Classification of Horticultural Crops in High Resolution Multispectral Imagery Using Deep Learning Approaches

Abstract

To match the ever increasing nutrition needs of growing population and dwindling agricultural labour force, agriculture sector has been progressively adapting farming practices with major technological interventions. Discrimination of different types of crops is a common requirement for mechanized crop management which includes tasks such as identifying pests, predicting crop yield etc. High resolution remote sensing data has been extensively used for various agricultural applications. Supervised classification has been the method of choice for crop discrimination in high resolution multispectral imagery. Considering the increasing complexity of agricultural lands wherein the co-occurrence of a host of crops and other urban landscape features is a common scenario evolved, classification of high resolution multispectral imagery using traditional supervised approaches is fraught with ambiguities and may lead to site-specific. We have explored the potential of high resolution multispectral satellite imagery (CARTOSAT-2D) for classifying horticultural crops under a complex landscape composition (rice, sugarcane, coconut, arecanut, forest, fallow land, urban, and water body) in Shivmogga region of Karnataka. Based on the selective ground truth data augmentation from implementing two traditional machine learning algorithms, we have adopted two popular deep learning (DL) network architectures: CNN based U-net and Seg net for the regional level classification of crops under complex agricultural settings. Validation of the results with independent ground truth data indicates fairly good accuracy of about 89 percent from the DL methods implemented. However, the accuracy of the results from the Random Forest algorithm is also competitive (about 80 percent) indicating the continued relevance of traditional ML models to crop classification at regional level.