A Decision Tree Approach for Spatially Interpolating Missing Land Cover Data and Classifying Satellite Images

Abstract

Sustainable Development Goals (SDGs) are a set of priorities the United Nations and World Bank have set for countries to reach in order to improve quality of life and environment globally by 2030. Free satellite images have been identified as a key resource that can be used to produce official statistics and analysis to measure progress towards SDGs, especially those that are concerned with the physical environment, such as forest, water, and crops. Satellite images can often be unusable due to missing data from cloud cover, particularly in tropical areas where the deforestation rates are high. There are existing methods for filling in image gaps; however, these are often computationally expensive in image classification or not effective at pixel scale. To address this, we use two machine learning methods—gradient boosted machine and random forest algorithms—to classify the observed and simulated ‘missing’ pixels in satellite images as either grassland or woodland. We also predict a continuous biophysical variable, Foliage Projective Cover (FPC), which was derived from satellite images, and perform accurate binary classification and prediction using only the latitude and longitude of the pixels. We compare the performance of these methods against each other and inverse distance weighted interpolation, which is a well-established spatial interpolation method. We find both of the machine learning methods, particularly random forest, perform fast and accurate classifications of both observed and missing pixels, with up to 0.90 accuracy for the binary classification of pixels as grassland or woodland. The results show that the random forest method is more accurate than inverse distance weighted interpolation and gradient boosted machine for prediction of FPC for observed and missing data. Based on the case study results from a sub-tropical site in Australia, we show that our approach provides an efficient alternative for interpolating images and performing land cover classifications.

Stochastic spatial random forest (SS-RF) for interpolating probabilities of missing land cover data

“…In order to produce our own missing data, we simulated missing data in cloud patterns on the images t + 1 and t + 2 . We simulated the missing data patterns independently of the NDVI based forest presence data using the process described in [16] based on [33]. We applied the SS-RF method to the same pixels and neighbourhoods in each image, identified by their geographical location (longitude and latitude), to ensure we were examining the same observations over time and could make fair assessments of model performance when interpolating the values at future time points t + 1 and t + 2.…”

“…This is the minimal information available for all images even when spectral data are missing. We demonstrated that this approach was surprisingly accurate [16]. However, like traditional compositing, all these statistical approaches to spatially interpolating missing data also do not measure the uncertainty of the interpolated values they produce.…”

“…In addition to compositing, there are statistical approaches to interpolating missing data in satellite images including canonical spatial kriging-based approaches [13,14] spectral-based and temporal-based models, and hybrid combinations of these [15]. In previous work [16] we developed a machine learning approach to spatial interpolation which is fast and accurate. Machine learning methods are popular and useful for remote sensing applications including for identifying clouds and cloud shadow as a preprocessing step to interpolation [11,17], and monitoring changes in forest cover [18].…”

“…The outputs of our method are a probability and an interpolated value. As an overview, our method fits a random forest with spatial information [16] to observed satellite images of the same area captured on previous, recent dates and constructs a Beta distribution to produce both a land cover classification and a probability of the classification of interest. The values from the Beta distribution can be used to interpolate missing data in images and provide a measurement of uncertainty of the interpolated values.…”

Optimal Tuning of Model Predictive Controller Weights Using Genetic Algorithm with Interactive Decision Tree for Industrial Cement Kiln Process

“…The weights or importance given to these metrics used in GA objective function is determined by using the IDT approach. The decision tree provides a decision support tool [38] for the selection of weights that can improve the objective function. It prompts for the operator response to determine the improvements in the performance metrics.…”