Governments and agricultural managers require information on the spatial distribution and area of cultivated crops for planning purposes. Governments can adequately plan the import and export of food products based on such information. Although some agricultural ministries annually commission their staff to map different crop types, these ground surveys are expensive and yet cover only a sample of farms. Remote sensing data, together with ancillary information, enable the determination of the spatial distribution of crops at varying spatial scales with relatively little financial resources. This can, however, be achieved with various degrees of uncertainty, as some crops are spectrally similar and pixel sizes sometimes bias the estimation of crop acreages.
Knowledge of the growth cycle or calendar of crops is essential for accurate interpretation of RS data (Forkuor et al., 2014; Peña-Barragán et al., 2011; Son et al., 2013). The cropping calendar entails the period from land preparation, planting, growth, pollination, senescence and harvesting. The spectral reflectance of crops at each of these growth stages is referred to as the temporal profile of the crop. This information is extracted from satellite images based on surveyed locations (boundaries) of the dominant crop types in the area of interest. Dominant crop types are surveyed during field campaigns that ideally should coincide with the period of satellite image acquisition.
For example in a study in Mali, the boundaries of 48 fields representing six dominant crops were mapped during an extensive field campaign, from which the temporal profiles of each crop were extracted. Figure 4.8 shows typical temporal profiles of dominant crops in Mali based on NDVI images for different crop types calculated from DigitalGlobe data. It shows the typical growth cycle of each crop during the cropping season (May – October).
(Source: STARS team in Mali)
Based on the potential uniqueness of these temporal profiles, UAV and satellite images can be classified to reveal the spatial distribution of crops in the area of interest.
Apart from spectral information, several other information layers can be added to improve the accuracy with which different crops can be identified. An example is textural information (Haack and Bechdol, 2000; Sheoran and Haack, 2013).
Texture represents the degree of local spatial variations in an image. Different crop types, by virtue of their spatial arrangement, have different textural properties. Derivation and addition of textural measures to the spectral information can therefore improve classification accuracies.
Both texture and context are adding important information for the classification of image segments. Context refers to the relation between coarse and fine image segments. Texture serves as a valuable parameter in addition to spectral reflectance for characterizing the different segments. The texture parameters, which worked best, were GLCM (Grey Level Co-occurrence Matrix) and GLDV (gray-level difference vector) (Conrad et al., 2010; Novack et al., 2011).
Despite acquiring sufficient UAV/satellite and field data, crop classification can be very challenging and result in low accuracies. At the core of this difficulty is high variability in the spectral characteristics of the crops under study. In other words, the temporal profiles of the crops under study, are ideally unique for each crop, but this is often not the case.
Figure 4.9 (above), for example, shows the monthly temporal profiles of the dominant crops in our test site in Mali. Each column (labelled with a month name) depict the spectral patterns extracted from five quadrats within a field. The figure shows high similarity in the profiles, which could make crop identification and separation from satellite/UAV images quite challenging.
This high spectral and spatial variability can be attributed to a number of reasons. These include:
- Overlaps in cropping calendar, especially in rainfed dominated agricultural areas where different crop types are planted and harvested around the same time, leading to similarities in their temporal profiles.
- Differences in management practices (e.g. tillage, weeding, fertilization, etc.) between and within fields result in high spectral and spatial variability.
- Variations in soil type, depth and fertility
- Intercropping, i.e. cultivation of different crop types on the same land
- Proximity of natural/semi-natural vegetation to cultivated areas
- Occurrence of excessive trees on agricultural plots
- Water accumulation
- Occurrence of pest and diseases in some portions of a field
In order to reduce the effects of the above-mentioned factors on crop classification, a number of measures can be pursued: These include:
- Inclusion of additional RS data, e.g. Synthetic Aperture Radar (SAR) data (Forkuor et al., 2014; McNairn et al., 2009).
- Landscape stratification based on soil, topography, climate, etc.
- Performing object, instead of pixel-based, image analysis (Peña-Barragán et al., 2011).
- Testing different classification approaches such as the sequential masking classification algorithm (Forkuor et al., 2015; Van Niel and McVicar, 2004).