E. Izquierdo-Verdiguier

A Cloud-Based Multi-Temporal Ensemble Classifier to Map Smallholder Farming Systems

Smallholder farmers cultivate more than 80% of the cropland area available in Africa. The intrinsic characteristics of such farms include complex crop-planting patterns, and small fields that are vaguely delineated. These characteristics pose challenges to mapping crops and fields from space. In this study, we evaluate the use of a cloud-based multi-temporal ensemble classifier to map smallholder farming systems in a case study for southern Mali. The ensemble combines a selection of spatial and spectral features derived from multi-spectral Worldview-2 images, field data, and five machine learning classifiers to produce a map of the most prevalent crops in our study area. Different ensemble sizes were evaluated using two combination rules, namely majority voting and weighted majority voting. Both strategies outperform any of the tested single classifiers. The ensemble based on the weighted majority voting strategy obtained the higher overall accuracy (75.9%). This means an accuracy improvement of 4.65% in comparison with the average overall accuracy of the best individual classifier tested in this study. The maximum ensemble accuracy is reached with 75 classifiers in the ensemble. This indicates that the addition of more classifiers does not help to continuously improve classification results. Our results demonstrate the potential of ensemble classifiers to map crops grown by West African smallholders. The use of ensembles demands high computational capability, but the increasing availability of cloud computing solutions allows their efficient implementation and even opens the door to the data processing needs of local organizations.

Kernel change discriminant analysis for multitemporal cloud masking

This paper presents a multitemporal feature extraction method based on kernels that is particularly designed for change detection. The method provides features that maximize specific changes between two dates while minimizing sources of errors, such as residual land-cover changes and misregistration errors, in the time series. The extracted features computed in the kernel feature space can deal with non-linear relations between samples at different dates. Moreover, no supervised information is required to find the changes of interest for the selected dates in the time series. The effectiveness of the proposed method is successfully illustrated in a cloud masking application using a Landsat time series. Results show that the proposed method provides the most discriminative features in terms of cloud detection when confronted with state of the art linear and nonlinear unsupervised feature extraction algorithms. In particular, extracted features with the proposed method enable automatic cloud detection in multispectral time series.

Advanced Feature Extraction for Earth Observation Data Processing

Extracting rich and semantically discriminative features from remote-sensing data is of paramount relevance to advance in the understanding, visualization, classification, and representation of the data. In this article we taxonomically categorize recent feature extraction methods that have been applied to remote-sensing problems, focusing on kernel-based methods, convolutional neural networks, and principal curves. We review the formulation and relevant aspects of these methods. Advantages and shortcomings are illustrated in real experiments dealing with multi- and hyperspectral image processing of varying spatial resolutions.

Integrating support vector machines and random forests to classify crops in time series of Worldview-2 images

Crop maps are essential inputs for the agricultural planning done at various governmental and agribusinesses agencies. Remote sensing offers timely and costs efficient technologies to identify and map crop types over large areas. Among the plethora of classification methods, Support Vector Machine (SVM) and Random Forest (RF) are widely used because of their proven performance. In this work, we study the synergic use of both methods by introducing a random forest kernel (RFK) in an SVM classifier. A time series of multispectral WorldView-2 images acquired over Mali (West Africa) in 2014 was used to develop our case study. Ground truth containing five common crop classes (cotton, maize, millet, peanut, and sorghum) were collected at 45 farms and used to train and test the classifiers. An SVM with the standard Radial Basis Function (RBF) kernel, a RF, and an SVM-RFK were trained and tested over 10 random training and test subsets generated from the ground data. Results show that the newly proposed SVM-RFK classifier can compete with both RF and SVM-RBF. The overall accuracies based on the spectral bands only are of 83, 82 and 83% respectively. Adding vegetation indices to the analysis result in the classification accuracy of 82, 81 and 84% for SVM-RFK, RF, and SVM-RBF respectively. Overall, it can be observed that the newly tested RFK can compete with SVM-RBF and RF classifiers in terms of classification accuracy.

On the use of guided regularized random forests to identify crops in smallholder farm fields

The smallholder farms located in sub-Sahara Africa are typically characterize by heterogeneous mosaic of crops, soils, weather and farm practices. Automatic crop identification over smallholder fields is challenging when one only uses the spectral information of very high spatial resolution image time series. The extraction of spatial-spectral information is important to reach classifier accuracy. We deploy cloud computing techniques to allow working with thousands of features derived from an image time series. However, this number of extracted features is forces one to select the most important features for identification routines. This paper introduces a simple feature selection method based on Random Forest - the Guided Regularized Random Forest (GRRF) - which reduces feature dimensionality without loss data information. Preliminary experiments show that we can reach an overall accuracy by around 63%, and the results using random forests trained by GRRF features improve by around 2.5% the results by a Random Forest classifier that uses all the features.

Identifying crops in smallholder farms using time series of WorldView-2 images

A high heterogeneity in farming factors (soils, weather, inputs, practice) characterizes the typical smallholder farm landscapes of sub-Saharan Africa. This complicates automatic classification to crop when using only spectral information of very high spatial resolution image time series. This work addresses the crop identification problem in smallholder landscapes through three steps: features extraction, feature selection and classification. Feature extraction is used to exted the spatial-spectral information of the farm fields, with a substantial number of features considered through cloud computing. Feature selection is based on correlation between the features and the labels of the fields crops and it is applied to reduce the dimensionality of the data without lose information. Finally, a random forest classifier is applied to identify a crop class per field. Good preliminary results were obtained reducing the number of features from 1638 to 66. The overall accuracy achieves 80% in the test set using a random forest classifier.

Introducing co-clustering for hyperspectral image analysis

This work introduces the use of co-clustering for hyperspectral image analysis. Co-clustering is able to simultaneously group samples (rows) and spectral bands (columns). This results in blocks, which do not only share spectral information (classical one way clustering) but also share sample information. Here, we propose using a co-clustering algorithm based on Information Theory - the optimal co-clustering is obtaining minimizing the loss of information between the original and the co-clustered images. A hyperspectral image (160000 samples and 40 bands) is used to illustrate this study. This image was clustered into 150 groups (50 groups of samples and 3 spectral groups). After that, blocks of the spectral groups was independently classified to assess the effectiveness of the co-clustering approach for hyperspectral band selection applications. Furthermore, the results were also compared with state-of-art methods based on morphological profiles, and the covariance matrix of the original hyperspectral image. Good results were achieved, showing the effectiveness of the Co-clustering approach for hyperspectral images in spatial-spectral classification and band selection applications.