Seasonal forest disturbance mapping using Sentinel-1/2 data

By Clara Siegert

Abstract
The clearing and degradation of forest is a process of global importance, that affects forests in their ecological, economical and social services to mankind. Disturbances such as wildfires, storms, droughts or excessive exploitation can ultimately lead to complete forest loss. Therefore, early detection of negative impacts is critical to protect and manage the fast shrinking forest cover. Satellite imagery time series offer a great potential to detect such disturbances. Looking in the past with the undisturbed condition as reference allows to detect impacts timely. However, the detection of disturbances is particularly difficult in deciduous forests with pronounced seasonality, because the shedding of the leaves can be confused with degradation.

The objective of this work was to determine the extent to which deseasonalisation through harmonic model fitting can improve the detection of disturbances. Specifically, it was investigated whether deseasonalisation works better with Sentinel-1 (S1) or Sentinel-2 (S2) data, and whether it performs better in forests with lower seasonality (tropical dry forest in Bolivia) than in forests with more pronounced seasonality (temperate deciduous forest in Romania).

A seasonal model was fitted over a 3-year period from 2016 to 2018 to deseasonalise 2019 and 2020. A 2019 wet/summer season image was used to collect training data for the forest and disturbance classes, which then were used to train different Random Forest (RF) models. Training data was collected from the original and deseasonalised data from both satellites to train two different models in each region. The RF models were tested and validated in one image during the dry/winter season and one during the wet/summer season.

Disturbance detection improved significantly with deseasonalised S2 images in both forest ecosystems. However, misclassifications still occurred in the dry/winter season because too few satellite images were available to fit an adequate harmonic model for the period 2016 - 2017, especially in the temperate deciduous forest.

More observations were available from the cloud penetrating S1 satellite, but despite that deseasonalisation did not significantly improve the detection of disturbances in either forest ecosystem. Fitting of the harmonic model was difficult because changing environmental conditions such as the amount of precipitation and moisture content in vegetation and soils resulted in large erratic fluctuations in the S1 time series signal. Because of this, only large scale clearcut disturbances could be detected by S1, while forest degradation could not be detected.

It was found that deseasonalisation using harmonic model fitting can significantly improve the detection of disturbances in tropical dry forests and temperate deciduous forests with S2 imagery. Forests with lower seasonality showed slightly better deseasonalisation results than forests with pronounced seasonality. Deseasonalised S1 data can also improve disturbance detection, but pre-processing and a longer time series is needed for harmonic model fitting to eliminate signal variability due to changing environmental conditions.