Detection and mapping of agriculture seasonal variations with deep learning–based change detection using Sentinel-2 data

• Published in: Arabian journal of geosciences 2022, Vol.15 (9), Article 825
• Main Authors: Singh, Gurwinder, Singh, Sartajvir, Sethi, Ganesh Kumar, Sood, Vishakha
• Format: Article
• Language: English
• Subjects: Accuracy; Agricultural land; Algorithms; Artificial neural networks; Cereal crops; Change detection; Classification; Climatology; Cultivation; Datasets; Deep learning; Detection; Earth and Environmental Science; Earth science; Earth Sciences; Feature extraction; Grain cultivation; Imagery; Land acquisition; Land use; Machine learning; Neural networks; Original Paper; Remote sensing; Satellite imagery; Seasonal variation; Seasonal variations; Seasons; Spaceborne remote sensing; Spring (season); Support vector machines; Wheat
• ISSN: 1866-7511; 1866-7538
• DOI: 10.1007/s12517-022-10105-6

Change detection is one of the vital ways to analyse the multitemporal variations over a specified period using remote sensing data. In recent years, deep learning (DL) algorithms have become the choice of many remote sensing researchers to solve the problems of conventional change detection methods and to improve their accuracy. In the present work, the DL classifier has been incorporated with the post-classification comparison (PCC), named DL-based change detection (DLCD), to extract the features from satellite imagery based on their spatial and spectral properties and detect the seasonal variability. For demonstration purposes, the dataset has been acquired over the agricultural land in Punjab State, India, using the Sentinel-2 optical dataset during the period 2017–2018. Due to the climatology of Punjab, this region is well-suited for wheat cultivation. Therefore, we have computed the change maps for the rabi seasonal crop (wheat) which is planted usually in October and grows throughout the winter season to be harvested in the spring season (April). To confirm the effectiveness of the proposed approach, the performance of DLCD has been cross-validated with random forest (RF)–based PCC, convolutional neural network (CNN)–based PCC and support vector machine (SVM)–based PCC. Experiential outcomes have shown that DLCD achieved a higher accuracy (94.8–97.2% in classified maps and 91.8–95% in change maps) as compared to the RF-PCC (87.6–90.2% in classification and 88–89.4% in change maps), CNN-PCC (90.4–93.4% in classified maps and 87.4–90% in change maps) and SVM-PCC (86–88.8% in classified maps and 86–88.8% in change maps). This study can be significant in terms of extraction of various crop types, water surfaces and manmade features, as well as various land-use patterns using DLCD.