Effect of Spatial Resolution on Land Cover Mapping in an Agropastoral Area of Niger (Aguie and Mayahi) Using Sentinel-2 and Landsat 8 Imagery within a Random Forest Regression Framework

Monitoring environmental changes over time requires image time series with extensive historical depth. However, high spatial resolution images often lack such depth. Additionally, some remote areas may suffer from either insufficient satellite coverage or a lack of high-resolution or high-quality imagery.  This study aims to investigate the impact of spatial resolution on image classification. Therefore, Landsat 8 and Sentinel-2 images from October to December 2020 were processed and classified using Random Forest regression in Google Earth Engine (GEE).  Training samples were collected from Collect Earth Online (CEO) to train the model. In addition to the spectral bands available, vegetation indices were considered to optimize classification results. The study revealed differences in land cover areas estimated by the two sensors. These differences are statistically significant at p < 0.001, although they are small. The validation results showed that the RMSE from Sentinel-2 is slightly lower than that from Landsat 8. Although small, this difference is significant at p < 0.05. This highlights two key points: (i) that spatial resolution positively influences the accuracy of image classification, especially when dealing with Landsat 8 and Sentinel-2 imagery; and (ii) that the difference between Landsat 8 and Sentinel-2 sensors is not too substantial in the context of a fragmented landscape, since it ranged from 0.03% to 3.94% across land covers. Therefore, Landsat imagery and, by extension, medium-resolution satellite imagery can still yield satisfactory land cover maps, especially in a patchy landscape such as the southeastern part of Niger.

Keywords: Stratified random sampling; Google Earth Engine (GEE); Random Forest; Collect Earth Online (CEO); Niger