Towards A Powered AI GIS Platform for Agricultural Implantation Identification

Abstract

With the expansion of agricultural areas and the diversity of crops, manually identifying and precisely locating these areas has become increasingly challenging and time-consuming. In this work, we propose a technique that combines the power of Deep Learning, specifically Convolutional Neural Networks (CNNS or ConvNet) for identifying agricultural areas using aerial imagery, and Geographic Information System (GIS) for spatial mapping. In this study, we use a dataset of scene images collected from different remote imaging sensors, and apply appropriate preprocessing techniques to improve the data quality. We trained four models on the preprocessed dataset, including our convolutional neural network (CNN) model that was built from scratch, and leveraging the power of transfer learning, we adapted the pretrained VGG16, DenseNet121, and ResNet50 models on the same dataset to establish the most appropriate model for our task. The pre-trained VGG16 model demonstrated the highest performance with an accuracy of 96%, followed by DenseNet121 at 94%. Our CNN model achieved an accuracy of 89%, whereas ResNet50 recorded the lowest accuracy of 62%. Then using GIS techniques, we developed a desktop application to efficiently classify and visualize geotagged images in the Assafia region map, with a particular focus on farmlands. This application leverages the predictive results of our customized VGG16 model, chosen for its exceptional performance in our specific task. Our work demonstrates the effectiveness of transfer learning techniques, particularly the VGG16 model, for accurate and timely land type classification. Additionally, it underscores the utility of GIS techniques in visualizing spatial data effectively.