Unsupervised LSTM-AE and Harmonic Models for Improving Timeliness of Forest Logging Notifications With Dense Satellite Image Time Series

The availability of free and open high resolution optical satellite imagery has greatly improved over the past few years, with the opening of the Landsat archives in 2008 and the launch of Sentinel-2 satellite in 2015. High resolution satellite imagery is widely used to monitor deforestation and forest degradation, evaluate carbon stocks worldwide, or map countries or continents. Satellite Image Time Series (SITS) analysis methods using all available data (Wulder et al., 2012) have defined the state of the art for continuous change monitoring and map updating (Zhu et al., 2014). However, reference data for accuracy assessment in time series analysis is often sparse, especially in the temporal domain or in case of rare phenomena such as storm damages, which motivates the development of unsupervised SITS analysis methods.

In 2010, Asta storm caused severe damages over 8.1 million m3 of forest in Finland, as well as to the country infrastructure (e.g. buildings, power lines and railroads), totaling financial losses over 20 million euros. The estimation of forest damages was carried out mainly by drive-by field visits or even involving flyovers by the Finnish air force, and was slow, expensive and inaccurate. In Finland, a significant part of the forest is owned by small private owners, who are required to report to the Finnish Forest Centre before any logging operation. However there is often a long delay (months or over a year) between a forest logging notification and the corresponding logging operation.

In this work, we propose an unsupervised SITS analysis model based on Long Short-Term Memory-Autoencoder (LSTM-AE) (Srivastava et al., 2015) network, and compare it with a well-known harmonic model trained in a quasi unsupervised way. Both models are used to demonstrate the benefits of dense SITS analysis for improving temporal and spatial accuracies of both forest logging notifications and storm damage mapping, compared to current operational practices in Finland.

All available 361 Landsat images acquired between 1997 and 2015 over adjacent footprints (paths 189 and 190, row 17) were downloaded and atmospherically corrected on USGS ESPA on-demand pre-processing platform. Scenes in path 190 were registered to a reference scene on path 189. All images were then stacked, keeping only pixels on the overlapping areas. Through this operation the effective temporal acquisition frequency was doubled, compensating for a low number of observations at high latitudes due to frequent cloud cover in summer and low sun angle in winter, that prevented radiometric correction.

The study area was located at Hyytiälä Forest Research Station in Finland, and entirely comprised in the overlap area between the two adjacent Landsat footprints. Reference data for accuracy assessment was collected from forest logging notifications of the Finnish Forest Centre, as reported by forest owners before normal logging operations (clearcuts, thinning cuttings) or consecutive to three major storm events between 2010 and 2015. In addition, clearcuts between 2010 and 2015 were visually delineated between two very high resolution images, with an improved spatial accuracy of the reference data.

We applied a LSTM-AE model, an autoencoder (AE) creating concise representations of time series data while preserving temporal information by using specific long short-term memory (LSTM) cells. Our version of the model was originally created for electronic health records - which are often sparse and contain both random and systematic biases - then adapted to SITS analysis which present similarities in terms of irregular temporal sampling and sparsity of break events.

As a baseline, the Continuous Change Detection and Classification (CCDC) algorithm (Zhu et al., 2014) was applied to the Landsat image stack (on the Short-Wave Infrared Band - SWIR 1 at 1.55-1.75 mm) in a quasi-unsupervised way, skipping the final supervised classification steps. The CCDC parameters were tuned interactively on pixels of known stable forest plots and clearcuts within the studay area. Once the parameters were set, change magnitude was computed automatically between the relevant dates corresponding to each event (storm or clearcuts). Finally, change magnitude maps were thresholded to filter out undesired changes and produce change maps.

The spatial accuracy of unsupervised forest change maps from LSTM-AE and CCDC were comparable, while the LSTM-AE approach was fully automatic and unsupervised. In addition, unsupervised change maps obtained from time series were more timely and temporally accurate than reference semi-supervised maps obtained from a bi-temporal change detection method used operatively. Results confirm the benefits of dense time series analysis using all available data instead of annual time series or bi-temporal change detection, for improved tolerance to missing observations and increased timeliness in the detection of forest changes.

Automatic analysis of dense image time series enables the production of accurate forest change maps, more frequently and cost-effectively than e.g. Lidar mapping (5-6 € /ha, currently used every 5 years or more to map forests over the whole Finland). The proposed methods are almost entirely automatic and thus suited to operational conditions, accurately updating logging notifications and storm damage maps.