README.md

---
license: apache-2.0
language:
- en
tags:
- Sentinel-2
- Sentinel-1
- Cropland
- Satellite_image
- Remote_sensing
size_categories:
- n>1T
---


# Sentinel Patch Imagery Dataset for China's Cropland

## 1. Dataset Description

- **Dataset Name:** Sentinel Patch Imagery Dataset for China's Cropland
- **Version:** V1.0
- **Authors:** Mo Wang, Jing Wang, Yunpeng Cui, Juan Liu, Li Chen, Ting Wang, Jinming Wu, Huan Li
- **License:** apache-2.0
- **Examples:** chunk_41 (Example_S1_chunk_41.zip.001 and Example_S2_chunk_41.zip.001) serves as example datasets as this chunk is small.

### 1.1. Overview

This project is inspired by Copernicus-Pretrain dataset (https://huggingface.co/datasets/wangyi111/Copernicus-Pretrain). However, we only focus on cropland. This dataset provides a curated collection of multi-sensor, multi-temporal image patches focusing exclusively on agricultural land across China in 2021. The dataset contains Sentinel-1 (Sentinel-1 SAR GRD: C-band Synthetic Aperture Radar Ground Range Detected, log scaling) and Sentinel-2 (Harmonized Sentinel-2 MSI: MultiSpectral Instrument, Level-2A (SR)) imagery covering China where cropland pixel percentage is no less than 80%. Sentinel-2 images are in area of 2640 m * 2640 m. Pixel dimensions are slightly greater than 264 * 264 dependding on their geolocation. Sentinel-1 images are standardized to a uniform pixel dimension 264*264 and a consistent Coordinate Reference System (CRS) for time-series analysis. The data acquisition is governed by a rigorous workflow designed to maximize geometric and temporal consistency across four key seasonal stages.

### 1.2. Data Sources

The raw data is sourced from the European Space Agency's (ESA) Copernicus Sentinel missions, accessed via the Google Earth Engine (GEE) platform.

| Sensor | GEE Collection ID | Bands | Resolution (Raw) |
| :--- | :--- | :--- | :--- |
| **Sentinel-2** (Optical) | `COPERNICUS/S2_HARMONIZED` | 13 bands (B1-B12) | 10 meters |
| **Sentinel-1** (SAR) | `COPERNICUS/S1_GRD` | 2 bands (VV, VH) | 10 meters |

### 1.3. Geographic and Temporal Coverage

* **Geographic Scope:** Agricultural areas across the entire territory of China.
* **Target Land Cover:** Each sample location is centered on agricultural land, ensuring a high proportion of agricultural land use (pixel percentage >80%).
* **Temporal Depth:** 4 distinct time steps (patches) are collected for each location and sensor, aligning with seasonal dates.
    * **Acquisition Window:** $\pm 30$ days around four seasonal reference dates (e.g., Solstices and Equinoxes).

## 2. Dataset Structure and Data Fields

The dataset is organized hierarchically, reflecting the grid location, unique sample ID, and temporal patches. The core data is composed of GeoTIFF files.

### 2.1. File Naming Convention (Post-Processed)

The final data structure is organized as:

`{base_output_dir}/{grid_code}_{grid_center_lon}_{grid_center_lat}/{point_id}/{gee_id}.tif`

* **`point\_id`:** Unique identifier combining the grid code and CSV file line index (e.g., `44RPQ_123`).
* **`gee\_id`:** The original GEE system index, which encodes the acquisition date and MGRS tile information (e.g., `20210621T030519_20210621T051147_T44RPQ`).

### 2.2. Standardized Patch Properties

All patches, regardless of the sensor, have been harmonized during post-processing to the following properties:

| Property | Value |
| :--- | :--- |
| **Pixel Dimensions (Height x Width)** | 264*264 |
| **Final Spatial Extent** | 2,640m x 2,640m |
| **Data Type** | **Float32** |
| **Coordinate Reference System (CRS)** | Harmonized to the **Majority CRS** (e.g., UTM) of the raw time series stack |
| **File Format** | GeoTIFF (.tif) |

## 3. Pre-processing and Standardization Workflow

The data preparation involves two major phases: GEE Acquisition and Local Post-Processing.

### 3.1. GEE Acquisition Filters

| Sensor | Acquisition Logic & Filters | Source |
| :--- | :--- | :--- |
| **S2** | **Cloud Filter:** <=20% | |
| | **Geometric Consistency:** Requires all 4 selected images for a location to share the **identical MGRS tile code**. | |
| **S1** | **Orbit Priority:** Prioritizes **DESCENDING** orbit images; falls back to **ASCENDING** if no valid DESCENDING image is found. | |

### 3.2. Local Post-Processing for Sentinel-1 GRD images

The following steps were applied to the raw downloaded GeoTIFFs:

1.  **Reference Center Calculation:** The geographic coordinates ($\text{lon}$, $\text{lat}$) of the center pixel of the **first image** in the raw sequence are used as the consistent cropping reference point.
2.  **CRS and Transform Alignment:**
    * The **majority CRS** and its corresponding Affine Transform (`majority\_transform`) are identified for the time series stack.
    * Any image with a non-matching CRS/Transform is **reprojected** to the majority standard.
    * **Resampling Method:** Nearest Neighbor.
3.  **Cropping:** All images are cropped to the 264*264 pixel size centered on the calculated reference point.
4.  **Data Type Conversion:** Final conversion to **Float32**.

## 4. Intended Uses and Limitations

### 4.1. Intended Uses

* **Multi-Modal Remote Sensing:** Training and benchmarking models that fuse SAR (S1) and Optical (S2) data for robust land cover classification and monitoring.
* **Time-Series Analysis:** Developing and validating spatio-temporal models (e.g., RNNs, 3D CNNs) for crop type identification and phenology monitoring based on consistent seasonal time steps.
* **Agricultural Intelligence:** Supporting research on agricultural classification, yield prediction, and land use mapping in China.

### 4.2. Limitations and Caveats

* **Temporal Gaps:** The data relies on a $\pm 30$ day window around seasonal peaks. The actual acquisition date can vary widely within this window, leading to some temporal jitter.
* **Resampling Method:** The use of Nearest Neighbor resampling for reprojection preserves original pixel values but may lead to minor geometric imperfections or stair-step effects on boundaries, especially if the raw image CRS significantly differed from the final majority CRS.
* **Cloud Cover (S2):** Although filtered, the S2 data can still contain up to $20\%$ cloud cover or shadow pixels. Further masking or cloud removal may be necessary for some applications.
* **S1 Orbit Mix:** While the S1 workflow prioritizes DESCENDING, some time steps may rely on ASCENDING images. Users should account for the potential change in look angle and its effect on SAR backscatter signatures.

## 5. Citing This Dataset

Wang, M., et al. (2025). Sentinel Patch Imagery Dataset for China's Cropland (Revision 8e988a5) [Data set]. Hugging Face. DOI: 10.57967/hf/6738