Skip to content Skip to sidebar Skip to footer

We finished 2018 with an article on precision agriculture and therefore it only seems appropriate to launch the new year with another piece on precision agriculture. Only this time we will be looking between the rows (excuse the pun). The paper from de Castro et al entitled “An Automatic Random Forest-OBIA Algorithm for Early Weed Mapping between and within Crop Rows Using UAV Imagery” looks at the automatic detection of weeds in various agricultural fields. I could not have found this at a better time since I have been discussing the use of Trimble eCognition in combination with UAVs in the agricultural sector with several customers. The work by de Castro et al demonstrates a great use of eCognition in the fast paced UAV data acquisition world. The study’s main objective “was the early season weed detection between and within rows in herbaceous crops, for the purpose of designing an in-season early post-emergence prescription map”. To tackle this problem, the authors developed an OBIA workflow based on a Random Forest (RF) classifier in combination with UAV-derived height (DSM) inputs.

Five study areas in southern Spain were selected for the study, 3 sunflower fields and 2 cotton fields. A key factor in the UAV-based data acquisition was the flexibility to fly the fields while the crops were in the optimal growth stages – all fields were captured while “in the 17, 18 growth stage of the BBCH  Biologische Bundesantalt, Bundessortenamt and Chemische Industrie) scale” and contained varying weed species.

View of 3 sunflower crops in early growth stages: a) S1-16, b) S1-17, c) S1-17

The input data was acquired with a microdrone MD4-1000 from 30 m and 60 m ground heights. The resulting orthomosaics had 0.6 cm and 1.2 cm respectively. DSMs were derived for each data set with resolutions of 1.2 cm (30 m) and 2.4 cm (60 m). Both data products were then combined within the Trimble eCognition Developer software.
The first step in the analysis within eCognition was the generation of a DTM layer. This was then combined with the spectral data (Red, Green, NIR) contained in the orthomosaic for segmentation. Subsequently, an nDSM was calculated by subtracting the DTM from the DSM giving the authors valuable information on object’s height above ground.

Shadow objects were then classified using an overall intensity threshold and removed from the subsequent analysis. Then, the image objects were split into 2 classes: vegetation and bare soil. Once separated, a crop row detection was performed; “a merging operation was performed to create lengthwise vegetation objects following the shape of a crop row. In this operation, two candidate vegetation objects were merged only if the length/width ratio of the target object increased after the merging” providing the bases for a seed and grow approach.

After establishing the vegetation objects, the rule set searched for crops and weeds within the rows, assuming that any vegetation growing outside the row is weed. Vegetation within the rows was classified as either crop or weed based on whether the vegetation had an “above average height” – such objects were then pre-classified as crop. According to the authors, “the average height of the crop row was shown to be a suitable feature for sample selection, as most weeds growing within the crop row in this scenario are generally smaller than the crop plants”.

The results of the pre-classification where then fed into a RF classifier model. A random sampling process was implemented and the training samples were then passed through the algorithm to train and classify the image. Subsequently, a classification enhancement was performed.

The percent weed coverage and weed density was calculated for each field. Better correlations were achieved with the 30 m flight data due to the improved resolution. “The OBIA algorithm was able to identify every individual plant in the field and estimate the plant height feature of each one”. The use of DSM data was critical for the analysis and feasible due to the amount of overlap between images.

Classification results for a) sunflower and b) cotton field

The authors then developed a series of prescription maps for weed eradication or treatment based on a presence of weed threshold – “any grid with a number of weeds equal to or higher than 1 was considered a treatment area” regardless of the weed size. The prescription maps can be used by the farmers to optimize their operational costs. The potential herbicide savings based on the prescription maps, “calculated in terms of untreated areas” were significant for cotton, 60 to 79% and between 27 and 37% for sunflower. I am curious if this approach can be applied to other crop types – possibly corn. I hope this inspires some of you to go out and try!

Follow Us on Social Media

For more news on eCognition