4.1.

Data preprocessing and downloading on Google Earth Engine (GEE)

The initial retrieval and pre-processing of satellite data will be executed on GEE. GEE is a cloud computing platform designed for large-scale geospatial data analysis, enabling efficient processing of petabytes of remote sensing data.

• This involves accessing and applying initial remote sensing corrections (e.g., cloud masking, for Sentinel-2, speckle filtering, terrain correction for Sentinel-1) to the required Sentinel-1 (radar) and Sentinel-2 (optical) image collections.

Sentinel 1 Data Preprocessing

C Band SAR data from Senitnel 1 satellites will be use as the primary input features for training the deep learning model. The basic know-hows of the Sentinel 1 SAR data have been already explained in the previous lecture sections. The extraction of Sentinel 1 data was done through GEE image collection ‘COPERNICUS/S1_GRD_FLOAT’. The collection already provides two single co-polarization VV, HH two dual band cross polarization HV, VH, and angle band. For the practical usage, especially for training machine learning model the bands VV and VH has been widely used in the literatures (Nikaein et al., 2023). The Sentinel 1 data on GEE archive already comes in a pre-processed through thermal noise removal, radiometric calibration, and terrain correction. However, as SAR backscatters are highly noisy, speckle filtering is an essential correction.

Speckle Filtering

Speckle is a salt-and-pepper-like granular noise inherent to SAR images, resulting from the coherent nature of radar waves and the constructive and destructive interference among backscattered signals from many elementary scatterers within a single resolution cell. This noise reduces the visual quality and makes image interpretation and automated classification difficult. Speckle filters (e.g., Lee, Gamma Map, or refined Lee filters) are applied to reduce this noise while attempting to preserve the integrity of edges and subtle features, thereby enhancing the image quality and improving the accuracy of subsequent analysis.

Terrain Correction

Terrain correction, specifically Radiometric Terrain Correction (RTC), is applied to Sentinel-1 SAR data to account for distortions caused by topography. The intensity of the backscattered radar signal is significantly affected by the slope and aspect of the terrain relative to the radar beam. For instance, slopes facing the sensor appear brighter (foreshortening), while those facing away appear darker (shadowing), regardless of the actual land cover. RTC uses a Digital Elevation Model (DEM) to normalize the backscatter intensity, effectively removing these geometric and radiometric distortions induced by topography. This correction is vital for ensuring that the processed SAR data accurately reflects the physical properties of the surface and is consistent across different scenes and times.

Addressing the necessity of the above preprocessing steps (Mullissa et al., 2021) proposed a comprehensive SAR pre-processing framework and a library to automate the preprocessing directly through GEE python API. For the preprocessing of Sentinel 1 data this pipeline is adapted.

Sentinel 2 Data Preprocessing

Cloud Masking (Sentinel-2)

Cloud masking is a crucial preprocessing step for optical satellite imagery, such as that from Sentinel-2. Clouds and their shadows obscure the ground, leading to incorrect spectral signatures and misleading analysis results. There are various cloud masking techniques in the literature which falls into 2 major categories, multi-temporal and mono-temporal approaches. Multi-temporal methods like MAJA and Tmask compare the current image with previous, cloud-free scenes of the same location. Although highly effective, their requirement for auxiliary historical data limits their applicability in real-time scenarios. On the other hand, Mono-temporal Approaches require only the single image, offering reduced complexity, and are subdivided into:

Physical Rule-Based Methods: Tools such as Sen2Cor, Fmask, and QA60 use fixed thresholds on spectral bands and indices for classification. While computationally efficient, they are typically less accurate and struggle to differentiate between thick and thin clouds.

Machine Learning (ML) Approaches: These treat masking as a pixel classification problem, learning relationships from training data. An example is s2cloudless, a computationally efficient pixel-based ML method that, in its common form, does not identify cloud shadows or distinguish between thick and thin clouds (Skakun et al., 2022; Wright et al., 2024). GEE also provides quality assurance bands (like the SCL—Scene Classification Layer) provided with the imagery to identify the cloud shadows which can be integrated with the s2cloudless masking approach to efficiently remove cloud and cloud shadows((Sentinel-2 Cloud Masking with S2cloudless | Google Earth Engine, n.d.)).

Given the need for thousands of images to train the deep learning model, GEE's capability for handling massive datasets is essential. Once the Sentinel 1 and Sentinel bands data has been pre-processed, those will be downloaded to the local storage and then further processed on the local infrastructure. One can run the script 0_0_2Band_TrainData_Download-S1.py, and 0_1_2Band_TrainData_Download-S2.py respectively located in the folder ./1_Preparation/ Before running the python scripts, please follow the instructions below to change the variables.

Data Download

1. Change the variables: Change the following variables in 0_0_2Band_TrainData_Download-S1.py:

   • GEE_SERVICE_ACC: replace the current value to your service account id, i.e.,
   • GEE_CREDS: replace the current with he path of your private key.
   • paths: dictionary of paths where crop names are the keys and the path to the geopackage files are the values to corresponding keys

2. Run the script: Run the Script by running the following command-
   python 0_0_2Band_TrainData_Download-S1.py
   This will download the pre-processed sentinel 1 imageries

3. Open the file 0_1_2Band_TrainData_Download-S2.py and change the following paramters accordingly.

   • GEE_SERVICE_ACC: replace the current value to your service account id, i.e.,
   • GEE_CREDS: replace the current with he path of your private key.
   • paths: dictionary of paths where crop names are the keys and the path to the geopackage files are the values to corresponding keys

4. Run the script: Run the Script by running the following command-
   python 0_1_2Band_TrainData_Download-S2.py
   This will download the pre-processed sentinel 2 imageries