Crop selection is one of the most important decisions in agriculture. With yield, crop decay, and income of the farmer down the line, payment for services becomes the paramount issue of choice. In earlier months of farming, the basis for crop choice would include a farmer's intuition and his knowledge of the farm. Reliability and accuracy issues arose because of climate change, soil health, and also market fluctuations. With the recent advancements in machine learning (ML) and artificial intelligence (AI), it has been possible to establish data-driven crop recommendation systems (CRSs). The crop recommendation systems used combinations of many input parameters, including but not limited to, soil nutrients (N, P, K), soil pH, temperature, humidity, rainfall, and few socio-economic indicators (e.g., Minimum Support Price). This review puts forward in view the data-driven CRSs concerning literature developed after the year 2022, which incorporates ML therein. The literature contains those approaches using classical ML algorithm(s), ensemble model(s), and DL model architectures. Furthermore, we investigated the sources of data that were used in existing crop recommendation systems, e.g., soil and climate databases; sensor networks based on.Internet of Things; remote sensing pictures; and market databases. As per the comparative study of the literature, these studies illustrated how, in a few case examples, ensemble models and hybrid models were far more accurate in prediction compared to machine learning and deep learning models. Also, in the case of multi-modal and big datasets, deep learning architecture brought in its own share of merit. Many challenges still remain in the promising field of ML-based CRS research despite all the opportunities that definitely exist; for example, data availability; generalizable geo-regions; black-box interpretability trustworthiness; and trustworthiness by end-users (farmers). The paper ends by considering directions that may include possible solutions with Explainable AI (XAI), AutoML, federated learning, blockchain security, and synthetic data generation towards achieving a robust and scalable CRS.