Mediterranean basin vegetation forecasting approaches: accuracy analysis & climate-land cover-sensor nexus impacts

Abstract

ABSTRACTFrom land degradation and desertification to cyclones and tropical storms, and so on, the repercussions of global change have become increasingly severe in recent years. Such environmental impacts require continuous assessment and monitoring. Thus, to study and analyse these impacts, a variety of time-series forecasting approaches have been developed, including statistical ones [i.e. Moving Average (MA), Auto-Regressive Integrated Moving Average (ARIMA), etc.], and machine-learning approaches such as Recurrent Neural Networks (RNN), Long-Short-Term Memory Network (LSTM) and Convolutional Neural Networks (CNN). In this study, accuracy of the most used forecasting approaches (i.e. MA, LSTM and Conv-LSTM) has been quantified, and three impacts (i.e. climate regions, land cover and satellite sensors) have been brought to light. Firstly, eight Mediterranean regions were selected (i.e. two hot arid regions, two cold arid regions, two regions with temperate hot summers and two regions with temperate warm summers) based on Köppen climate classification. Secondly, 654 hyperspectral images retrieved from three different satellites (i.e. Sentinel-2, Landsat-8 and MODIS) from 2016 to 2022 were computed in order to predict 29 vegetation biophysical indices (i.e. NDVI, GNDVI, EVI, CVI, etc.). Accordingly, more than 18,000 images were computed, resulting in 696 time series forecasted using the aforementioned approaches. Finally, 2088 forecasted time series have been determined, and their accuracy has been compared. As a result, Landsat-8 and Sentinel-2 images had the highest forecasting accuracy in the three approaches, reaching 86% of the computed indices being over 50% accurately predicted in LSTM model, while MODIS data had the highest forecasting accuracy only in MA model, with a percentage of 72% of the computed indices being over 50% accurately predicted. Furthermore, we have identified that region climate impacts vegetation forecasting accuracy. For instance, arid regions showed low accuracies across all models, while temperate regions showed higher accuracies in the Mediterranean region.KEYWORDS: Mediterranean basintime seriesstatistical forecastingmachine learning forecastingforecasting accuracyremote sensingclimate impacts Highlights Forecasting of 29 vegetation biophysical indices of three different satellites in eight Mediterranean regions using statistical and ML models.Statistical and ML forecasting models accuracy quantification disclosing three impacts: climate, land cover and satellite sensor.Forecasting using very high-resolution satellite images gives more accurate results than moderate ones.AcknowledgementsThis work is supported in part by the University Abdelmalek Essaâdi Thematic project, 2019–2022, Monitoring land cover dynamics of the TTA region in the context of global warming: Contribution of Earth's observation satellites and high-performance computing.Disclosure statementNo potential conflict of interest was reported by the authors.