Multi-Decadal Remote Sensing of Land Cover and Surface Water Dynamics in a Data-Limited Monsoon Lake Watershed in Myanmar
Year: 2026
Type: article
Abstract: Long-term monitoring of land use/land cover (LULC) and surface water is challenging in monsoon lake watersheds with limited reference observations. This study constructed annual LULC in the Inle lake watershed, Myanmar, for 2000–2025 using Landsat annual median composites, six reflectance bands and six spectral indices. Random Forest, Support Vector Machine (SVM) and XGBoost were evaluated using spatially grouped internal testing of purposively selected reference samples. SVM achieved the highest 26-year mean internal overall accuracy (71.49%), kappa (0.620) and macro-F1 (71.79%). A separate stratified probability sample of 600 fixed pixels was assessed for 2000, 2010, 2020 and 2025 to estimate map accuracy, class areas and paired net changes within a common 5438.20 km2 domain. Area-weighted overall accuracy ranged from 58.7% to 71.4%, while Built-up user’s accuracy remained low at 12.4–22.9%. Reference-based 2000–2025 changes were Agriculture +450.1 km2 (95% sampling interval: +203.4 to +696.8), Forest −631.5 km2 (−857.3 to −405.7) and Water +42.3 km2 (−23.5 to +108.0). The Water interval included zero, and the reference assessment did not establish a 2020–2025 recovery. Annual trend significance for Agriculture and Built-up changed under assumed classification error correction scenarios. Within the Landsat 8-only period (2013–2025), mapped Forest still declined, whereas mapped Agriculture showed no monotonic trend. Within the fixed open water outline of Inle lake, mapped water declined slowly but consistently (−0.07 km2 yr−1; p < 0.001), the Moebye reservoir showed no trend, and most of the mapped decline occurred in the surrounding sanctuary wetlands. Most of the reference-based agricultural increase and forest loss occurred between 2000 and 2010. The results support agricultural expansion and forest loss but leave watershed-wide surface water change uncertain. Distinguishing raw mapped trajectories from reference-based estimates and sampling uncertainty is essential for interpreting long-term watershed change.
Source: Remote Sensing
DOI: 10.3390/rs18213837
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,258
OpenAlex ID: W7225635452