A Training Sample Migration Method for Classification of Vegetation Cover Using LSTM Model and Remote Sensing Data (Case Study: Ravansar County, Kermanshah Province, Iran)

Abstract

Improving the quality of land-use change classification, with an emphasis on rangeland vegetation, by using time-series satellite imagery and migration data for training can lead to proper management of natural lands. Rangelands, as one of the most important ecosystems on Earth, play a crucial role in maintaining ecological balance and ensuring food security. Training data (reference data) are a fundamental factor in determining the quality of information generated by machine learning methods, as they serve as the basis for future decisions. Using training data migration methods helps reduce data collection costs and enables timely access to reliable training data. This study aims to improve classification quality by applying a training data transfer technique based on an artificial intelligence algorithm, Sentinel-2 satellite image time-series data, and distance-based criteria. The study area is located in Ravansar, Kermanshah province, in the west of Iran. Ravansar has various biological regions involving urban areas, agricultural farms, rangelands and geological formations. To this end, migration data samples from 2017 to 2021 were determined to promote the accuracy of vegetation cover classification using three methods include Euclidean Distance (ED), Spectral Angle Distance (SAD) and Dynamic Time Warping (DTW). This research indicated that the vegetation plant classes require more time data, while other classes need no time series. In addition, using the remote sensing data of the Sentinel-2 sensor for the migration of training samples indicated the efficiency of the proposed methods for producing up-to-date maps without extensive land surveys. The proposed algorithm, using DTW, SAD, and ED flexible distance measures, Sentinel-2 image time series data, and providing optimal thresholds for different feature spaces, achieved an accuracy of up to 76%. This capability, the ability to extract useful information from multi-temporal data of the Sentinel-2 sensor without the need to use the entire time interval, is an important advantage. This method provides valuable insights into changes and dynamics of plant cover, as well as the identification and classification of vegetation, which are essential for the optimal management of natural resources and food security. The final map, incorporating plant cover and other land-use classes, serves as a practical tool for monitoring annual vegetation changes and assessing ecological shifts, thereby supporting policymakers and researchers in strategic planning and management.