Abstract

STRAND: New sources and tools for effectiveness: AI and Machine Learning in Statistics

TITLE: Application of Deep Learning for Land Cover Classification using Freely available Satellite images
Accurate and up-to-date land cover information is essential for agricultural planning, environmental monitoring, and evidence-based policy making in Rwanda, where land resources are limited and rapidly changing. Traditional land cover mapping approaches based on field surveys and manual image interpretation are costly, time-consuming, and updated infrequently, often resulting in outdated datasets. This project addresses these limitations by developing an automated, end-to-end land cover classification system that combines deep learning techniques with freely available Earth observation data.
This study presents the development and implementation of a deep learning model for land cover classification using freely available Sentinel-2 and Sentinel-1 satellite images from Google Earth engine.
The main aim of this project was to develop an automated pipeline for land cover classification and produce relevant statistics using freely available satellite images by applying deep learning models. This will help in timely planning and monitoring and more frequent agriculture survey sampling frame revision using updated land cover maps. Moreover, this will reduce time and costs aligned with this task of manually upgrading.
Multispectral Sentinel-2 imagery and ESA WorldCover reference data were used to generate high-resolution land cover maps at 10 m spatial resolution. A comprehensive preprocessing pipeline was implemented, including spatial harmonization, cloud masking, vegetation index computation, tiling, stratified sampling, and normalization. The study evaluated multiple deep learning semantic segmentation architectures, U-Net, U-Net++, DeepLabV3+, PSPNet, and MAnet—using pretrained encoder backbones and transfer learning. Model performance was assessed using pixel accuracy, macro F1-score, and mean Intersection over Union (mIoU).
Among the tested models, U-Net with a ResNet-50 backbone achieved the best overall performance, demonstrating reliable classification of major land cover classes such as cropland, forest, built-up areas, and water bodies. The selected model was operationalized through a user-friendly web application that enables non-technical users to select a district and year, automatically triggering satellite data retrieval, prediction, visualization, and land cover statistics generation.
The results demonstrate the feasibility of integrating deep learning and satellite imagery into an operational land cover monitoring system. The proposed solution supports more frequent updates of land cover maps, reduces time and costs, and provides a scalable foundation for future enhancements, including multi-temporal analysis, cloud deployment, and validation using field data.