Integrated Novelty Detection Systems: From Sensor Networks to Multi-Novelty Computer Vision

Irené Tematelewo

cs.LGOct 14, 2026

Integrated Novelty Detection Systems: From Sensor Networks to Multi-Novelty Computer Vision

Authors: Irené Tematelewo

Organizations: Oregon State University

Abstract

Novelty detection is crucial in various technological and scientific domains. Its importance spans from ensuring the reliability of sensor networks to enhancing the adaptability of computer vision systems. As AI systems become increasingly sophisticated and complex, effective anomaly detection is essential for maintaining reliability and accuracy. This work addresses novelty detection in weather sensor networks and computer vision. In weather sensor networks, particularly for monitoring rainfall, detecting blocked rain gauges can significantly impact data reliability, food security, and disaster management. This work evaluates four existing quality control methods for detecting blocked rain gauges and demonstrates their limitations, particularly in detecting clusters of faulty sensors. By comparing methods such as the Conditional Mixture Model (CMM), PSWQ (Fz-filter), Kernel Regression, and Gaussian Method, we identify their shortcomings and propose three novel quality control techniques that integrate probabilistic fault detection and consistency-based diagnosis framework to improve timely detection of blocked sensors, while reducing false alarms. In the realm of computer vision, current systems often fail to handle multiple types of novelty due to their focus on single-novelty scenarios. We highlight the need for integrated systems capable of handling multiple types of novelty. To support this, we introduce the Novel Snapshot Serengeti (NOVEL-SS) dataset, derived from the Snapshot Serengeti camera trap data, which includes various forms of novelty. We also present the DCA (Detect, Characterize, and Accommodate) baseline system, designed to handle multiple novelty by detecting, characterizing, and accommodating them. This work aims to advance the robustness and adaptability of novelty detection systems across both weather sensor networks and computer vision applications.

Figures & tables

Figure 1: Correlation coefficient (rho, orange) between precipitation readings at the CHIC and NINN stations (the closest pair of stations) in the Oklahoma Mesonet as a function of the length of time over which precipitation is aggregated. The blue curve shows the fraction of missing values.

Figure 2: Historical training and testing rainfall distribution in the Oklahoma Mesonet and the Kenya dataset.

Figure 3: Geographical distribution of the 23 stations in the Kenya test dataset.

Figure 4: Geographical distribution of stations in the Oklahoma Mesonet test dataset.

Figure 5: Definitions of TP, FP, FN, and TN. The rectangles indicate time intervals during the year. During green intervals, the station is working properly. During red intervals, it is faulty. The arrows indicate time points at which a QC method might introduce an alarm. The two arrows in the dashed box at right indicate the status if no alarm is declared during the year.

Panel of Figure 6 — Figure 6: Detection performance on the TAHMO Kenya and Oklahoma networks with 5% and 10% faulty rain gauges that are spatially random. Each point plots the FAR and average detection delay for a choice of $z_{min}$ that achieves 5%MAR. Shapes correspond to different methods; colors correspond to different values of $z_{min}$ . Yellow shading highlights the region of 0-20% FAR. Plots in the left column show performance on the TAHMO Kenya network, and plots in the right column show performance on the Oklahoma network. Rows from top to bottom represent 5% and 10% faulty stations.

Panel of Figure 7 — Figure 7: Detection performance on the TAHMO Kenya and Oklahoma networks with 15% and 20% faulty rain gauges that are spatially random. Plots in the left column show performance on the TAHMO Kenya network and plots in the right column show performance on the Oklahoma network. Rows from top to bottom represent 15% and 20% faults.

Panel of Figure 8 — Figure 8: AMOC and FAR-MAR trade-off of algorithm performance for the TAHMO Kenya and Oklahoma networks with 5% faulty rain gauges that are spatially random and 3 days of evidence. Plots in the left column show performance on the TAHMO Kenya network and plots in the right column show performance on the Oklahoma network.

Faults	FZ ok	CMM ok	KR ok	GM ok	FZ ke	CMM ke	KR ke	GM ke
5%	50	10	16	16	130	20	70	75
10%	55	12	13	10	160	50	70	112
15%	55	12	15	10	–	25	34	–
20%	55	12	15	12	–	30	60	62

Table 1: Lowest detection delay at 5%-20% faults (FAR

\leq

0.20, MAR

\leq

0.05) for Oklahoma and Kenya networks. The boldface and underline highlight the best and the second best performing QC methods.

Delay	FZ ok	CMM ok	KR ok	GM ok	FZ ke	CMM ke	KR ke	GM ke
15	0.65	0.15	0.50	0.60	0.97	–	0.78	–
30	0.22	0.00	0.03	0.45	0.87	0.15	0.45	–
45	0.17	0.00	0.03	0.22	0.30	0.15	0.45	–
60	0.03	0.00	0.00	0.07	0.25	0.15	0.15	0.45
75	0.01	0.00	0.00	0.04	0.21	0.15	0.15	0.40
90	0.00	0.00	0.00	0.03	0.20	0.07	0.15	0.40

Table 2: Lowest false alarm rate at 5% faults, 0.05 MAR and selected average Detection Delays. Oklahoma and Kenya networks are noted by the column name subscript

Panel of Figure 9 — Figure 9: Effect of more distant neighbors on the TAHMO network. The plots on the left column top to bottom show the performance when the neighbors are chosen with minimum distance to the target station of 0 km, 5 km and 10 km, and the plots on the right (top to bottom) show performance when the neighbors are at a minimum distance of 15 km, 30 km and 50 km. Fault rate is $15\%$ ; spatially randomly distributed.

Panel of Figure 10 — Figure 10: Effect of more distant neighbors on Mesonet. The plots on the left column top to bottom show the performance when the neighbors are chosen with minimum distance to the target station of 0 km, 5 km and 10 km, and the plots on the right (top to bottom) show performance when the neighbors are at a minimum distance of 15 km, 30 km and 50 km. Fault rate is $15\%$ ; spatially randomly distributed.

Method	$0$ km	$5$ km	$10$ km	$15$ km	$30$ km	$50$ km
GM ok	66	75	78	-	-	-
KR ok	62	68	82	82	103	103
CMM ok	27	46	74	75	75	78
Fz-filter ok	-	-	-	-	-	-
GM ke	118	-	-	-	-	-
KR ke	41	73	73	73	-	-

Table 3: Average detection delay in the Oklahoma Mesonet and the TAHMO Kenya networks for increasingly distant neighbors (

d_{min}=

0km, 5km, 10km, 15km, 30km and 50km) at a maximum MAR of 5%, and a maximum FAR of 20%. The best performing QC method in boldface.

Panel of Figure 11 — Figure 11: Effect of spatially-clustered faults on detection performance for the TAHMO Kenya network with $15\%$ faulty stations. The plots on the left column show the performance when the cluster of faults are inside the area (SCF-IN) and the plots on the right are when some stations are at the periphery (SCF-OUT).

Panel of Figure 12 — Figure 12: Effect of spatially-clustered faults on the detection performance for the Oklahoma Mesonet network with $15\%$ faulty stations. The plots on the left column show the performance when the cluster of faults are inside the area (SCF-IN) and the plots on the right are when some stations are at the periphery (SCF-OUT).

	FAR $\leq 20\%$							FAR $\leq 10\%$
Method	2	3	4	5	6	7	8	2	3	4	5	6	7	8
GM IN	68	57	76	52	60	55	54	68	57	76	52	60	55	54
KR IN	84	76	88	70	77	83	74	84	76	88	70	77	83	74
CMM IN	50	37	37	38	-	-	-	50	37	67	46	-	-	-
Fz-filter IN	-	-	-	-	-	-	-	-	-	-	-	-	-	-
GM OUT	65	53	65	55	48	55	67	70	75	65	62	69	72	67

Table 4: Average detection delay at a maximum MAR of 5%, and a maximum FAR of 20% for spatially clustered faults SCF-IN and SCF-OUT in the Oklahoma network. The columns show the cluster size and the rows are the QC methods. The boldface and underline highlight the best and the second best performing QC methods.

	FAR $\leq 20\%$				FAR $\leq 10\%$				FAR $\leq 5\%$
Method	2	3	4	5	2	3	4	5	2	3	4	5
GM IN	81	93	49	104	137	93	113	161	-	-	-
KR IN	65	65	64	87	65	73	64	102	71	-	83	102
CMM IN	45	63	-	-	70	63	-	-	-	-	-	-
Fz-filter IN	-	-	-	-	-	-	-	-	-	-
GM OUT	108	-	-	-	108	-	-	-	-	-	-	-

Table 5: Average detection delay at a maximum MAR of 5%, and a maximum FAR of 20% for spatially clustered faults SCF-IN and SCF-OUT in the TAHMO Kenya network. The columns show the cluster size and the rows are the QC methods. The boldface and underline highlight the best and the second best performing QC methods.

Panel of Figure 13 — Figure 13: AMOC and FAR-MAR trade-off of algorithm performance for the Oklahoma network with 15% faulty rain gauges that are spatially random, neighbors $\leq$ 30 km are excluded, for two and three days of evidence. Plots in the left column show performance with two days of evidence and and plots in the right column show performance with three days of evidence

Panel of Figure 14 — Figure 14: AMOC and FAR-MAR trade-off of algorithm performance for the Kenya network with 15% faulty rain gauges that are spatially clustered with size=2 in the interior of the network, for two days of evidence.

Panel of Figure 15 — Figure 15: AMOC and FAR-MAR trade-off of algorithm performance TAHMO Kenya network. The left hand plots represent 5% faulty rain gauges that are spatially random with four days of evidence. The right hand plots represent 20% faulty rain gauges that are spatially random with five days of evidence.

Day	Rain S a	Rain S b	Rain S c	Rain S d	Rain S e
$1$	$0.0$	$0.0$	$5.1$	$2.4$	$3.3$
$2$	$0.0$	$0.0$	$0.0$	$0.0$	$0.0$
$3$	$0.0$	$0.0$	$3.5$	$6.2$	$7.0$
$4$	$0.0$	$0.0$	$2.0$	$4.1$	$5.0$
$5$	$0.0$	$0.0$	$4.0$	$5.5$	$1.0$

Table 2: Illustrative artificial rainfall (mm) observed over the sensor network in Figure 3

Figure 4: Illustrative Bayesian network.

Table 3 : Truth table of S a $\bm{\text{b}_{a}}$ $\text{b}_{b}$ $\text{b}_{c}$ consist. 0 0 0 0 0 0 1 1 0 1 0 0 0 1 1 1 1 $\ast$ $\ast$ 1	Table 4 : Truth table of S b $\bm{\text{b}_{b}}$ $\text{b}_{a}$ $\text{b}_{d}$ consist. 0 0 0 0 0 0 1 1 0 1 0 0 0 1 1 1 1 $\ast$ $\ast$ 1	Table 5 : Truth table of S c $\bm{\text{b}_{c}}$ $\text{b}_{a}$ $\text{b}_{e}$ consist. 0 0 0 0 0 0 1 0 0 1 0 1 0 1 1 1 1 $\ast$ $\ast$ 1
Table 6 : Truth table of S d $\bm{\text{b}_{d}}$ $\text{b}_{a}$ $\text{b}_{b}$ consist. 0 0 0 0 0 0 1 0 0 1 0 0 0 1 1 1 1 $\ast$ $\ast$ 1	Table 7 : Truth table of S e $\bm{\text{b}_{e}}$ $\text{b}_{c}$ $\text{b}_{d}$ consist. 0 0 0 1 0 0 1 1 0 1 0 1 0 1 1 1 1 $\ast$ $\ast$ 1

Table 8: Truth tables. The first column of each subtable is the Boolean state of the target sensor. The symbol

\ast

means any value in {0,1}, and the “consist.” column is the consistency of the target with those of its neighbors having a Boolean state of 0.

Figure 5: Historical training and testing rainfall distribution in the Oklahoma Mesonet and the Kenya dataset.

Method	FAR $\leq 5\%$	FAR $\leq 10\%$	FAR $\leq 15\%$	FAR $\leq 20\%$
PCSD-SAT	-	58	58	58
PCSD-JIGM	44	44	44	44
PCSD-JIKR	36	36	36	36
GM	66	66	66	66
KR	63	63	63	63
CMM	42	23	23	23

Table 9: Average detection delay in the Oklahoma Mesonet network at a maximum MAR of 5%, and a maximum FAR of 5%, 10%, 15%, and 20%. The best performing QC method in boldface.

Method	FAR $\leq 5\%$	FAR $\leq 10\%$	FAR $\leq 15\%$	FAR $\leq 20\%$
PCSD-SAT	-	-	-	-
PCSD-JIGM	48	48	48	39
PCSD-JIKR	71	67	48
GM	-	-	-	118
KR	51	51	41	41
CMM	70	51	38	38

Table 10: Average detection delay in the TAHMO Kenya network at a maximum MAR of 5%, and a maximum FAR of 5%, 10%, 15%, and 20%. The best performing QC method in boldface.

	FAR $\leq 20\%$							FAR $\leq 10\%$
Method	2	3	4	5	6	7	8	2	3	4	5	6	7	8
PCSD-SAT IN	-	-	-	-	-	-	-	-	-	-	-	-	-	-
PCSD-JIGM IN	71	67	48	44	47	48	42	71	67	71	44	47	48	42
PCSD-JIKR IN	56	43	73	45	53	33	35	56	43	73	45	99	33	70
GM IN	68	57	76	52	60	55	54	68	57	76	52	60	55	54
KR IN	84	76	88	70	77	83	74	84	76	88	70	77	83	74

Table 11: Average detection delay at a maximum MAR of 5%, and a maximum FAR of 20% for spatially clustered faults SCF-IN and SCF-OUT in the Oklahoma network. The columns show the cluster size and the rows are the QC methods. The boldface and underline highlight the best and the second best performing QC methods.

	FAR $\leq 20\%$				FAR $\leq 10\%$				FAR $\leq 5\%$
Method	2	3	4	5	2	3	4	5	2	3	4	5
PCSD-SAT IN	-	-	-	-	-	-	-	-	-	-	-	-
PCSD-JIGM IN	47	75	75	128	47	93	75	-	47	87	83	140
PCSD-JIKR IN	61	62	63	-	67	88	63	140	-	-	-	-
GM IN	81	93	49	104	137	93	113	161	-	-	-
KR IN	65	65	64	87	65	73	64	102	71	-	83	102

Table 12: Average detection delay at a maximum MAR of 5%, and a maximum FAR of 20% for spatially clustered faults SCF-IN and SCF-OUT in the TAHMO Kenya network. The columns show the cluster size and the rows are the QC methods. The boldface and underline highlight the best and the second best performing QC methods.

Figure 1: Example no-novelty and novelty images from the Novel-SS dataset.

Figure 2: Schematic diagram of the training and execution phases of the Novel-SS task.

Novelty types	Categories
Known species	Elephant, Gazelle grants, Gazelle Thomsons, Giraffe, Hyena Spotted, Lion female, Other bird, Topi, Wildebeest, Zebra
Novel species	Baboon, Buffalo, Cheetah, Dikdik, Eland, Guineafowl, Hartebeest, Impala, Jackal, Koribustard, Mongoose, Ostrich, Reedbuck, Warthog, InsectSpider
Known activities	moving, standing
Novel activities	eating, resting
Known envs	daytime
Novel envs	dawn/dusk, night time, fog, snow

Table 1: Known and novel species, activities and environments.

Figure 3: Architecture of the DCA-SS system. Initial training is not shown. The network iterates over each bounding box in the image. The resulting predictions for species and activity are aggregated by probabilistic inference to produce predictions for the species counts, species presence, and activity presence tasks.

Trial Type	Queries	NU	DCA	NU	DCA	NU	DCA	NU	DCA
		Animal Counts				Species Presence
	DCA	(MAE $\downarrow$ )				( $F_{1}$ $\uparrow$ )
	Feedback	Known		Novel		Known		Novel
No Novelty	570	0.28	0.34	–	–	0.83	0.84	–	–
Novel Species	720	0.30	0.40	2.58	2.56	0.84	0.88	0.05	0.07
Species Combination	735	0.28	0.44	7.64	2.45	0.85	0.82	0.46	0.60

Table 2: Animal counts and species presence

F_{1}

during the test phase. The best value for each trial/metric is in boldface.

	DCA	Activity Presence ( $F_{1}$ $\uparrow$ )
	Feedback	Known		Novel
Trial Type	Queries	NU	DCA	NU	DCA
No Novelty	570	0.95	0.88	–	–
Novel Activities	735	0.87	0.86	0.17	0.19
Activity Combination	0	0.86	0.86	0.68	0.70
Dawn/Dusk	735	0.85	0.85	0.67	0.69

Table 3: Activity presence

F_{1}

during the test phase. The best value for each trial/metric is in boldface.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

		no-novelty	Novel species	Novel comb species	Novel activities	Novel comb activities	Novel env
Train	spe	all known	-	-	-	-	-
	act	{ St } or { Mo }	-	-	-	-	-
	env	{ day }	-	-	-	-	-
Valid	spe	-	{ Ko } or { Ela }	{ ( Ze, GT ) , ( Ze, Gi ) , ( OB, Wi ) , ( OB, GT ) }	{ Wi } or { To }	{ El }, { Wi } or { To }	{ Ze } or { Wi }
	act	-	{ St } or { Mo }	{ St } or { Mo }	{ Ea }	{ St } and { Mo }	{ St } or { Mo }
	env	-	{ day }	{ day }	{ day }	{ day }	{ fog, snow, dawn/dusk }

Appendix

Table 1: Summary of species (spe), activities (act) and environment (env) in the training, validation and test data sets of the Novel-SS. Notation {(A, B)} means that A and B are simultaneously present in the image while {A, B} means that at least one of A and B is present. The red color mean the species, activity or environment is novel. Species abbreviation : Ko =Koribustard, El =Elephant, Ela =Eland, Ze =Zebra, Gi =Giraffe, Wi =Wildebeest, To =Topi, GT =Gazelle Thomsons, OB =Other bird. Activities abbreviation : St =standing, Mo =moving, Ea =eating, Re =resting.

Figure 1: Distribution of bounding boxes counts in the Novel-SS.

Panel of Figure 2 — Figure 2: Images illustrating the behavior of the Megadetector.

Panel of Figure 3 — Figure 3: Statistics of images in the Novel-SS.

Panel of Figure 4 — Figure 4: Statistics of images in the Novel-SS (cont.).

Species	Zebra	Wildebeest	Gazelle Thom.	Gazelle Grants	Other bird	Topi	Giraffe	Elephant	Hyena spotted	Lion female
Zebra		547 ; 727	42 ; 55	12 ; 4	25 ; 32	18 ; 9	21 ; 19	8 ; 6	0 ; 4	0 ; 2
Wildebeest	547 ; 727		20 ; 11	11 ; 10	19 ; 25		4 ; 2	4 ; 0	0 ; 1	0 ; 3
Gazelle Thom.	42 ; 55	20 ; 11		50 ; 12	109 ; 17	10 ; 11	6 ; 6	0 ; 3	3 ; 0
Gazelle Grants	12 ; 4	11 ; 10	50 ; 12		9 ; 3	2 ; 0	0 ; 3
Other bird	25 ; 32	19 ; 25	109 ; 17	9 ; 3		3 ; 0	1 ; 8	2 ; 7	2 ; 3
Topi	18 ; 9		10 ; 11	2 ; 0	3 ; 0

Appendix

Table 2: Validation and test image counts for novel combination of known species novelty type. The cells with yellow background are the combination of species in the validation set and the ones with white background are the test set. The red colored numbers are the counts of images with animals engaged in the activity ” standing ” and the blue ones are those with animals engaged in the activity ” moving ”. Four validation images (not included in the table) contain 3 species: 2 of them contain the triplet (giraffe, wildebeest, zebra) and 2 contain the triplet (gazelle Thomsons, gazelle Grants, zebra). The test set on the other hand contains only 1 image with 3 species (other bird, wildebeest and zebra).

Panel of Table 3 — Table 3: Validation and test image counts for novel activities novelty type. The cells with yellow background are the species and activities in the validation set and the ones with white background are the test set. The red colored activities are the novel activities and the lack ones are the known activities.

Figure 5: Comparison of Mean Average Error for Selected Trials of Various Continual Learning Configurations

Figure 6: Comparison of Species Presence

F_{1}

for Selected Trials of Various Continual Learning Configurations

Figure 7: Comparison of Activity Presence

F_{1}

for Selected Trials of Various Continual Learning Configurations

Fault Detection

Fault detection identifies abnormal behavior in machines, sensors, infrastructure, software, and autonomous systems before it causes harm or disruption.

See all 105 papers

A Sensor-Adaptive Incremental Learning Framework for Artifact Detection in Satellite Precipitation Data
Sep 1, 2026Andres F. Monsalve, Hernan A. Moreno, Christian D. KummerowRemote Sensing Incremental Learning
University of Texas at El Paso, El Paso, TX, USA · Colorado State University, Fort Collins, CO, USA
VNDUQE: Information-Theoretic Novelty Detection using Deep Variational Information Bottleneck
May 12, 2026Aryan Gondkar, Hayder Radha, Yiming DengUncertainty Quantification Information Bottleneck
Nondestructive Evaluation Lab, Department of Electrical and Computer Engineering Michigan State University East Lansing, MI · Department of Electrical and Computer Engineering Michigan State University East Lansing, MI
Neuromorphic Object Detection: An In-Depth Study and Future Directions
Jul 26, 2026Jianing Li, Dianze Li, Arren Glover +5Neuromorphic Computing Event-Based Object Detection
Beijing Key Laboratory of Brain-inspired Spiking Large Models, School of Computer Science, Peking University, Beijing 100871, China · Peng Cheng Laboratory, Shenzhen 518000, China · Event-Driven Perception for Robotics, Italian Institute of Technology, Genoa 16163, Italy +6

Sep 1, 2026J/K move · Enter open · S save

A Sensor-Adaptive Incremental Learning Framework for Artifact Detection in Satellite Precipitation Data

Andres F. Monsalve, Hernan A. Moreno, Christian D. Kummerow

University of Texas at El Paso, El Paso, TX, USA · Colorado State University, Fort Collins, CO, USA

Remote Sensing Incremental Learning Anomaly Detection

Open paper AskPDF arXiv

Historically, retrieving rainfall data from satellite imagery has been the domain of space agencies. However, in recent years, the development of cheaper, more compact satellites (SmallSats) capable of detecting rainfall proxies has led to a significant increase in private-sector initiatives for satellite launch and surface precipitation products. This rapid growth has yet to be matched by data validation efforts. Consequently, the need for a robust tool to detect anomalies in near-real-time data before it is disseminated to the public has become critical. In this paper, we present the development of an anomaly-detection system to identify artifacts in global satellite-based rainfall products. The developed framework leverages pre-trained computer vision models and incorporates scarce human-labeled data to detect specific anomalies. Our proposed anomaly detection strategy is tested on data from the Special Sensor Microwave Imager (SSMI) and the Special Sensor Microwave Imager/Sounder (SSMIS). Results demonstrate the efficacy of our approach at separating regular orbits from artifact-containing orbits for each satellite, with performance comparable to state-of-the-art in-place methods. Additionally, the framework offers explainability and the capacity for iterative refinement following false-positive or false-negative classifications.

Faults	FZ ok	CMM ok	KR ok	GM ok	FZ ke	CMM ke	KR ke	GM ke
5%	50	10	16	16	130	20	70	75
10%	55	12	13	10	160	50	70	112
15%	55	12	15	10	–	25	34	–
20%	55	12	15	12	–	30	60	62

Faults	FZ ok	CMM ok	KR ok	GM ok	FZ ke	CMM ke	KR ke	GM ke
5%	50	10	16	16	130	20	70	75
10%	55	12	13	10	160	50	70	112
15%	55	12	15	10	–	25	34	–
20%	55	12	15	12	–	30	60	62

Abstract

Fault Detection

Explore similar work

A Sensor-Adaptive Incremental Learning Framework for Artifact Detection in Satellite Precipitation Data

VNDUQE: Information-Theoretic Novelty Detection using Deep Variational Information Bottleneck

Neuromorphic Object Detection: An In-Depth Study and Future Directions

Faults	FZ ok	CMM ok	KR ok	GM ok	FZ ke	CMM ke	KR ke	GM ke
5%	50	10	16	16	130	20	70	75
10%	55	12	13	10	160	50	70	112
15%	55	12	15	10	–	25	34	–
20%	55	12	15	12	–	30	60	62