Crop Type Identification and Crop Area Prediction Accuracy Using Convolutional Neural Network Model in Three States of Northwest Nigeria
Accurate crop identification and yield estimation are vital for effective agricultural planning and food security in Nigeria. Traditional statistical approaches often fall short due to data inconsistency, non-linearity, and limited scalability. This study applies Convolutional Neural Networks (CNNs) to detect three major crops maize, rice, and soybean using satellite imagery, UAV data, and ground-truth observations across Kano, Kaduna, and Katsina states during the 2023 cropping season. 960 farmers per crop were sampled from 96 communities, each contributing a 0.153-hectare plot. Data was collected at four crop growth stages and processed into tensors for CNN model training. Results showed variation in detection accuracy across six LGAs. Karaye had the best model performance, where we observed the lowest mis-detection incidence rate (14.3%) and highest detection accuracy for soybean (61.0%) and rice (54.1%). In contrast, maize was consistently more difficult to detect, with misdetection rates reaching 34.8% in some of the LGAs. The pooled model recorded a moderate error rate of 21.6%, highlighting the general feasibility of the CNN approach. Key challenges included detection imbalance, spectral overlap among crops, and ecological complexity. Despite these, the CNN model demonstrated strong potential for scalable, image-based crop monitoring, primarily when supported by localized high-quality field data. The study recommends enhancing model performance through improved data balance, higher-resolution imagery, synchronized image capture with crop growth stages, and location-specific model training. These measures can significantly improve detection accuracy and support the integration of machine learning tools into Nigeria’s agricultural monitoring systems.
