Mapping Drainage Ditches in Forested Landscapes Using Deep Learning and Aerial Laser Scanning

Digging drainage ditches is common forestry practice across northern European boreal forests and in some parts of North America (Lõhmus et al. 2015). Ditching helps with lowering the groundwater level in the wet parts of the forest to improve soil aeration and support tree growth (Laurén et al. 2021; Sikström and Hökkä 2016). Some of the most drained countries include Finland, the Baltic States, and parts of Sweden, where the drained forest stands comprise 20%–25% of the total forest area (Nieminen et al. 2018). However, the extensive use of ditches over a long period of time has resulted in a major change in wetland and soil hydrology and impacted the overall ecosystem functions of these regions (Kuglerová et al. 2020).

Recently, the intensive ditching practice has been identified as posing multiple environmental risks, particularly for degradation of wetland and soil, greenhouse gas emissions (Audet et al. 2017; Peacock et al. 2021), increased nutrient and sediment loadings to water bodies, and biodiversity loss (Holden et al. 2004; Lepistö et al. 2021; Lidman et al. 2017; Lõhmus et al. 2015; Nieminen et al. 2018). Research and actions were set up to minimize environmental loss and restore degraded land in the ditched forest landscape. However, such initiatives are significantly constrained by the lack of accurate and site-specific information of ditch networks. For example, a comparison between a national field inventory (Ståhl et al. 2011) and the best available high-resolution map of Sweden suggests that only 9% of the ditches that were inventoried by the national field inventory are mapped on current maps. In addition, 68% of the field-mapped channels were ditches, highlighting the scale of this lack of knowledge. A study in northern Sweden documented that accounting for all ditches nearly doubled the size of the stream network within a $68\text{-}{\mathrm{km}}^2$ catchment (Hasselquist et al. 2018). These findings are a clear indication that there is a discrepancy between the potential significance of artificial water bodies, such as drainage ditches, and their low representation in scientific research and policy (Koschorreck et al. 2020). The increasing understanding of the environmental risks associated with forest ditches, together with the poor representation of ditch networks in existing maps of many forest landscapes, makes detailed mapping of these ditches a priority for sustainable land and hydrological management.

The availability of high-quality remote sensing data has enabled successful mapping of different landscape features; accordingly, satellite-based hydrological analysis has become an effective tool for the assessment of water resources across the globe (McCabe et al. 2017). In open landscapes, such as agricultural lands or open peatlands, ditches can be detected from satellite images or aerial photos (Ayana et al. 2017). But mapping ditches in a forest landscape remains a significant challenge, given that such features are obscured by the tree canopy cover (Benstead and Leigh 2012; Hasselquist et al. 2018). A common solution for mapping small-scale watercourses in such landscapes (Benstead and Leigh 2012) is to model the accumulated flow from a digital terrain model and set a stream initiation threshold for extracting the stream networks. The stream initiation threshold refers to how large a drainage area is required to initiate a stream. However, the formation of stream heads varies depending on climate, topography, and soil conditions (Elmore et al. 2013; Jensen et al. 2017; Julian et al. 2012; Russell et al. 2015). Moreover, in landscapes where channels have been altered by humans, such as in the case of ditches, channel formation is no longer controlled by natural erosion. Hence, the flow accumulation–based hydrological modeling approach does not adequately capture a heavily modified ditch network in forested landscapes.

More recently, deep learning approaches have emerged as a mainstay of data processing and analysis in the field of water resources (Sit et al. 2020). Airborne laser scanning (ALS) data of high point density are being integrated with powerful machine learning and statistical tools to map ditches. Most ditch detection studies used ALS-derived digital elevation models (DEMs) but focused only on smaller study areas of up to 150 ha (Cazorzi et al. 2013; Rapinel et al. 2015; Roelens et al. 2018a, b). In addition, those ditch detection studies were predominantly performed in open areas, such as French marshes (Rapinel et al. 2015) and vineyards (Bailly et al. 2008), Belgian grasslands and peri-urban areas (Roelens et al. 2018b), Italian and US agrarian landscapes (Cazorzi et al. 2013; Passalacqua et al. 2012), Chinese rice fields (Qian et al. 2018), or near roads in Finland (Kiss et al. 2015). In such open landscapes, ditch detection is comparatively less complex, as tree cover is rare, and ditches are usually well maintained in these populated, easily accessed, managed landscapes.

These ditch detection methods using ALS data in open landscapes generally use data filtering to highlight ditches: for example, wavelet transformation (Bailly et al. 2008), relative elevation attribute (Cazorzi et al. 2013; Roelens et al. 2018b), geometric and Laplacian curvature (Passalacqua et al. 2012), linear filter (Rapinel et al. 2015), or topographic position index and standardized elevation index (Kiss et al. 2015). Data filtering follows a binary classification of raster layers into ditches and nonditches, while the classification can be performed on ALS point clouds, pixels, or linear objects depending on the chosen method. Different classification approaches have been used for binary classification, such as manual setting of thresholds using expert knowledge (Kiss et al. 2015; Passalacqua et al. 2012), logistic regression (Roelens et al. 2018b), and application of machine learning models (Bailly et al. 2011; Roelens et al. 2018a).

Deep learning presents a new avenue for data-driven hydrological analysis and modeling, especially with high-resolution imagery. Semantic image segmentation using deep learning is on the rise for many applications, from autonomous driving to virtual or augmented reality systems (Garcia-Garcia et al. 2017). However, the use of deep learning for image analysis in hydrological research remains limited. Deep learning techniques use multilayer models that can effectively capture the underlying complexity in heterogeneous hydrological systems (Sit et al. 2020). Deep learning can be effective for extracting small-scale hydrological features and can reduce prediction uncertainty by being insusceptible to raw and noisy data (Shen et al. 2018). Here, we developed a novel methodology by combining two state-of-the-art technologies—ALS and deep learning—for detecting drainage ditches in forested landscapes. We then validated the methodology using data from multiple countries in Northern Europe. To the best of our knowledge, this approach to detecting ditches has not been reported in the literature previously.

We trained a deep neural network on high-density ALS data and manually digitized ditches from 10 regions across Sweden. We set aside 20% of the data for testing the final model. In addition to this testing data, the model was applied to four additional test areas spread across Sweden, Finland, Latvia, and Poland [Fig. 1(a)].

Most previous ditch detection studies focused only on study areas smaller than $100{\mathrm{km}}^2$ within predominantly open areas. Here, we introduce a large dataset with 1,607 km of mapped ditch channels spread across $344{\mathrm{km}}^2$ in a landscape mainly dominated by forest. We demonstrate that our approach of combining deep learning with high-density ALS data to map drainage ditches has accuracy equal to or higher than that of previous studies. However, the studies by Graves et al. (2020), Kiss et al. (2015), Larson and Trivedi (2011), Rapinel et al. (2015), Stanislawski et al. (2018) reported only recall, which made them difficult to evaluate. Additional metrics on false positives would be helpful for a comparison with our results. It needs to be emphasized that we used only one topographical index to achieve this high accuracy, while other studies, such as Roelens et al. (2018a), relied on multiple indices for mapping drainage ditches in a smaller landscape (i.e., $68{\mathrm{km}}^2$). This substantially enhances the utility of our methodology, especially for implementing it at large spatial scales with limited computational resources. The model could map an area of $1{\mathrm{km}}^2$ in 8.6 s using a GeForce GTX 1080 Ti Graphics Card. Still, it will be worth exploring whether adding more indices besides HPMF can improve the prediction accuracy. For example, a recent study in Finland by Bhattacharjee et al. (2021) demonstrated that aerial photos may contain important information for mapping drainage ditches on peatlands. However, in areas hidden beneath canopy, it might be better to include topographical indices such as impoundment size index implemented in Whitebox tools 1.4.0 (Lindsay 2018), which can be included as multiband images. Thus, it seems worthwhile to explore these indices as additional covariates in the deep neural network model for future analysis.

Considering the inherent challenges of the study region’s glaciated and forested landscape, the performance of our ditch mapping approach was highly successful. For instance, the landscape presented substantial challenges due to landscape features that appear similar to ditch channels in the ALS data. These include natural streams, historical channels from the latest deglaciation, ravines on sediment soils, and variability in larger-scale topography from steep mountains to flat mires or smaller agricultural areas interspersed in the forest landscape. Although our model produced a fair amount of false positives in the sedimentary deposits in the Swedish test catchment, the overall performance of the model was notably better than the previous ditch mapping studies that were conducted in relatively less challenging, more homogeneous landscapes. Some of the false positives in our prediction might even be actual ditches that are missing or misplaced in the ground-truth digitized data [blue lines in Figs. 3(d and f)]. This is likely because of the differences in digitization approaches in the international datasets. Hence, the accuracy measures calculated from the international validation datasets should be interpreted cautiously. The high MCC and recall values of ditch prediction in the international test sites indicate that the developed methodology performs well in all countries around the Baltic Sea, although the model was developed based on training data from Sweden. The trained model can be adjusted for local conditions with transfer learning using local data.

Visual inspection of the test sites revealed that some of the false positives were natural stream channels, as demonstrated in Fig. 3(b). This suggests that it could be possible to train a deep neural network to detect natural stream channels in addition to drainage ditches. This needs to be explored in the future with additional ground-truth data on natural stream channels. A model that can map natural stream channels would complement traditional topographical modeling of accumulated flows from digital terrain models, especially if the model can map headwater stream channels, as these are often missing on current maps (Benstead and Leigh 2012), and deal with road embankments, which make small streams difficult to map using traditional topographical modeling of low accumulation (Lidberg et al. 2017).

The management of water systems is key to sustainable development, but headwater streams (Bishop et al. 2008) and artificial drainage ditches are often not included in monitoring programs (Koschorreck et al. 2020). This has large implications for future research and practical land-use management, as the ditch networks have legacy effects on greenhouse gas balance, water resources, and biodiversity. Ditches possess particular characteristics that often make them emit large amounts of greenhouse gases (Peacock et al. 2021). Specifically, the accumulation of sediment and the development of anoxia result in favorable conditions for the production of the potent greenhouse gas methane (Roulet and Moore 1995). Emissions of nitrous oxide (${\mathrm{N}}_2\mathrm{O}$) and carbon dioxide (${\mathrm{CO}}_2$) can be disproportionately large compared with the small areal extent of ditch surfaces (Audet et al. 2017; Peacock et al. 2019). In countries such as Sweden and Finland, which have significant areas of land drained for agriculture or forestry, these greenhouse gas emissions can make nonnegligible contributions to national greenhouse gas budgets (Koschorreck et al. 2020). Further, it is unclear whether peatland restoration or ditch cleaning to maintain forest growth is the best strategy to avoid negative environmental outcomes in the future. Mapping the drainage ditches in the forest landscape is necessary to address these important sustainability questions.

Mapping drainage ditches is an important first step in finding effective landscape and hydrology management strategies. We showed that semantic image segmentation with deep learning from high-resolution ALS data can be used to detect previously unmapped drainage ditches in forested landscapes in the Baltic Sea region with an overall accuracy of 99% and an MCC of 0.78. This novel technique requires only one topographical index, which makes it possible to implement on large scales with limited computational resources. Our method performs better on most of the metrics than previous ditch detection studies and at least equally well on all others, despite a more varied and challenging landscape in our test data dominated by forests. Visual inspection indicated that this method also classifies natural stream channels as ditches, which suggests that a deep neural network can be trained to detect natural stream channels in addition to drainage ditches.

Some or all data, models, or code generated or used during the study are available in a repository online in accordance with funder data retention policies. The code and a reproducible example are available from github (Lidberg et al. 2021). The data are available from Mendely data (Ågren et al. 2022), doi: 10.17632/zxkg43jsx8.1.

We thank Liselott Nilsson, Eliza Hasselquist, Gudrun Norstedt, Lars Strand, Anders Hejnebo, Björn Lehto, Magnus Martinsson, Marcus Björsell, Tobias Johansson, Andrew Landström, and Catarina Welin for spending 513 h digitizing the ditches used to train the model in this study. Further, we thank Jerry Lidberg for lending us his gaming graphics card. The study was funded by KEMPE, WASP-HS, SMHI, Vinnova (2014-03319), Formas (2019-00173 and 2021-00115), and the Interreg project: Water Management in Baltic Forests.