Joaquin-Gajardo/nigeria-crop-mask

Code for: Evaluating the role of training data origin for country-scale cropland mapping in data-scarce regions: A case study of Nigeria # :earth_africa: Nigeria Crop Mask :seedling: This repository contains the code and data supporting Gajardo et. al (2025), for training a pixel-wise LSTM binary classifier to predict cropland vs non-cropland from remote sensing data and using it to generate two cropland maps for Nigeria for the year 2020. The code is largely based on the work of Kerner et. al (2020) from NASA Harvest, who build a similar cropland mask for Togo. Links: - :pencil: **Paper:** Gajardo et. al (2025), *Evaluating the Role of Training Data Origin for Country-Scale Cropland Mapping in Data-Scarce Regions: A Case Study of Nigeria.* - :pencil: **Preprint:** Gajardo et. al (2025), *Evaluating the Role of Training Data Origin for Country-Scale Cropland Mapping in Data-Scarce Regions: A Case Study of Nigeria.* - :clapper:**Demo:** a Google Earth Engine (GEE) app to interactively visualize the output maps map and compare them to the ESA WorldCover 2020 land cover map, size-by-side. - :open_file_folder:**New Nigeria cropland dataset:** Google Drive link with the new hand-labelled binary Nigeria cropland dataset (labelled by one expert via remote photointerpretation) and the respective `hd5f` data arrays. - :eyes: **Maps:** Google Drive link with the final cropland and cropland probability maps for Nigeria 2020. ## :open_file_folder: Data The data used to train the LSTM model combines a new hand-labelled dataset of crop and non-crop labels distributed throughout Nigeria (figure below) with a subset of the global Geowiki cropland dataset to predict the presence of cropland in a pixel time series. The pixels time series consists of 12 monthly composites of remote sensing data at 10 m resolution, including Sentinel-1 and Sentinel-2 satellite images, as well as meteorological and topographic data. The training and inference data is processed using the CropHarvest Python package. ## :hammer: Setup The code was developed and tested on a Linux-based workstation using Python 3.7. For setting up …