What data fusion adds

The Swiss land-use statistics describe more than 4.1 million sample points on a regular 100-metre grid. Land use captures a site's function; land cover describes its visible surface. An aerial image alone cannot always resolve that distinction: a lowland meadow and an alpine pasture can look similar while belonging to different land-use classes.

The research prototype therefore combines two processing stages. Convolutional neural networks analyse RGB and false-colour infrared aerial images. A random forest fuses their class probabilities with Landsat time series, terrain elevation, canopy height and cadastral information. The time series contribute seasonal vegetation patterns, while the spatial layers help interpret the decisive sample point.

Evidence [1]

Table of SWISSIMAGE aerial datasets with RGB and infrared examples, followed by three image tiles with marked reference points.
Aerial imagery for ADELE: RGB and false-color infrared SWISSIMAGE data across acquisition periods and resolutions, with 50 × 50 m image tiles centered on Swiss land-use statistics reference points. The examples illustrate the distinction between land cover and land use. Meyer et al. (2025), Fig. 2. Image and source ↗

Results and practical value

On the representative evaluation dataset, the full fusion reaches 83% overall accuracy for land use and 86% for land cover. The corresponding RGB image models reach 53% and 72%. This comparison demonstrates the benefit of complementary data for this task. The findings were published in the European Journal of Remote Sensing in 2025.

Overall accuracy conceals differences between common and rare classes. The paper therefore also reports class-level F1 scores and discusses small, heterogeneous or visually similar categories. In the reported ADELE deployment, only sufficiently confident no-change cases are accepted automatically. The 98% precision target applies to this selected subset, rather than every class and sample point.

Evidence [1]

Bar charts comparing CNN and random-forest F1 scores for common land-use and land-cover classes, with contributions of the input data groups.
What data fusion contributes: F1 scores for the five most common land-use and land-cover classes, comparing CNN variants with the full random-forest model. Below: estimated contributions of the input data to random-forest classification. Meyer et al. (2025), Fig. 8. Image and source ↗
Methods in detail
  1. Classify RGB and false-colour infrared aerial imagery with neural networks.
  2. Fuse class probabilities with Landsat time series, elevation, canopy and cadastral data in a random forest.
  3. Evaluate overall accuracy and class-level F1 scores on a representative dataset.

Publications and sources

  1. PublishedMeyer et al. (2025), European Journal of Remote Sensing