In-Field Crop-Weed Classification Using Remote Sensing and Neural Network
Seyit Kerimkhulle, Zhandos Kerimkulov, Dias Bakhtiyarov, Nazerke Turtayeva, Jong Kim
Abstract
With the development of the latest remote sensing techniques, farmers have more control over their fields, yield forecasting and maximization. In this respect, the classification of weeds is an interesting point for exploration. Since weed growth is one of the main constraints for field productivity, this research aimed to observe a weed spread behavior, develop an in-field crop-weed classification model and test it in the farmlands. Thus, the model is based on the study of multispectral satellite images and the LSTM neural network at two-time intervals of 2 weeks each. It detects weed outbreaks at a radius of 3 to 6 meters with an accuracy of 94-96%. The labelling is based on empirical data of the spectral signature of crops and weeds obtained during a series of field trips to the Akmola region of the Republic of Kazakhstan. These results are a step forward towards the creation of a scalable crop-weed classification model. Direct application can benefit the agricultural sector and economy of the country.