Finding seasonal crops

Finding 1 crop versus many by Naturinda Evet

According to research, quantification of crop acreage has been highly dependent on the use of machine learning algorithms and training data. Here at Fieldy, we have tried to explore the use of information on the crop phenology patterns to quantify crop acreage. Crop phenology refers to the physiological development stages of crop growth from the planting time to harvest time. This is especially helpful for areas with less or no training data or for application on a large scale. This was applied in rice mapping and identification of seasonal crop land.

SAR for rice

Rice mapping was carried out using Synthetic-Aperture Radar (SAR) data. SAR data represents the backscattered signals reflected from the earth's surface. The backscatter is affected by the size and structure of the objects such as vegetation on the earth's surface and the moisture content. For example, when the SAR signals from a sea are compared with those from a vegetated area, the signals reflected from the sea will be much lower than those reflected from the vegetated area.

Basing on this information, we decided to use SAR data to identify rice fields in Bhola, Bangladesh. We downloaded Sentinel 1 images for each month in the rice-growing season that is February, March and April 2021 and studied the temporal behaviour of the time-series signals. 

It was observed that the rice plots have a different temporal behaviour for the SAR signals compared to other crops. This is because rice grows in flooded areas, it’s usually covered by water in the early planting stages and as it grows, the leaves grow longer and bigger and cover water. Maximum, minimum, and the difference statistical values were calculated for each time series and used to identify the rice fields. An area of 89.111 SQKM was identified as rice fields for the year 2021 in the growing season of February to April.

Finding cropland

The identification of seasonal crop land was greatly enthused by the devastating impacts of COVID 19 on many sectors in different countries as one of the major affected sectors was the agricultural sector. The measures that were put in place to contain the spread of the virus led to less productivity which disrupted the growing and supply of agricultural products especially for seasonal crops. As Fieldy, we decided to develop a method to quantify the acreage of seasonal crops. This was done using Normalised Difference Vegetation Index (NDVI). NDVI is a graphical indicator that represents vegetation health based on how plants reflect particular ranges of the electromagnetic spectrum like near infrared. NDVI can be used in the study of the spatial temporal changes in the vegetation health.

An NDVI based method was developed to detect seasonal crops using freely available Sentinel 2 images. This method majorly uses NDVI thresholds and the cropping pattern of seasonal crops to quantify crop acreage. This is based on the fact that seasonal crops are grown and harvested a number of times in a year and therefore there is a high deviation in the NDVI values compared for the year. A wetland mask is also used to extract out all the wetland areas from the study area. This method is advantageous in a way that the method can be applied in different areas and at a large scale. It was tested in a 14,681sqkm area in Uganda and it achieved a performance of 67.90% accuracy with maize having the highest accuracy of 71.09%. However, the thresholds used were gotten basing on maize fields’ data therefore an increase in accuracy could be achieved if other types of seasonal crops are put into consideration. One of the limitations of this method is the unavailability of all images for each month in the year because of the cloud cover that affects Sentinel 2 images. This could be solved by exploring the use of Sentinel 1 data to apply this method.