Selecting appropriate variables for detecting grassland to cropland changes using high resolution satellite data

Grassland is one of the most represented, while at the same time, ecologically endangered, land cover categories in the European Union. In view of the global climate change, detecting its change is growing in importance from both an environmental and a socio-economic point of view. A well-recognised tool for Land Use and Land Cover (LULC) Change Detection (CD), including grassland changes, is Remote Sensing (RS). An important aspect affecting the accuracy of change detection is finding the optimal indicators of LULC changes (i.e., variables). Inappropriately selected variables can produce inaccurate results burdened with a number of uncertainties. The aim of our study is to find the most suitable variables for the detection of grassland to cropland change, based on a pair of high resolution images acquired by the Landsat 8 satellite and from the vector database Land Parcel Identification System (LPIS). In total, 59 variables were used to create models using Generalised Linear Models (GLM), the quality of which was verified through multi-temporal object-based change detection. Satisfactory accuracy for the detection of grassland to cropland change was achieved using all of the statistically identified models. However, a three-variable model can be recommended for practical use, namely by combining the Normalised Difference Vegetation Index (NDVI), Wetness and Fifth components of Tasselled Cap. Increasing number of variables did not significantly improve the accuracy of detection, but rather complicated the interpretation of the results and was less accurate than detection based on the original Landsat 8 images. The results obtained using these three variables are applicable in landscape management, agriculture, subsidy policy, or in updating existing LULC databases. Further research implementing these variables in combination with spatial data obtained by other RS techniques is needed.

Impact of input data (in)accuracy on overestimation of visible area in digital viewshed models

Viewshed analysis is a GIS tool in standard use for more than two decades to perform numerous scientific and practical tasks. The reliability of the resulting viewshed model depends on the computational algorithm and the quality of the input digital surface model (DSM). Although many studies have dealt with improving viewshed algorithms, only a few studies have focused on the effect of the spatial accuracy of input data. Here, we compare simple binary viewshed models based on DSMs having varying levels of detail with viewshed models created using LiDAR DSM. The compared DSMs were calculated as the sums of digital terrain models (DTMs) and layers of forests and buildings with expertly assigned heights. Both elevation data and the visibility obstacle layers were prepared using digital vector maps differing in scale (1:5,000, 1:25,000, and 1:500,000) as well as using a combination of a LiDAR DTM with objects vectorized on an orthophotomap. All analyses were performed for 104 sample locations of 5 km², covering areas from lowlands to mountains and including farmlands as well as afforested landscapes. We worked with two observer point heights, the first (1.8 m) simulating observation by a person standing on the ground and the second (80 m) as observation from high structures such as wind turbines, and with five estimates of forest heights (15, 20, 25, 30, and 35 m). At all height estimations, all of the vector-based DSMs used resulted in overestimations of visible areas considerably greater than those from the LiDAR DSM. In comparison to the effect from input data scale, the effect from object height estimation was shown to be secondary.