Scholarlink

<p>Temporal deep learning models have been evaluated for their efficiency in classifying land-use and land-cover dynamics in Western Bahia, Brazil. Examined through the utilization of MODIS (Moderate Resolution Imaging Spectroradiometer) MOD13Q1 EVI (Enhanced Vegetation Index) time-series data in conjunction with LEM labels. A comprehensive benchmark is introduced that integrates preprocessing, sequence modelling, and evaluation. Five architectures -LSTM (Long Short-Term Memory), Bi-LSTM (Bidirectional LSTM), GRU (Gated Recurrent Unit), Attention- LSTM and a lightweight hybrid Transformer -LSTM framework are systematically trained and evaluated using an 80/10/10 series-level split. Evaluation employs metrics such as accuracy, macro-F1, precision, recall, ROC (Receiver Operating Characteristic),AUC (Area under the curve), confusion matrices, and qualitative error inspection. The reported test results demonstrate an accuracy of 99.0% and a macro-F1 score of 0.97, indicating that errors are primarily found among classes that are phenologically similar and are influenced by class imbalance. This work provides a clear and reproducible baseline for MODIS-based land use and land cover mapping in Western Bahia, facilitating method comparison and practical application. Expected developments encompass the integration of multiple sensors, spatially constrained validation for robust leakage assessment, adaptation across different domains, and refined uncertainty calibration for optimal threshold determination.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Land use and land cover dynamics study in Bahia’s Western Region, Brazil: a remote sensing and deep learning approach

  • 期刊论文

  • 发表日期：

    2025年12月26日

查看全文
  • Himanshu Sahu,
  • Pushpalata Pujari

摘要

Temporal deep learning models have been evaluated for their efficiency in classifying land-use and land-cover dynamics in Western Bahia, Brazil. Examined through the utilization of MODIS (Moderate Resolution Imaging Spectroradiometer) MOD13Q1 EVI (Enhanced Vegetation Index) time-series data in conjunction with LEM labels. A comprehensive benchmark is introduced that integrates preprocessing, sequence modelling, and evaluation. Five architectures -LSTM (Long Short-Term Memory), Bi-LSTM (Bidirectional LSTM), GRU (Gated Recurrent Unit), Attention- LSTM and a lightweight hybrid Transformer -LSTM framework are systematically trained and evaluated using an 80/10/10 series-level split. Evaluation employs metrics such as accuracy, macro-F1, precision, recall, ROC (Receiver Operating Characteristic),AUC (Area under the curve), confusion matrices, and qualitative error inspection. The reported test results demonstrate an accuracy of 99.0% and a macro-F1 score of 0.97, indicating that errors are primarily found among classes that are phenologically similar and are influenced by class imbalance. This work provides a clear and reproducible baseline for MODIS-based land use and land cover mapping in Western Bahia, facilitating method comparison and practical application. Expected developments encompass the integration of multiple sensors, spatially constrained validation for robust leakage assessment, adaptation across different domains, and refined uncertainty calibration for optimal threshold determination.