Maize (Zea mays
L.) crop has an increasing importance in the economy of Vietnam. According to the latest national statistical survey [1
] by 2018 the total maize area cultivated across the country was roughly 1,104,000 hectares, from which close to 10% is located in Dak Lak. Over the past 20 years, the maize cropping area has almost doubled, leaping from 663,000 hectares in 1997 [2
]. This increase was mostly driven by the growing demand from some of the largest animal feed, meat [3
] and food companies that have been operating in Vietnam since 1986.
In Dak Lak, maize fields are located either on hillslopes or in valleys, depending on the water demand of the crop, which is mostly grown under rainfed conditions. Only soils with high fertility and organic matter are used, due to the economic and terrain difficulties faced by farmers to fertilize crops. Many of the maize fields lie within the boundaries of either forest production land or fallow land, and hence it is crucial for forest management and land administration units to know how these lands are managed by locals and if there is any sign of forest encroachment. It is also important for the government and agroindustry to know where they could expect the most productive maize-producing regions to expand their activities. Most of the time, this information would come from either the official statistical data, which are often not up-to-date and commonly released two years later, or from estimates provided by local agencies.
Traditional methods for estimating cropland areas, such as the census household-based survey or the cadastral ground survey, are very time-consuming and may not reflect the real situation on the ground by the time the data are released. A less cost-and-time consuming and robust method that can provide objective and near real-time monitoring of crop types and their phenology stages is thus in high demand [4
]. From remote sensing, the annual cycle of vegetation phenology can be inferred and characterized by four transition dates that define the most important phenological stages of vegetation dynamics at annual time scales [8
]. These transition dates are known as green-up, maturity, senescence and dormancy. Today, with the growing development of advantaged technology, earth observation from satellites and other space-borne instruments have provided a spatially explicit and efficient way for cropland monitoring not only at large extents (e.g., regional) but also at local levels (e.g., field plot) [9
To map and monitor the cropland at a regional scale, remotely sensed data have to provide wide geographical coverage, high temporal and adequate spatial resolutions, and must be available at minimum cost [12
]. Extensively used sources of remotely sensed data such as the National Oceanic and Atmospheric Administration (NOAA) for the advanced very high resolution radiometer (AVHRR) and Landsat have some of these characteristics, but neither can meet the requirement for both spatial and temporal coverages, hence they are limited for such purposes [13
]. For detailed crop mapping, Landsat Thematic Mapper (TM)/Enhanced Thematic Mapper Plus (ETM+)/Operational Land Imager (OLI) data have demonstrated their suitability [14
]. In areas where crops are produced over large areas with relatively homogeneous cover and uniform cropping season, Landsat data are a reliable source [16
]. In areas where information on crops and cropping seasonality are poor, Landsat might only be able to distinguish cropland from non-cropland [17
]. Moreover, one should take into account that in humid tropical regions many crops have only a 3–4 month growing season, during which cloud cover occurs frequently. As a consequence, Landsat data become scarce, and often only few images with less than 10% cloud coverage become available each year. This is not sufficient to detect crop phenology and changes in crop cover [18
]. Improvement to overcome these disadvantages can be made when Landsat data are combined with other multitemporal sensor data that have a higher revisiting frequency at a larger spatial footprint (i.e., MODIS) and with data with a higher spatial resolution (i.e., Sentinel-2) [20
]. Although with large pixels, AVHRR data have almost daily coverage of the surface of the Earth, which can be used to generate weekly to 10-day maximum value composites, allowing capture of vegetation phenology via the normalized difference vegetation index (NDVI) or similar indices at national [23
] and global scales [25
]. The advantage of AVHRR is its temporal resolution at the cost of a limited spectral and spatial resolution. Its 1-km spatial resolution means that there is a high chance of mixed pixels in agricultural areas with small fields and complex land cover patterns [29
]. Turner et al. [31
] stated that in order to map in detail the variability and complexity of cropland systems at landscape levels, higher resolution than the AVHRR data are required.
MODIS, an abbreviation of the Moderate Resolution Imaging Spectroradiometer, is an instrument on board NASA’s Terra and Aqua spacecraft. MODIS provides daily coverage of the earth surface at 250-m resolution. It offers an opportunity for more detailed monitoring and mapping land use land cover (LULC) covering large territories [32
]. Liu et al. [33
] used MODIS 8-day Enhanced Vegetation Index (EVI) to map cropping patterns, which were defined by the planting sequence and spatial arrangement of crops by field [34
], of Henan province in China for three consecutive years. A threshold-model method was used to retrieve phenological metrics, producing an overall accuracy of approximately 84% in extracting the cropping patterns. Temporal NDVI and EVI from MODIS TERRA provided a successful discrimination in phenology and the ability to map large areas of soybean and maize in the state of Paraná, Brazil for the 2010/2011 and 2011/2012 crop years [35
]. Vintrou et al. [36
] used a landscape stratification of MODIS MOD13Q1 16-day NDVI time series at 250-m pixel resolution to produce a crop map for Mali. They concluded that the MOD13Q1 NDVI-based method produced a better map, compared to the other global products that were available at that time. Time series of optical remote sensing vegetation index (VI) data are often noisy, due to cloud contamination. To understand crop phenology and cropping season variations through interpreting time series of optical remote sensed VI data, noise-reduction methods to reconstruct those temporal VI data are widely used. Use of filter-based methods that employ a predefined filter to fill the gaps and smooth the multi-temporal data in a local moving window, such as the Savitzky–Golay (SG) filter, are very common [37
]. The SG filter performs a least-squares-fit over a set of consecutive values within a fixed window by a polynomial of a certain degree [42
]. Chen et al. [37
] embedded the SG filter into a time series model to reconstruct MVC SPOT-VGT image data of China. They were able to conclude that SG is, although simple, a very robust method to smooth noisy NDVI data and to fill out missing pixel values. A study by Zhou et al. [43
] compared the performance of four different remote sensing time series reconstruction methods over space and biome types. He found that for tropical and subtropical regions, the SG approach yielded the best results.
To translate the crop phenological data constructed from VI time series into a cropping pattern map, different classification approaches can be selected, following either an unsupervised or a supervised approach. The Support Vector Machine (SVM) classification algorithm is a supervised approach where samples must be taken to train the VI time series. An SVM network is a learning machine for two-group classification problems [44
]. An SVM classifies data by finding the best hyperplane that separates all data points of one class from those of the other class. Foody and Mathur [45
] used discriminant analysis, decision tree, feed forward neural network and SVM algorithms to classify crop types in a village in eastern England from airborne thematic mapper imagery. They found that of the four algorithms, SVM provided the best overall classification accuracy for mapping winter–spring crops. A similar conclusion concerning SVM classification accuracy was drawn from a study conducted in England and Spain by Pal and Mather [46
]. They compared SVM with artificial neural network and maximum likelihood algorithms and showed that the SVM achieves a higher degree of classification precision than either the machine learning (ML) or the artificial neural network (ANN) classifier, and that the SVM can be used with limited training datasets and high-dimensional data.
In Dak Lak, due to the complexity of the landscape and the recent intensification of practiced cropping systems, and the absent of any prior maize cropping patterns map, we applied the SG filter on MODIS Terra MOD13Q1 Enhanced Vegetation Index (EVI) data to reconstruct the temporal EVI data and employed the SVM classification system on the SG smoothed-and-gap filled EVI time series to identify maize cropping patterns.
The Savitzky–Golay filtering algorithm proved to have great potential for reconstructing the EVI profiles of seven vegetation cover types in Dak Lak, where cloud coverage is a prominent problem. By interpolating the missing EVI-value pixels and reconstructing the EVI curves based on the temporal behavior of the greenness values from each vegetation types, the Savitzky–Golay algorithm can delineate the maize cropping seasons from the others. This is important due to the extensive farming practices and complex cropping systems in a very dynamic and mosaic landscape of Dak Lak, which make it hard to identify crops and where they are grown, especially for those classified as annual cash crops. This is the reason why the official land use map [54
], which is released once every 5 years, cannot provide detailed information on the kind of annual cash crops that are currently cultivated in the province (Figure 2
). In fact, this most up-to-date official map is only able to provide rough information on where one would expect to find annual cash crops, including maize, and sometimes no crop would be shown since the land was already left fallow.
The MOD13Q1 EVI does have advantages for monitoring seasonal growth and changes of vegetation covers at large scale, but its 250-m spatial resolution has the disadvantage that is larger than the size of field. This means any maize field smaller than that pixel size will be missed when mapping due to the mixed EVI signal/value with other dominant vegetation cover types. Moreover, the scattered maize fields in this mosaic landscape would also be missed during the SVM classification process due to the nature of the linear SVM algorithm.
The linear kernel SVM classifier was able to map the maize cropping patterns with a producer’s accuracy of 74% but missed almost 26% of the validation GPS fields for the other annual cash crops class and rice class. This could be due to narrow margin hyperplanes separating maize crops from the others, and especially the ‘1-season maize’. Improvement could be made by having a better representation of training samples with the help of more representative GPS maize fields distributed over the landscape, or by comparing with other nonlinear kernel SVM classifiers.