Multi-Class Land Detection Analysis: An Approach for Farm, Forest and Rocky Land Classification Using Instance Segmentation from Satellite Images

Land management plays a vital role in the world's environment, economy and agriculture. In this era of dynamic technology, we aim to optimize land use for various purposes. Detecting land types from satellite images helps to utilize farmland effectively during the monsoon season and natural disasters. We introduce a new dataset collected from the Google Maps satellite view. We processed 2,000 images and after preprocessing, 1915 images were selected for research. Following dataset collection and preprocessing, we applied several pre-trained models, including YOLOv5 Segmentation, YOLOv8 Segmentation, Detectron2 and TensorFlow MobileNetV2 Lite. The initial results from YOLOv5 object detection did not meet expectations, as it failed to detect rocky land and missed some areas of farm and forest land. However, after applying Contrast Limited Adaptive Histogram Equalization (CLAHE) to all images, we again deploy YOLOv5 Instance Segmentation, which then successfully detected all three classes. After that, we analyzed the same dataset with the YOLOv8 model, which effectively detected all three classes. Unfortunately, Detectron2 did not provide the expected results for this dataset. Finally, we deploy TensorFlow MobileNetv2 Lite, which achieved 94 % accuracy and accurately detected farm, forest and rocky land individually. The model was able to detect all three classes from the combined annotated images. Overall, the results from YOLOv5, YOLOv8 and Tensorflow MobileNetv2 were satisfac-tory, while Detectron2 did not perform as expected.