Abstract
Higher overall accuracy does not necessarily mean more complete landslide mapping at every location. Across eight rainfall-related events, we examine the benefits of earlier satellite imagery using U-TAE, a neural network that combines images acquired on different dates, as the main model. We compare overall accuracy with the fraction of mapped landslide area detected within local image patches. Mean average precision increased from 0.333 with a post-event image to 0.379 with a pre/post pair and 0.419 with a longer sequence. Adding earlier images to the pair improved mean average precision in seven events but slightly reduced it in Chimanimani. However, the average increase in the detected fraction across locations and regions was 1.1 percentage points (95% interval −1.3 to +3.3). This provides no clear evidence that earlier images increased the average fraction of mapped landslide area detected. The estimated gain in Italy decreased from 3.0 to 0.3 percentage points when the decision thresholds changed, indicating that the apparent improvement depended on how the predictions were converted into a landslide map. The benefit of earlier images was also greater in patches containing more mapped landslide area in some regions, but was not clearly related to the strength of surface change in patches with sufficient mapped area and usable optical observations. An auxiliary comparison with U-Net, a convolutional image-segmentation model, also showed local detection differences alongside higher overall accuracy. Local detection also varied when the same models were retrained using unchanged images. Checking local detection changes alongside conventional accuracy measures helps assess whether adding earlier imagery reduces missed landslide area across locations.