Remote sensing and crop recognition: Improving the information system supporting smallholders in Mali
Creating a transparent information system is crucial to improve the position of smallholder farmers in Mali. For instance, knowledge as a result of crop monitoring can lead to sustainable intensification, a decreasing yield gap and higher quality products. One of the first steps in improving this information system is crop recognition. In the STARS project remote sensing images are used to recognize and classify crops on smallholder farms. Crop variability, crop similarity at certain growing stages, different land preparation practices and landscape characteristics were, however, posing challenges for a successful classification (Figure 1). To come to a successful classification, this research tested the influence of different classification methods, vegetation indices [VI], strata and spatial/temporal resolutions.
- Figure 1. Within-field variability due to trees and a variable crop performance (dark red circles are trees, other red pixels represent crops or weed, green areas represent bare soil and the bright green spots are most likely caused by ants eating the crops). WorldView-2 image: 03-09-2015.
The data collection of this research took place in Mali at two sites. Both sites were part of the STARS project and millet, sorghum, peanuts, cotton and maize were the five dominant crop types. The first study area was ICRISAT’s research station in Samanko. This area was controlled by researchers and crops grew under controlled conditions. A single crop in this part was expected to show a relatively homogeneous spectral and temporal signature, as there were hardly any within-field irregularities. This area was used to discover small differences between similar crops, to create an extensive temporal dataset and to investigate the performance of different classification methods on well managed crops. The second area covered an area of about 30 square kilometers around Sougoumba, a small village near the border of Burkina Faso. The heterogeneous landscape made this area suitable for this study; different soil types, field sizes and planting densities were all occurring. In addition, there were differences in elevation, as part of the area was covered by a flood plain and part by an old plateau.
For the first study area an extensive Unmanned Aerial Vehicle [UAV] dataset was created. Therefore, an octocopter Geo-x8000 mounted with a Tetracam was flown over the area four times a week, at heights of 35, 90 and 185 meter (spatial resolutions of respectively 2, 5 and 10 cm on the ground). For the second study area very high resolution images of the growing seasons in 2014 and 2015 were made available.
Five different supervised classification methods were tested on their suitability (minimum distance, maximum likelihood, spectral angle mapper, k-nearest neighbour [K-NN] and regression tree). The K-NN was the superior classification method in both study areas, because crop classes were non-normally distributed, overlapping and contained outliers (to which this method is relatively insensitive).
In addition, the accuracy of a classification using a single image was compared to the accuracy using a multi-temporal dataset. To create a multi-temporal dataset vegetation indices were used. For both study areas, spectral bands of the images were combined to a single index without losing relevant information. Therefore, five VIs were tested in this research (NDVI, SAVI, TSAVI, PVI, WDVI). Because the satellite and UAV images were converted to a single band per date, the same classifiers as with the single image classification could be used on the multi-temporal dataset. For Samanko, the study area with the well-managed crops, a single image was sufficient. Crops were separable and, as a result, the overall classification accuracy with mono-temporal classification was 83 percent. However, for the heterogeneous area of Sougoumba a multi-temporal classification was superior (Figure 2). With a multi-temporal dataset spectrally similar crops can distinguish themselves from each other in phenological development. Therefore, the overall accuracy of a multi-temporal K-NN classification using the most suitable soil-adjusted VI (PVI) was 22 percent more accurate than the single image classification.
Figure 2. Classification improvements using the K-NN classifier (and PVI dataset for the temporal classification).
Because crop and landscape variability were expected to pose challenges for the classification of Sougoumba, the effect of a stratification was researched. In short, stratification can create a larger inter-variability between and a smaller within-variability of classes by dividing the study area in a meaningful way. After the division, the separate parts are classified individually. In this study, a stratification of Sougoumba based on soil type was most suitable and increased the overall accuracy from 58 to 63 percent. Mainly because in these soil strata, other landscape characteristics which caused variability were included (elevation and distance to homestead).
To decrease the effect of within-field variability on the classification result, a per-field majority filter was applied. Noise, caused by aberrant crops or weed was decreased and, therefore, the overall accuracy increased. Due to the per-field aggregation it was also possible to detect the misclassified fields in Sougoumba. In these fields, within-field variation was clearly visible. Most likely this was caused by crop-management (e.g. different sowing, harvesting, re-planting and weeding dates/practices). In addition, field-boundaries in the training dataset could be misplaced, as within some fields the multi-temporal dataset clearly indicated different phenological trends (Figure 3 & Figure 4).
- Figure 3. Temporal within-field (Peanut) variability in Sougoumba, visualised with the PVI. The temporal profile of the three markers is visualised in Figure 4. Based on the satellite images of 2015, from left to right; 03-06-2015, 09-07-2015, 03-09-2015, 25-09-2015, 19-10-2015 and 05-11-2015.
At last the influence of different spatial and temporal resolutions was tested. Spatial aggregation proved to be suitable when there are many extreme observations within the classes (overall accuracy increased to 68 percent). Furthermore, decreasing the temporal resolution from eight to five images acquired at different dates was not affecting the classification accuracy. For classification, it is only needed to have imagery at dates at which crops start to deviate from each other. When both optimal resolutions were combined the accuracy slightly decreased, which was caused by the loss of temporal information (a decrease of 1 percent).
With a maximum overall accuracy of 68 percent, the classification method in this research was successful and reasonably accurate. Some small improvements in the training dataset would increase the accuracy even more. Within-field variation on smallholder farms, however, remains a challenge. As the classification accuracy is decreasing with an increasing class variability. Still, crop monitoring is possible in the region of Sougoumba. Moreover, the approach followed by this research is flexible and can be applied to other study areas. For STARS this means that the creation of a transparent information system supporting smallholder farmers in Mali is one step closer.
The full-text version of Wilmar van Ommeren's thesis is not available yet.