Search

This work deals with the evaluation of the potential of multi-temporal and multi-sensor (optical and radar) images of the satellites Sentinel-1 and Sentinel-2 satellites for land cover mapping and coastline evolution monitoring in a tropical ecosystem over the period 2015 to 2020. The study area chosen is the Cameroon Estuary. This area represents a major ecological transition in the sub-region with the presence of protected natural reserves (mangroves, dense forest, wetlands) but also a strong anthropic activity (construction, agriculture, degraded forest). The methodological approach consisted of a chain of pre-processing and visual image analysis, followed by a combination of the bands from each sensor, a supervised Random Forest classification to draft a land use map and finally a digitization of the coastline. Overall, the results show that the classification with the Sentinel- 2 images using 10 bands and adding 4 vegetation indices is slightly more accurate (95.75 %) than the classification with the 13 initial bands (91.78 %). The classification performed with the Sentinel-1A multi-temporal image series (dual polarization (VV, VH) and VV/VH ratio) leads to an accuracy of 78.44 %. The combination of Sentinel-2 A&B bands (10 bands and 4 indices) and Sentinel-1 improves the results and leads to an accuracy of 98.76 %. Our results also show that the use of multi-temporal image time series significantly improves the classification accuracy compared to the use of a single image (mono-date), and this for both sensors, i.e. an additional gain of 13 % and 10 % respectively for Sentinel-2 and Sentinel-1. However, this gain remains small for temporally stable classes. The results of the coastline evolution analysis show that the Cameroon estuary is disturbed by different levels of erosion (Cape Cameroon, North Manoka Island, Sanaga River mouth), and accretion (Limbe, and South Manoka Island) and also by weak internal variations. The fusion of optical and radar remote sensing data in the discrimination of land use classes, has shown that the construction zones are the most vulnerable to coastal erosion. On the other hand, the presence of vegetation (mangrove, forest) stabilizes and protects the coast from possible risks of this nature. [ABSTRACT FROM AUTHOR]

In this article, we analyze the possibility to improve the classification of crops in an irrigated area of northern Morocco based on the combination of the multi-temporal data of two satellites (Sentinel-1 and Sentinel-2) with the inclusion of nine indices. The area concerned (Loukkos), in addition to its strategic position, is characterized by a Mediterranean climate with a strong ecological value. It presents an intense agricultural activity with a great diversity of cultures whose functioning could be affected by climate change. To quantify water requirements, we used the series of satellite images acquired during 09/2017 to 08/2018. Maps produced for three classification levels illustrate our approach. The study showed that the 10 optical channels, at 10 and 20 m spatial resolution, of the data acquired by Sentinel-2 provide good performance, with a kappa index > 85% for all (sub)classes and an overall accuracy > 86%. These performances are superior to those obtained with radar data acquired by Sentinel-1, with F-score deviations of 9% lower on average, and up to 29% (on cork oak/SSC level). Neither the addition of optical radiometric indices, nor the optical and radar combination data provide a significant improvement in performance obtained with Sentinel-2 data. In order to exploit the data obtained, future work would focus on studying the temporal profiles of each type of crop. [ABSTRACT FROM AUTHOR]

In remote sensing, there is a large number of algorithms to classify a satellite image. Among these classification algorithms, the Random Forest appears to be particularly powerful. The objectives of this study are to evaluate (1) the importance of image selection for the level of accuracy of the training model and (2) the nature of the relationship between the level of accuracy of the model and the overall accuracy of the thematic map resulting from the classification of the satellite image with this algorithm. Based on a Landsat 8 OLI image taken over a tropical mountain area : the West Cameroon region, 35 models were built and tested. Analysis shows that the level of overall accuracy of the Random Forest results are closely dependent on the accuracy of the training model used to classify the satellite image, and on the choice of the images used to train this model. Moreover, the selection of these images is itself dependent on the quality of the Regions of interest (ROI) that will be used to build the model. It is therefore important to place particular emphasis on the quality of the input data in order to guarantee satisfactory results with this algorithm. Otherwise, the performance of this algorithm could be inferior to that of other classification algorithms. [ABSTRACT FROM AUTHOR]

Monitoring forest state is required to estimate the levels of greenhouse gases emissions from deforestation. The objective of this study was to estimate deforestation trends for 2005-2010 and 2010-2013 periods in the eastern rainforests ecoregion of Madagascar. A supervised classification approach of multi-date Landsat images using the Random Forest algorithm was used. A complete 30 m map of the forest cover change was produced on 20.5 million hectares. The result was validated by photo interpretation of nearly 11 000 points from a regular sampling grid on the basis of SPOTand Landsat satellite images acquired in 2010-2013. The overall mapping accuracy was estimated to be about 90%. The area of Madagascar rainforests has been estimated at 4.5 million hectares in 2005, 4.4 million hectares in 2010 and 4.3 million hectares in 2013. The annual rate of deforestation has increased from 0.5%. to 0.9 % between the two periods. This study could be considered as a reliable cost-effective and reproductible approach for monitoring forests in Madagascar. [ABSTRACT FROM AUTHOR]