Table 3.

Qualitative comparison of smart environment-related works.

	Use Case	Ref	Contribution	AI Role (At the Edge)	AI Algorithm	Dataset	AI Placement	Employed Technology	Platform	Metrics	Benefits AI-Edge	Drawbacks
Smart environment	AQM	[20]	Predicting of futureindoor status of PM10 and PM2.5	Prediction	LSTM	Data from Seoul, Korea	Edge device, cloud	Federated learning	TensorFlowKeras	RMSE	Minimize load Hight accuracy	Does not consider all factors in prediction
		[21]	Green energy-based wireless sensing network for air-quality monitoring	Prediction	LSTM	Airbox system dataset	Edge device, cloud	Federated learning	Not mentioned	MAE-loss RMSEEnergy thresholdsaving, ratio error rate	Communication efficiencyPreserving data privacy Low computational complexity	Slightly lower accuracy
		[22]	Location awareenvironment sensing	Prediction	k-means, LSTM, CNN (ResNet)	WA dataset Outdoor image datasets	Edge device, cloud	Distributedcomputing cluster	Federated learning	Accuracy, avg. sum of squared errors, silhouette coefficient	High accuracy	Homogeneous nodes only considered
		[23]	Distributed data analysis for air prediction	Preprocessing	K-means SVM, MLP, DT, KNN, NB	U.S. Pollution Data Kaggle	Edge devices, cloud	Distributed computing	IFogSim toolkit-YAFS-	AccuracyPrecision recallF1-Score	Data reductionLow response time reduction	Not consider mobility of nodes
		[24]	On-device air-quality prediction	Prediction	CNN, LSTM	Dataset from University of California–Irvine (UCI) Machine Learning Repository page	Edge devices(RPi3B+, RPi4B)	Posttraining quantization Hardware accelerator	TensorFlow Lite	RMSE, MAEexecution time	Low-complexity model latency	Accuracy degradation
	WQM	[25]	Onboard sensor classifier for the detection of contaminants in water	Classification	EA PCA	Real-world dataset	Edge device (sensors)	Low-cost model	Not mentioned	Accuracy F-score TP TN FP FN	High accuracy	Low accuracy for unlabeled data
		[27]	Online water-quality monitoring	Prediction	BPNN	Real-world dataset	Edge gateway	Low-cost model	Not mentioned	Data transmission response time	Low-complexity model accuracy, data transmission reduction	Accuracy needs to improved
	WQM	[26]	Real-time water- quality monitoring	Preprocessing prediction	PCA LR MLP SVM SMO Lazy-IBK, KStar RF RT	Data of sewage water-treatment plant of the institute, data collected from river Ganga	Edge device (Raspberry Pi)	Transfer learning	Python, Weka	Correlation coefficient MAE RMSE-RAE RRSE Edge response time	Less response time	Communication cost not considered
	SWM	[28]	Smart water saving and distribution	PredictionDecision making	FFN MDN	Real-world dataset	Edge server	SofT computing blockchain	Python	MSE accuracy	Effective decision-making	Accuracy needs to be enhanced
Smart environment	UM	[29]	Reduce data and improve data quality or underwater	Data (fusion, reduction)	BPNN evidence theory	Western Pacific measurement information	Fog gateway Cloud	Edge preprocessing	Not mentioned	Time consumption Redundant data volume R, MAE, MSE SMAPE	Low communication costHigh accuracy	High delay
		[30]	Real anomaly detection errors in underwater vehicles	Network management, data reduction classification, decision-making	YULO (CNN), RL	Real-world dataset	Edge device (Raspberry Pi) Fog gateway	Hardware accelerator, pretrained CNN	Not mentioned	Accuracy, latency, recall	High accuracy, less latency	Accuracy degraded
		[31]	Low delay for Seawater quality prediction	Data reductionPrediction	PCA RVM	Real-world dataset	Mobile edge computing	Low-cost model	Not mentioned	CD MAE RMSE	Higher prediction Low time consumption	High-cost model
		[32]	Downlink throughput performance enhancement	Resource allocation Classification	DRL DNN	Real-world dataset	Edge device (IoUT devices)	Federated learning	Not mentioned	Downlink throughput channel usage Convergence rate	Low complexity	–