Sub-Pixel Crop Type Classification Using PROBA-V 100 m NDVI Time Series and Reference Data from Sentinel-2 Classifications

Abstract

This paper presents the results of a sub-pixel classification of crop types in Bulgaria from PROBA-V 100 m normalized difference vegetation index (NDVI) time series. Two sub-pixel classification methods, artificial neural network (ANN) and support vector regression (SVR) were used where the output was a set of area fraction images (AFIs) at 100 m resolution with pixels containing estimated area fractions of each class. High-resolution maps of two test sites derived from Sentinel-2 classifications were used to obtain training data for the sub-pixel classifications. The estimated area fractions have a good correspondence with the true area fractions when aggregated to regions of 10 × 10 km 2 , especially when the SVR method was used. For the five dominant classes in the test sites the R 2 obtained after the aggregation was 86% (winter cereals), 81% (sunflower), 92% (broad-leaved forest), 89% (maize), and 67% (grasslands) when the SVR method was used.including the linear mixture model (LMM) [4], artificial neural network (ANN) [9-11], regression tree [5], fuzzy classification [6,12], and support vector machine (SVM) [13].Liu and Wu [11] argued that the non-linear models, especially neural network-based models, outperformed the traditional linear unmixing models. Support for such a conclusion is given in Verbeiren et al.[1] who compared ANN and LMM in an attempt to generate a sub-pixel map from SPOT-VEGETATION 1 km data across Belgium. The authors showed that the ANN approach outperformed LMM and that, for the major classes, the acreage estimates obtained via ANN, when aggregated to the level of the administrative regions, were in good agreement with the true values. Also, a multilayer perceptron (MLP) neural network regression algorithm has been shown to outperform the regression tree algorithm [5]. Atkinson et al. [9] also obtained better results with ANN than with the other tested methods but pointed out that its successful implementation depends on accurate co-registration and the availability of a training data set. Liu et al.[14] compared a linear spectral unmixing model, a supervised fully-fuzzy classification method and a SVM to generate a fraction map and achieved the most accurate fraction result using SVM. Six machine learning methods were compared in a recent study [15] based on multiple criteria, where the authors found that, in general, no method performs best for all evaluation criteria. However, when both time available for preprocessing and the magnitude of the training data set are unconstrained, support vector regression (SVR) and least-squares SVM for regression clearly outperform the other methods.Regarding the satellite imagery widely used for agricultural monitoring, SPOT-VEGETATION (SPOT-VGT) sensors provided one of the longest time series of multispectral reflectance since 1998. The mission was succeeded in 2013 by PROBA-V (PRoject for On-Board Autonomy-Vegetation), a small satellite commissioned by the European Space Agency. The sensor on-board PROBA-V generates products at three different ...

Pixel and Object-Based Machine Learning Classification Schemes for Lithological Mapping Enhancement of Semi-Arid Regions Using Sentinel-2A Imagery: A Case Study of the Southern Moroccan Meseta

“…However, the generalized lithological maps generated from various fractional images engenders often difficulties in producing homogeneous maps [19], [20], and also produce an inconsistent generalized thematic results [21]. This technique is seldom applied in lithological mapping but widely used in crop mapping [22], [23]. It was acknowledged in previous studies [24], [25], that a pixel is not the optimal spatial unit for lithological mapping.…”

Multi-temporal Crop Type and Field Boundary Classification with Google Earth Engine

“…Immitzer et al [9] confirmed the high values of the red-edge and shortwave infrared (SWIR) bands for vegetation mapping but the usage of these raw bands additionally improved the classification tasks in comparison with those achieved by single index-based classification studies. Dimitrov et al [10] too used Sentinel-2 as training data to apply sub-pixel classification to PROBA images with a resolution of 100 m/pixel. The chosen machine learning methods, i.e., SVM and RF, provided classification accuracies ranging from 67% for grasslands to 92% for broad-leaved forests.…”

Early Crop-Type Mapping Under Climate Anomalies

“…For instance, Saini et al [7] tested single-date images from Sentinel-2 using support vector machines (SVMs) and Random Forest (RF) and showed that the OA's of the classifications achieved by RF and SVM using one Sentinel-2 image were 84.22% and 81.85%, respectively [7]. While some studies used only one image for classification [7][8][9][10][11][12][13], other researchers have preferred multi-temporal information [14][15][16][17][18]. A long short-term memory neural net (LSTM) achieved an overall accuracy of up to 89.7% with Sentinel-2 time series [19,20].…”

“…A long short-term memory neural net (LSTM) achieved an overall accuracy of up to 89.7% with Sentinel-2 time series [19,20]. Another study based on support vector regression (SVR) and artificial neural network (ANN) achieved classification accuracies of up to 92% for broad-leaved forests [18]. In general, the multi-temporal approach captures better spectral characteristics over the growing season and improves crop types distinction.…”

“…In most studies, indices such as the NDVI were used to derive plant-specific differences [3,5,17,18,[24][25][26]. Belgiu et al [17] described an object-based approach combined with NDVI time series.…”