Machine learning techniques for estimating crop areas at the sub-pixel level.

Accurate forecasts of local and regional agricultural production are essential for agricultural market contractors and operators to assist prize agreements as early as possible in the crop growing season. These forecasts are also helpful for societies to anticipate to limited food availability. Satellite remote sensing is a fairly recent but already established technology for large-scale agricultural production forecasting thanks to its ability to provide synoptic, spatially explicit and repetitive observations of the Earth’s vegetative cover.

From the currently available satellite missions, sensors characterized by both low spatial and high temporal resolution (e.g. MODIS, SPOT-VGT) are most suitable for crop production forecasting over large areas. However, the characteristics of these satellite data combined with those of the large study areas they are mostly applied to, give rise to three specific modeling requirements:

·         To allow their use over large areas covered by large numbers of image pixels, production forecasting algorithms must be selected that minimize time-consuming human intervention and maximize automation;

·         The selected algorithm(s) must be able to handle data with a low signal to noise ratio and a small number of reference observations;

·         The low spatial resolution of the data calls for the use of sub-pixel assessment strategies. The use of endmembers (pure class signals) should however be avoided, as the availability of pure pixels is limited.

Machine learning methods have already displayed their ability to fulfill these three requirements, thus making them potentially suited for large-scale production forecasting. The main aim of this dissertation was to make a scientific analysis of the potential of machine learning methods for crop production forecasting at regional scales.

In this research, the cropped area was assessed through a sub-pixel land cover classification procedure. Four issues, that are each associated with a specific section of the sub-pixel land cover classification process, were subjected to an in-depth analysis.

A first issue that was tackled was the comparison of six machine learning algorithm for sub-pixel land cover classification and the identification of the most appropriate one. The machine learning methods – multilayer perceptron, support vector regression, least-squares support vector machine, bagged regression trees, random forest and boosted regression trees – were evaluated and compared with regard to six criteria that reflect their classification accuracy, their requirements for preprocessing, their robustness and their reference data requirements. Based on these criteria, the machine learning methods were ranked according to three scenarios that correspond to different decision making situations and differ in their weighting of the selected criteria. This assessment designated no overall winner, i.e. no method performed best for all scenarios. However, when both the processing time and the number of reference points were unconstrained, the support vector machines clearly outperformed all the other methods.

In several studies focused on sub-pixel land cover classification, it is stated that for the estimated land cover fractions to have a sound physical meaning, they must be non-negative and sum to one. The second issue this dissertation zoomed in on was the effect of explicitly including these so-called ‘fractional abundance constraints’ in sub-pixel land cover classification. Our results indicated that machine learning methods slightly benefit from constraining their outputs, but for most methods the obtained improvements were not significant. Embedding the constraints into the training phase of an artificial neural network yielded more accurate results than a post-processing approach that rescaled the model outputs to be non-negative and sum to one.

As the dimension of the (input) feature space increases, the accuracy of a prediction model will generally increase until an optimum is attained. Beyond the optimum number of input features, the model accuracy may either remain constant or even decrease. This phenomenon is often referred to as ‘the curse of dimensionality’ and was the third issue covered here. Different feature selection measures were applied to the problem of sub-pixel area estimation of grassland and maize, combined with two tree-based modeling approaches. For each measure, a simple single-feature search procedure was set against an iterative alternative. Our results showed that the feature importance measures embedded in the machine learning algorithms outperform the model-free filters. Also, as expected, iterative selection outperformed the single-feature evaluation in identifying the best-performing feature combination.

When validating sub-pixel classifications, most studies rely on overall accuracy statistics like RMSE or R². While these measures are suitable for simple model comparison, they fail at identifying the exact confusion patterns between the different classes. Former attempts at expanding the hard confusion matrix to make it applicable to sub-pixel classification did not lead to general acceptance, as there was no consensus about how to calculate the off-diagonal elements. The ‘STATCON’ expansion presented in this dissertation calculates the off-diagonal elements not by the commonly used Bayesian operators but by an approach based on the ‘linear statistical relation’ between the over- and underestimation of the area fraction values for the different classes. Based on a proof of concept using artificial datasets, we found STATCON to be a promising competitor to the currently available alternatives.

All previous issues were directly related to the task of crop area estimation. With regard to crop yield, we examined the potential of the boosted regression tree and the random forest approaches for (early-season) district level wheat yield prediction. The resulting models were found to deliver early-season estimates that are accurate enough to support decision making in the agricultural sector and to allow an operational use of the forecast model. To attain maximum prediction accuracy, incorporating predictors from the end of the growing season is however required.

Overall, our results showed that machine learning methods are capable of delivering fairly accurate predictions of crop area and yield at the spatial level of the municipality or district. At the pixel level, they reached a (sub-pixel) land cover classification accuracy comparable to that of parametric classification techniques reported in the literature. Selecting the most relevant input features and the most appropriate machine learning model were identified as key issues for maximizing the prediction accuracy. The newly developed STATCON matrix for inter-class confusion detection showed potential for identifying problematic classes, which is important for correctly focusing future model improvement efforts.