Semantic segmentation of landcover for cropland mapping and area estimation using Machine Learning techniques

The paper has focussed on the global landcover for the identification of cropland areas. Population growth and rapid industrialization are somehow disturbing the agricultural lands and eventually the food production needed for human survival. Appropriate agricultural land monitoring requires proper management of land resources. The paper has proposed a method for cropland mapping by semantic segmentation of landcover to identify the cropland boundaries and estimate the cropland areas using machine learning techniques. The process has initially applied various filters to identify the features responsible for detecting the land boundaries through the edge detection process. The images are masked or annotated to produce the ground truth for the label identification of croplands, rivers, buildings, and backgrounds. The selected features are transferred to a machine learning model for the semantic segmentation process. The methodology has applied Random Forest, which has compared to two other techniques, Support Vector Machine and Multilayer perceptron, for the semantic segmentation process. Our dataset is composed of satellite images collected from the QGIS application. The paper has derived the conclusion that Random forest has given the best result for segmenting the image into different regions with 99% training accuracy and 90% test accuracy. The results are cross-validated by computing the Mean IoU and kappa coefficient that shows 93% and 69% score value respectively for Random Forest, found maximum among all. The paper has also calculated the area covered under the different segmented regions. Overall, Random Forest has produced promising results for semantic segmentation of landcover for cropland mapping.