str,]): """Torchgeo returns a dictionary with 'image' and 'label' keys, but engine expects a tuple""" x,mask = batch["image"], batch["groudtruth"] x = self.norm(x) x_out, mask_out = self.transform(x, mask.unsqueeze(0)) return x_out.squeeze(0), mask_out.squeeze(0), batch["meta"] class SenBenchAirQualityS5PDataset: def __init__(self, config): self.dataset_config = config self.img_size = (config.image_resolution, config.image_resolution) self.root_dir = config.data_path self.modality = config.modality self.mode = config.mode def create_dataset(self): train_transform = RegDataAugmentation(split="train", size=self.img_size) eval_transform = RegDataAugmentation(split="test", size=self.img_size) dataset_train = SenBenchAirQualityS5P( root=self.root_dir, split="train", modality=self.modality, mode=self.mode, transforms=train_transform ) dataset_val = SenBenchAirQualityS5P( root=self.root_dir, split="val", modality=self.modality, mode=self.mode, transforms=eval_transform ) dataset_test = SenBenchAirQualityS5P( root=self.root_dir, split="test", modality=self.modality, mode=self.mode, transforms=eval_transform ) return dataset_train, dataset_val, dataset_test