ViT-ChangeFormer: A deep learning approach for cropland abandonment detection in lahore, Pakistan using Landsat-8 and Sentinel-2 data

Abstract

Cropland abandonment poses significant environmental, economic, and social challenges globally. As urbanization encroaches on agricultural areas, understanding the dynamics of abandoned croplands and accurately classifying and detecting them are essential for informed sustainable land use and effective policy development. However, traditional methods struggle to identify abandoned croplands due to temporal variability, limited spectral data and challenges in land cover variations. To address these challenges, we introduced an innovative deep learning approach that combines a Vision Transformer (ViT) with ChangeFormer for the classification and change detection of cropland abandonment using Landsat-8 and Sentinel-2 datasets in Lahore, Pakistan. We employed ViT for image classification, enhancing its efficacy through the incorporation of Vegetation Indices (VIs). This integration led to notable improvements in F1 score and Overall Accuracy (OA), elevating them from 86% and 88%to 92% and 95% respectively. Subsequently, ViT-generated classified rasters facilitated in identification of abandoned lands using ChangeFormer model. The direct comparison showcased a significant enhancement in ChangeFormer's performance, with F1 score and OA escalating from 91% and 90% to 97.5% and 96%, respectively. The improvment was particularly evident when testing ChangeFormer with ViT-generated rasters compared to raw imagery for binary change detection. The study identified 32,043 ha of abandoned cropland (14,613 in 2019 and 17,430 in 2024), with 16.35% converted to built-up areas in 2024. Urbanization was the primary driver, followed by conversions to barren land and water bodies. While our approach improves cropland abandonment detection, addressing unavailability of high-resolution imagery, computational costs, and integrating socio-economic and climate factors could enhance its accuracy and effectiveness.