ALFICC - Active Learning For Improved landCover Classification

This repository is intended to share some results and materials from my Master's dissertation project. These deliverables are intended for open access. Everyone is free to use them for research purposes, as long as this work and the resources' origin are properly cited.

Introduction

The goal was to use Sentinel-2 Surface Reflectance geospatial data in order to identify cashew orchards with Machine Learning (ML) tools across a region of interest (ROI) comprised of two parts: the country of Guinea-Bissau, and the neighboring region of the Republic of Guinea up to the Kogon river. Active Learning techniques were employed to develop an optimized training dataset from samples scattered across the ROI. With this data, we created land cover maps based on ML predictions for the entire ROI.

The labels were acquired for the year of 2021. For more details about the project, the data and application, the full dissertation can be found in the UPorto open repository

Contents

The contents shared in this repository regard the following:

• The two datasets of labeled samples that were created during the project's execution.
• Cashew land cover maps obtained with ML model predictions.

Datasets

Two new datasets of labeled pixels scattered across the ROI were created during the project's execution. Namely:

• Random Sampling dataset: corresponds to $4498$ samples randomly selected in the ROI.
• Margin Sampling dataset: $1816$ points selected by the margin sampling heuristic. This dataset shows a smaller geographical distribution and sample diversity than the Random Sampling one, but demonstrated its usefulness for ML purposes.

The labeling process included eight different classes, from 1 to 8: Closed Forest, Open Forest, Mangrove, Savanna (sparse vegetation), Cashew, Non-Forest, Water, and Non-Cashew. We note that class 8 - Non-Cashew contains pixels that are evidently not cashew orchards, but either contain a mixture of the other six classes (for example, Savanna and Non-Forest), or correspond to samples that could not be classified as any of the aforementioned categories. Therefore, class 8 samples are very useful for a binary classification setting: when trying to simply distinguish between cashew and non-cashew classes, they can be part of a broader Non-Cashew class, which also includes Forest, Mangrove, and so on. However, in other scenarios, these samples should be discarded from further consideration.

The .csv files include information about the pixels' coordinates in the EPSG:4326 projection and category following the aforementioned system. The pixels were acquired in $3 \times 3$ polygons or patches, with each patch being assigned a different number, which is represented in column polygonID - this can be important to analyze spatial distribution and correlation of the acquired samples. Besides these, the first 67 columns of randomsampling.csv and marginsampling.csv regard the training features that were considered for ML tasks. They correspond to a subset of a group of features formed by the spectral bands directly obtained from Sentinel-2 measurements, spatial features obtained with the GLCM method, and temporal CCDC coefficients. normalizers.csv stores the min, max, mean and std values for each of the 67 features, for feature normalization purposes.

Land cover maps

roi.geojson is a geodataframe that stores information about the two entries that constitute the ROI - Guinea-Bissau and the Republic of Guinea section. Multiple land cover maps were obtained during the project's execution, with two being shared in this repository:

• initial_map_7class: This map was obtained in earlier stages of the project's pipeline. It was created using the predictions of an ML classifier, which predicted the class of each pixel in the region, only between class 1 and class 7 - class 8 (Non-Cashew) was discarded for this setting, as we weren't working with a binary classification approach. We also note the base data where the ML model was trained was a proprietary dataset, shared by RSeT, and not one of the datasets obtained during the dissertation.
• final_map: This was the last map to be created, and was formed by ML predictions of a model trained on the entire margin sampling dataset. This algorithm only tried to identify cashew and non-cashew, so class 8 was added for this setting.

Map accuracy was determined by evaluating the true label of test pixels and comparing them to their representation on the land cover maps. The initial map showed 82.2% balanced accuracy and 88.4% Cashew F1-score, while the final map results in 94.0% balanced accuracy and 89.5% Cashew F1-score. Note that the test sets used to evaluate map accuracy were not the same for both maps: a test set for the first map was created based on approximately 20% of the RSeT proprietary dataset, while the final map's test set was roughly 50% of the random sampling dataset.

Programming requirements

Scripts have been developed using JavaScript (in Google Earth Engine's code editor) and Python.

• For the Python scripts, the following packages were used:
  • numpy
  • pandas
  • sklearn
  • geopandas
  • rasterio
  • rioxarray

---

Reference:

Pereira, M. (2024). Active Learning for Improved Landcover Classification [Master's thesis, Faculty of Sciences - University of Porto]. Repositório Aberto UPorto. https://hdl.handle.net/10216/164196

• Author: Miguel Ribeiro Pereira, MSc, Faculty of Sciences, University of Porto, Portugal
• Supervisors: João Pedro Pedroso, CMUP - University of Porto, Portugal
• Co-Supervisor: Sofia Cardoso Pereira, PhD, INESC TEC and Faculty of Engineering - University of Porto, Portugal