The influence of texture on canopy height prediction

← Previous research: Thesis on the influence of spectral values

1 Introduction

The global canopy height model (Lang et al., 2023b) predicts canopy heigth using only spectral imagery as input data. The authors claim the model has learned spectral and textural features, while also including geographical coordinates (Lang et al., 2023a). The first aspect has been analysisd in my Thesis, while Vinzenz HD Zerres analysed the role of latitude on prediction. Here the influence of texture is analysed and documented.

Lang, N., Jetz, W., Schindler, K., & Wegner, J. D. (2023b). A high-resolution canopy height model of the Earth [Source Code]. https://github.com/langnico/global-canopy-height-model

Lang, N., Jetz, W., Schindler, K., & Wegner, J. D. (2023a). A high-resolution canopy height model of the earth. Nature Ecology & Evolution, 7(11), 1778–1789. https://doi.org/10.1038/s41559-023-02206-6

2 Texture Manipulation

To isolate the effect of texture on prediction, a systematic manipulation of the input data is performed (See spectral methods). Pixels were shuffled across individual Sentinel-2-Scenes in their X and Y dimension while keeping the spectral bands (Z dimension) aligned. Two separate manipulations were performed and compared, once shuffling the pixels across the whole Scene and once shuffling pixels within a local 512x512 subtile, keeping global context but changing local texture. The manipulation was performed to different degrees ranging from 10 % to 80 % shuffled pixels.

2.1 Global pixel shuffle

The effects of texture manipulation to different degrees are illustrated both on average and for all individual sample tiles in Figure 1. On average there is no meaningful difference with a very weak positive change across all manipulation degrees. Between the sample tiles big differences are visible, with Malaysia having a strong negative response and USA West and Finland showing the highest positive responses.

Figure 1: Average signed difference to original prediction in meters, by tile and average across all tiles combined.

The average initial canopy heights show a very different distribution based on their response after manipulation. Areas showing an increase in prediction have a lower average initial canopy height while negative responses could be observed more in higher initial canopy heights.

Figure 2: Mean canopy height split by direction of response after manipulation. Negative difference in red and positive difference in blue.

2.2 Local pixel shuffle

3 Technical implementation