Determination of land use and land cover change using multi-temporal PlanetScope images and deep learning CNN model

Abstract Land use and land cover (LULC), an important component of Earth observation technology, plays a crucial role in image classification. Detailed LULC information is of great importance for monitoring agriculture and wetlands, urban and rural planning, land resource management, and the monitoring and detection of climate change-induced changes. To create a detailed LULC, high-resolution satellite imagery, ground truth data, and a model that provides high accuracy are required. For this reason, 8-band PlanetScope images with 3 m spatial resolution, Land Parcel Identification System (LPIS) physical blocks (1:5000 scale), and a convolutional neural network (CNN) model were used as ground truth data. While using the CNN model, hyperparameter optimization was performed to determine the most suitable parameters for achieving high classification accuracy. As a result of the classification, the overall accuracy value was 93.14%, while the accuracy value for each class (arable land, artificial surface, forest, grassland, shrubland, tree crops, and water) was around 90%. Although this study focuses on the use of CNN for LULC mapping, one of its main objectives is to create a high-resolution LULC map to serve as a basis for updating LPIS, a reference system for the management and control of support payments, which is one of the main components of the Integrated Administration and Control System (IACS). LPIS is generally required to be updated every five years according to European Union legislation. As a result of this study, the LULC classification (2023) and LPIS physical blocks (2015) were compared, and areas of change as well as areas needing updates were identified. This study introduces an innovative approach by integrating high-resolution, multi-temporal PlanetScope imagery with LPIS data using a deep learning-based CNN model for LULC classification. Unlike previous studies that primarily rely on medium-resolution imagery or traditional machine learning methods, this study leverages the spectral and temporal advantages of PlanetScope’s 8-band imagery to enhance classification accuracy. Moreover, this research provides a novel methodology for LPIS updating, offering a scalable and automated framework that can be applied at a national scale. By combining CNN-based classification with LPIS change detection, this study establishes a systematic approach for identifying areas requiring updates, contributing to more efficient land resource management and agricultural policy implementation.