Tahmi, Sadiq ， Rui, Xiaoping ， Guechi, Imen ， Hannache, Faris ， Wang, Xuege ， Chebika, Housseyn
2026-05-01 ENVIRONMENTAL MODELLING & SOFTWARE 2026 200(卷), null(期), (null页)
Reliable long-term land use and land cover (LULC) datasets remain scarce in semi-arid regions. This study proposes a scalable and transferable deep learning framework that leverages Landsat archives and Deep Neural Networks (DNNs) to generate a continuous annual LULC dataset from 1985 to 2025 in M'sila, Algeria, as a representative semi-arid case study. The framework incorporates an additional refinement step by integrating the Spectral Angle Mapper (SAM) to generate spectrally stable training samples. The results demonstrate that the proposed framework effectively captures complex spectral relationships, enabling the generation of the first temporally consistent annual LULC dataset spanning 40 years for the study area. Moreover, generalization tests conducted on a geographically similar region indicate satisfactory transferability, with some limitations arising from interactions between land cover patterns and varying geographic conditions. These results demonstrate the potential of deep learning with satellite data for generating robust long-term LULC datasets for environmental monitoring.