Skip to content

Land Cover Prediction from Satellite Imagery Using Machine Learning Techniques

Notifications You must be signed in to change notification settings

iremozcann/Land-Cover-Prediction-Using-Machine-Learning

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Land-Cover-Prediction-Using-Machine-Learning

Climate change has become a serious threat to humanity. Global land cover maps serve as an important means to tackle this problem. ESA has produced Global Land Cover Map from Sentinel-2 data. In this project, land cover maps were generated for unseen data by using a limited number of training samples. Gibraltar is selected as a study area as there are wide variety of land cover types.

alt text

About Data Files

  • Original data files are TIFF image files. They are in 2D spatial files. Numpy and CSV files are converted to suitable formats for SKLearn classifiers
  • Train.csv contains both Labels(Code) and Data. Test.csv contains only test data.
  • Data is UInt16, and Classlabels are Uint8.
  • Training data: S2A_MSIL1C_20220516_TrainingData.tif
  • Training labels: S2A_MSIL1C_20220516_Train_GT.tif
  • Test: S2B_MSIL1C_20220528_Test.tif
  • Class List:

alt text

Reference: https://esa-worldcover.s3.eu-central-1.amazonaws.com/v200/2021/docs/WorldCover_PUM_V2.0.pdf

Note: If you open CSV files with Excel It will truncate data exceeding 1 million'th row.

Land Cover Map from Sentinel-2 data

Each Sentinel-2 satellite is equipped with a single multi-spectral instrument (MSI) that has 13 spectral channels in the visible/near infrared (VNIR) and short-wave infrared (SWIR) spectrums. Continued collaboration with the SPOT-5 and Landsat-8 missions is possible inside the 13 bands thanks to the 10-meter spatial resolution, with a primary goal of land classification.

Machine Learning Methods

The objective of the project is to estimate the land-cover changes by employing machine learning techniques.In this study, various machine learning techniques such as Random Forest, K- Nearest Neighbor ( KNN) has been used for land cover prediction from satellite imagery. Furthermore, user-defined hyperparameters are optimized to obtain higher accuracy results. The Grid Search method was used to find the best parameters of each algorithm. Additionally, the k-fold cross-validation approach (k = 10, iteration = 10) is employed to obtain more precise and objective results.

References

  • Das, Tapan Kumar, Dillip Kumar Barik, and KVG Raj Kumar. "Land-Use Land-Cover Prediction from Satellite Images using Machine Learning Techniques." 2022 International Conference on Machine Learning, Big Data, Cloud and Parallel Computing (COM-IT-CON). Vol. 1. IEEE, 2022.
  • https://esa-worldcover.s3.eu-central-1.amazonaws.com/v200/2021/docs/WorldCover_PUM_V2.0.pdf
  • Castro Gomez, M. G. (2017). Joint use of Sentinel-1 and Sentinel-2 for land cover classification: A machine learning approach. Lund University GEM thesis series.

Releases

No releases published

Packages

No packages published