Abstract
In this study, we aim to estimate mountain grassland yield losses due to drought events using Sentinel-2 and Sentinel-1 satellite data. Leaf Area Index (LAI) is used as a proxy for yield, due to its high correlation with aboveground grassland biomass. LAI can be estimated rapidly and accurately from Sentinel-2 multispectral data [1]. However, the frequent unavailability of cloud-free imagery poses a significant challenge, creating spatial-temporal gaps. To address this limitation, there is growing interest in incorporating Sentinel-1 Synthetic Aperture Radar (SAR) backscatter data to perform spatial gap-filling, enhancing the temporal resolution of grassland LAI derived from Sentinel-2 and enabling continuous monitoring of grassland conditions all over the growing season [2, 3]. Addressing all season grassland monitoring challenges in alpine regions using backscatter data requires innovative solutions to integrate domain-specific ancillary information. Alpine grasslands are characterized by complex topography, seasonal variations and heterogeneous agricultural management, making traditional monitoring methods less effective [4]. The use of Sentinel-1 SAR data, with its ability to penetrate cloud cover and capture surface characteristics regardless of weather conditions, offers an approach to overcome these challenges. However, effective utilization of SAR data requires advanced machine learning methods that can exploit complexity of the backscatter signals in alpine grasslands. Incorporating domain knowledge and multi-source information into machine learning frameworks, makes it possible to enhance the accuracy and scalability of grassland monitoring [5, 4]. In this research, we propose a domain specific self-attention based convolutional network to extract feature from SAR data and its derivatives to capture long-range dependencies in the data and studied its impact in spatial gap-filling for grassland monitoring. We have considered Sentinel-1 and Sentinel-2 data over the Trentino-South Tyrol region, in north-eastern Italy for the year 2023. The effectiveness of the proposed method for predicting LAI has been validated by testing it against Sentinel-2-derived LAI and ground-measured LAI. In the preliminary results, we achieved an R2 ranging from 0.15 to 0.42 depending on the features selection on 40% of the test sets between the predicted LAI and Sentinel-2 LAI, and observed a positive correlation with ground-measured LAI.
Keywords:
References
[1] M. Castelli, G. Peratoner, L. Pasolli, G. Molisse, A. Dovas, G. Sicher, A. Crespi, M. Rossi, M. H. Alasawedah, E. Soini, et al., “Insuring alpine grasslands against drought-related yield losses using sentinel-2 satellite data,” Remote Sensing, vol. 15, no. 14, p. 3542, 2023.
[2] I. Tsardanidis, A. Koukos, V. Sitokonstantinou, T. Drivas, and C. Kontoes, “Cloud gap-filling with deep learning for improved grassland monitoring,” arXiv preprint arXiv:2403.09554, 2024.
[3] R. Cresson, N. Nar¸con, R. Gaetano, A. Dupuis, Y. Tanguy, S. May, and B. Commandre, “Comparison of convolutional neural networks for cloudy optical images reconstruction from single or multitemporal joint sar and optical images,” arXiv preprintarXiv:2204.00424, 2022.
[4] R. Guo, J. Gao, S. Fu, Y. Xiu, S. Zhang, X. Huang, Q. Feng, and T. Liang, “Estimating aboveground biomass of alpine grassland during the wilting period using in situ hyper- spectral, sentinel-2 and sentinel-1 data,” IEEE Transactions on Geoscience and Remote Sensing, 2023.
[5] E. Chiarito, F. Cigna, G. Cuozzo, G. Fontanelli, A. Mejia Aguilar, S. Paloscia, M. Rossi, E. Santi, D. Tapete, and C. Notarnicola, “Biomass retrieval based on genetic algorithm feature selection and support vector regression in alpine grassland using ground-based hyperspectral and sentinel-1 sar data,” European Journal of Remote Sensing, vol. 54, no. 1, pp. 209–225, 2021.