How can the shapes and distribution of deforested areas inform us about the agents of changes on the ground?

Figure 1. Relationships between four types of deforestation spatial patterns, visually differentiated using Google Earth Imagery, and related land-use agents. Photos provided by the Terra-i project.

A first look: relevance of deforestation spatial patterns research

Google Maps can clearly introduce us to what the spatial configurations of land cover are. Zooming in to a particular forested region like the Amazon allows us to visualise differences in the shapes (composition) and distribution (configuration) between deforested areas (Figure 1). These structural types of deforestation, commonly referred to as “deforestation spatial patterns”, may be linked with known land-use processes and agents that promote forest disturbances at the ground. In this way, the identification and mapping of types of deforestation patterns is key to providing multi-level, flexible information of use to forest conservation groups, land managers and related practitioners. Moreover, knowledge on the spatial configuration and composition of deforestation patterns assists in the modelling of future land-use scenarios.

Methodology: what’s new?

Several techniques have been developed to facilitate the understanding of deforested elements and their dynamics in landscapes (Figure 2). Of these methods, window analysis, has been widely used not only to characterise the spatial configuration and composition of deforested areas, but also to determine the agents associated with them: small-scale farmers, large plantations, or cattle ranchers, among others. The characterisation of spatial pattern typologies using window analysis has been strongly supported by the development of specialised software for assessing landscape fragmentation, such as the FRAGSTATS analysis package. This software has helped to strengthen the analysis of landscape structure using landscape pattern indices called “landscape metrics”.

Figure 2. Some examples of methodologies for analysing spatial patterns and elements of deforested areas. Refer to the sources provided for further information on these methodologies.

The workflow behind the window analysis approach is a type of “Knowledge discovery in databases” (KDD) process (Figure 3). KDD includes multidisciplinary activities that aim to develop new information from existing databases, and its major application areas are in marketing, fraud detection, telecommunication and manufacturing. KDD encompasses data storage and access, algorithm scaling (data mining) for application to massive data sets and results interpretation.

Figure 3. An overview of the steps in the KDD Process. Source: Fayyad et al. (1996).

In the case of pattern analysis research, input data basically consist of a set of landscape fragmentation properties (landscape metrics) that can be extracted from the deforested elements using a known window size to create a grid. After data extraction and preprocessing, a range of data mining algorithms for classification can be used; however, the vast majority of pattern analyses have been supported by the use of decision trees. Alejandro’s research focuses on the implementation of feed-forward artificial neural networks that tend to be more robust than decision trees. Artificial neural networks improve their performance by ‘learning,’ a process that may continue even after the training set has been analysed. Some disadvantages of artificial neural networks are the difficulties in explaining their workings to end users (unlike decision trees, which are easy to understand), overfitting, or failing to converge in the learning phase (they do not guarantee convergence or optimality), among others.

Enhancing the robustness of deforestation pattern analysis research

The perfect algorithm for a KDD project does not exist; it depends on the problem context and database features. Alejandro’s research has identified the following steps as key to enhance the robustness of the KDD process for mapping deforestation spatial patterns:
Preprocessing the input data to avoid missing values and noise, reduce of data dimensions (e.g. PCA analysis) and/or remove highly correlated or redundant variables;
Using bagging or cross-validation techniques to train artificial neural network models and reduce overfit;
Tuning and testing a range of values over the main artificial neural network parameters such as the number of hidden layers and weight decay;
Selecting the best artificial neural network models according to robust classifier performance metrics such as the Area Under the Receiver Operating Characteristic curve (AUC-ROC), which is not influenced by the costs of misclassification;
Plotting the structure of the selected artificial neural network model in a friendly format that includes the relative importance of different variables so as to explain to end users what could be behind “the black box”.

Preliminary results

A KDD project implemented in a pilot area (approximately 403,200 sq. kilometers) in Brazil showed how powerful artificial neural networks can be for classifying five types of deforestation patterns (four of which are illustrated in Figure 1). Using a window size of approximately 30 km by 30 km, a training set was randomly created, selecting 15 representative grids (samples) by pattern (Figure 4). From this validation set, 5 samples were used to train the model and the remaining 10 samples were used to test and assess the model’s performance. The input data consisted of approximately 113 input variables (FRAGSTAT-like metrics at patch, class and landscape level in grids extracted from 2004-2012 aggregated detections of the Terra-i dataset). In regard to artificial neural network structure, a set of values was defined for three parameters: the number of neurons (2 to 40 with an increment of 5) in the hidden layer, the penalization parameter associated with the weight decay (0.2 to 0.5 with an increment of 0.1) and the activation function (linear, entropy or censored).

As result of the iterative KDD process, 423 unlabelled grids were automatically classified (Figure 4) with a reported AUC value of 0.65 for the Terra-i dataset. Although the results were promising in the pilot area, a set of challenges (see next section) will likely be encountered when applying the method to a larger area, where the robustness of the methodology workflow needs to be tested and assessed.

Figure 4. Training and validation sets (top) and results (bottom) of the artificial neural network structure that provided the best AUC value for the classification of five types of deforestation patterns: consolidated (c), diffuse (d), fishbone (f), geometric (g) and linear (l). Underscore indicates validation set.

Challenges

Although there are many challenges to this approach from both the geospatial (management and preprocessing of large spatial datasets) and computational (database manipulation and parallel processing) perspectives, the most relevant from the research perspective are:
Assessing the influence of spatial resolution from the two deforestation datasets (Terra-i and GFC with 250m and 30m, respectively) in the final classification results;
Identifying and comparing the deforestation spatial patterns by five increasing window sizes (8, 15, 30, 60 and 120 km);
Defining optimal artificial neural network models from input variables that have ecological importance so as to reduce the number of inputs tested in the pilot area (more than a hundred to less than 30);
Using the optimal artificial neural network models to determine the temporal dynamics (by three-year intervals) of the types of deforestation spatial patterns by window size.
This research is supported by the usage of software such as ArcInfo and R, which provide a broad set of libraries for spatial analysis and database management. The former is being used to perform vector and raster operations (masking datasets with the target region and creating the forest/non-forest maps) as well as producing cartographical outputs. The latter allows for extracting the FRAGSTAT-like metrics and performing data mining processes using the caret R-package.

Alejandro is grateful to Mark Mulligan (dissertation supervisor at King’s) and the Terra-i team for their support, especially in the optimisation of technical and research aspects of this work.