4.4. Yield estimation

Crop yield determination is crucial for planning the food security situation of individual farmers as well as a district or even a whole country. Agriculture, as the backbone of many developing economies (especially in Africa), provides substantial portions of their Gross Domestic Product (GDP). Thus, possibilities of obtaining yield estimates with reasonable accuracy prior to harvest is very important for such economies, since timely interventions can be made in case of poor yields.

Analysis of remote sensing data alone, or in combination with other ancillary data (e.g. soil moisture), permits the determination of crop yield prior to harvest period (Dempewolf et al., 2014; Morel et al., 2014). The ability of RS to provide information on crop status and health (through e.g. NDVI or LAI) is a key contribution to the estimation of potential crop yield. Two methodological approaches for estimating crop yield with RS data are discussed here.

The first is a simplistic approach whereby an empirical relationship is developed based on correlation between a computed vegetation index (e.g. NDVI) and in-situ measurement of yield at harvest time (Bolton and Friedl, 2013; Tucker et al., 1985). Ideally, the acquisition date of the image(s) from which the vegetation index is computed should coincide with the time of in-situ yield measurement. However, this is not always the case. Other indices such as LAI have also been explored in crop yield estimation.

Figure 4.12 Source: http://www.pecad.fas.usda.gov/
 

The downside of this approach, however, is that the established relationship is often only valid for the particular field, crop type and RS data acquired on specific date(s). This is because growth seasons differ in terms of precipitation, temperature, crop cultivated, fertilizer application, etc. Consequently, no two seasons can be the same. Thus, application of the developed relationship to other seasons may fail to give the desired results. Although the relationship can be made stronger by including more historical data, the results will mostly be sub-optimal. Nonetheless, this method may seem suitable and appropriate in data scarce areas with minimal data.   

Another approach to estimating/predicting crop yield is through crop models. Better predictions can be achieved through models by considering the factors that affect crop growth and yield for a year of interest. Information such as meteorological and climatic data (surface temperature, rainfall, etc.), soil properties and farming practices are combined with spatially explicit remote sensing derived information such as slope and vegetation indices (NDVI) to model crop growth and eventually make estimations of the final crop yield (Dorigo et al., 2007 – data assimulation).

Rodrigues et al. (2015) mapped the within field yield variability in a wheat field in Mexico using high resolution proximal soil sensing and hyperspectral RS data. The proximal soil sensing was carried out using a dual-dipole EM38 Mk2 sensor (Geonics, Mississauga, ON, Canada) conductivity measured simultaneously in the 0-0.75 and 0-1.5 m range. The correlation among yield, apparent soil electrical-conductivity and several narrow-band spectral indices that are known to be related to stress-detection photosynthetic indicators based on pigment were tested. Figure 4.13 shows a map of the within field spatial variability of yield, which ranged from 4.6 to 8 t ha-1.

Figure 4.13 Spatial variability of crop (wheat) yield at field scale in Mexico (Source: CIMMYT).
 

In order to understand the possible causes of this spatial variability in yield, the highest and lowest yield areas were identified on the yield map and the corresponding reflectance values (from the hyperspectral image) were extracted for the whole cropping season. Based on this, a plot of the reflectance profile of the highest and lowest yield areas was made (Figure 4.14).

The figure shows that the highest (green lines) and lowest (red lines) yield region followed similar spectral behavior from 400 to 770 nm, but differences in the reflectance at each yield level emerged from 770 nm to 840 nm (near infrared, NIR).

Zarco-Tejada et al. (2005) found similar reflectance behavior for low and high growth areas of cotton. Such differences in reflectance may open possibilities for nutrient stress diagnosis at the flowering stage while there is still time for crop management, or even at early stages such as GS31.

Figure 4.14 Reflectance profile from the highest and lowest yield areas on a test field across crop cycle (Source: CIMMYT).
 

Findings of the study demonstrates the potential of the use of multiple VIs together with proximal soil sensing data for predicting yield spatial variability.