Enhanced Convolutional-Neural-Network Architecture for Crop Classification

Automatic crop identification and monitoring is a key element in enhancing food production processes as well diminishing the related environmental impact. Although several efficient deep learning techniques have emerged field of multispectral imagery analysis, classification problem still needs more accurate solutions. This work introduces competitive methodology for from satellite mainly using an enhanced 2D convolutional neural network (2D-CNN) designed at smaller-scale architecture, novel post-processing step. The proposed contains four steps: image stacking, patch extraction, model design (based on 2D-CNN architecture), post-processing. First, images are stacked to increase number features. Second, input split into patches fed model. Then, constructed within small-scale framework, properly trained recognize 10 different types crops. Finally, step performed order reduce error caused by lower-spatial-resolution images. Experiments were carried over so-named Campo Verde database, which consists set captured Landsat Sentinel satellites municipality Verde, Brazil. In contrast maximum accuracy values reached remarkable works reported literature (amounting overall about 81%, f1 score 75.89%, average 73.35%), achieves 81.20%, 88.72% when classifying crops, while ensuring adequate trade-off between multiply-accumulate operations (MACs) accuracy. Furthermore, given its ability effectively classify two sequences, this may result appealing other real-world applications, such urban materials.