Crop Mapping Using Advanced Deep Learning Framework in Remote Sensing Data

Abstract

An essential step in the agricultural monitoring system is identifying the crop type. However, the limiting physical features of images, as well as the lack of sufficient detail in a single-temporal image, limit their potential for crop mapping. Furthermore, time-series Synthetic Aperture Radar (SAR) data may not be compatible with the existing approaches. Therefore, a novel approach to crop-type identification needs to be put forth in order to solve the aforementioned problems. To this extent, the presented proposed framework introduces a deep learning (DL) technique using the advanced framework for accurate crop mapping. At first, the input dataset is pre-processed based on data mining techniques like normalization and cleaning. Afterward, Bidirectional Gated Auto encoders (BiGAE) features are extracted from the pre-processed data. Consequently, feature selection uses an opposition learning-based mud ring algorithm (Opp-MR) to reduce the redundant data. Then, the selected data are clustered based on relevance using an adaptive Kernel fuzzy clustering (AkFC) technique. Finally, the goshawk integrated convolutional attention efficient net (GICANet) is performed using an advanced DL framework to map crops accurately. The performances are evaluated using the Python simulation platform. The proposed method improves the overall accuracy by 97.74%, whereas the existing models like ResNet, DenseNet, EfficientNetB0, EfficientNetB5, and EfficientNetB7 have obtained only lesser performance. The proposed GICANet classifier outperforms the other approaches utilizing two error metrics, RMSE (0.5) and MAE (1.5).